doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.ppat.1007935 | Individual liver plasmacytoid dendritic cells are capable of producing IFNα and multiple additional cytokines during chronic HCV infection | Plasmacytoid dendritic cells (pDCs) are “natural” interferon α (IFNα)-producing cells. Despite their importance to antiviral defense, autoimmunity, and ischemic liver graft injury, because DC subsets are rare and heterogeneous, basic questions about liver pDC function and capacity to make cytokines remain unanswered. Previous investigations failed to consistently detect IFNα mRNA in HCV-infected livers, suggesting that pDCs may be incapable of producing IFNα. We used a combination of molecular, biochemical, cytometric, and high-dimensional techniques to analyze DC frequencies/functions in liver and peripheral blood mononuclear cells (PBMCs) of hepatitis C virus (HCV)-infected patients, to examine correlations between DC function and gene expression of matched whole liver tissue and liver mononuclear cells (LMCs), and to determine if pDCs can produce multiple cytokines. T cells often produce multiple cytokines/chemokines but until recently technical limitations have precluded tests of polyfunctionality in individual pDCs. Mass cytometry (CyTOF) revealed that liver pDCs are the only LMC that produces detectable amounts of IFNα in response TLR-7/8 stimulation. Liver pDCs secreted large quantities of IFNα (~2 million molecules of IFNα/cell/hour) and produced more IFNα than PBMCs after stimulation, p = 0.0001. LMCs secreted >14-fold more IFNα than IFNλ in 4 hours. Liver pDC frequency positively correlated with whole liver expression of “IFNα-response” pathway (R2 = 0.58, p = 0.007) and “monocyte surface” signature (R2 = 0.54, p = 0.01). Mass cytometry revealed that IFNα-producing pDCs were highly polyfunctional; >90% also made 2–4 additional cytokines/chemokines of our test set of 10. Liver BDCA1 DCs, but not BDCA3 DCs, were similarly polyfunctional. pDCs from a healthy liver were also polyfunctional. Our data show that liver pDCs retain the ability to make abundant IFNα during chronic HCV infection and produce many other immune modulators. Polyfunctional liver pDCs are likely to be key drivers of inflammation and immune activation during chronic HCV infection.
| This is a detailed characterization of human liver plasmacytoid dendritic cells from patients with a chronic viral infection. It revealed that these rare innate immune cells can become point sources of multiple immune activators and pro-inflammatory mediators. This study adds new information about the fundamental properties of pDCs, which are traditionally known as “natural interferon producing cells.” In fact, these cells produce an array of bioactive molecules and may play an important role in organizing the liver’s immune response.
| Plasmacytoid dendritic cells (pDCs) are rare innate immune cells that comprise about 0.5% of peripheral blood mononuclear cells (PBMCs). They migrate into tissues and are known as “natural” producers of interferon alpha (IFNα). pDCs constitutively express toll-like receptor (TLR)-7 and TLR-9, as well as interferon regulatory factor (IRF)-7, enabling them to detect viral nucleic acids and to quickly secrete type I IFNs (IFNα and IFNβ), which bind neighboring cells and induce hundreds of IFN stimulated genes (ISGs), initiating antiviral defenses.
The activity of pDCs during HCV infection remains obscure. Several groups examined pDC frequency and function during chronic infection. Nearly all found a reduced frequency of pDCs in blood [1–7]. Some reported that circulating pDCs are functionally intact [6, 7], but the majority reported impairment after stimulation with various TLR ligands [1–5], which was attributed to toxic effects of tumor necrosis factor α (TNFα) [5] and direct inhibitory effects of HCV proteins [8] [9]. In contrast to these inhibitory effects, HCV RNA stimulates pDCs by activating TLR-7 and RIG-I [10–13].
Past studies of patients and chimpanzees provide circumstantial evidence that liver pDCs do not produce IFNα during HCV infection. IFNA mRNA levels are not consistently elevated during acute [14] or chronic [15, 16] infection, and liver IFNA2 mRNA levels rose to detectable levels only after HCV was cured [17], suggesting that HCV may shut down IFNα production. The absence of detectable IFNA mRNA was initially puzzling because ISGs are highly induced in HCV-infected liver [18], but the discovery of type III IFNs (IFNλs) provided a possible explanation for the seeming paradox because these cytokines up-regulate many of the same genes as IFNα [19]. These investigations left the question of pDC functionality during HCV infection unanswered.
We explored an alternative explanation: the possibility that intrahepatic pDCs remain functional during chronic HCV infection but generate an IFNA mRNA signal that is too low to be detected in extracts of whole liver tissue. To improve the signal-to-noise ratio, liver mononuclear cells (LMCs) were purified and examined in parallel with whole liver tissue and PBMCs. We found that liver pDCs retain the ability to produce large quantities of IFNα and made more IFNα per cell than blood pDCs. Liver pDC frequency had strong positive correlations with whole liver expression of the “IFNα-response (I)” pathway of blood transcriptomic (BT) modules and with three monocyte-specific modules, indicating that pDCs are active in vivo and have effects on other liver immune cells. Single-cell mass cytometry (CyTOF) revealed that liver pDCs are the only LMCs that make IFNα and demonstrated that most IFNα-producing pDCs are polyfunctional and a single IFNα+ pDC makes several additional cytokines/chemokines. Individual liver BDCA1 DCs were similarly polyfunctional. These findings demonstrate that intrahepatic pDCs and BDCA1 DCs can be intense point sources of a constellation of immune activators and they establish that liver pDCs remain competent for IFNα production despite chronic exposure to viral products.
Medical record data, liver, and blood were obtained from 19 patients with chronic HCV infection who were undergoing liver transplantation. The median age was 62 years [interquartile range (IQR), 59–65]; 79% were male (S1 Table). The median natural model for end stage liver disease (MELD), which assesses the amount of liver damage, was 18 (IQR, 13–32). LMCs and PBMCs were prepared by density gradient centrifugation and either examined immediately, “ex vivo”, by flow cytometry, microarray, and RT/PCR or they were incubated for four hours with TLR ligands (or media alone) prior to analysis (Fig 1A). Total ex vivo liver mRNA of 11 of the 19 patients was analyzed by microarray and RT/qPCR (Fig 1B). LMCs of three additional anonymous HCV+ patients were analyzed by CyTOF (Fig 1C).
pDC frequencies in CD45+ PBMCs and LMCs were determined by flow cytometry using the gating strategy in S1 Fig, although this gating strategy does not rule out the possibility of including preDCs within the pDC population [20]. The pDC frequency was the same in liver and blood, suggesting that pDCs do not concentrate in the liver (Fig 2A), but the median fluorescent intensity (MFI) of HLA-DR on the liver pDCs was higher (Fig 2B), indicating greater activation. The impact of pDCs on surrounding liver cells was analyzed by examining correlations between pDC frequency and transcriptomic data of 11 whole livers. Four modules had a strong correlation (R2≥0.5) with liver pDC frequency (Fig 2C–2E): “IFNα response (I)” (Fig 2D), “Monocyte surface signature” (Fig 2E), “Enriched in activated dendritic cells/monocytes,” and “Enriched in monocytes (surface).” These results indicate that liver pDCs are active in vivo. Fifty-four percent of the genes in these four blood transcription (BT) modules are part of the Interferome [21]. The genes in these and other BT modules are listed in S4 Table. Liver pDC frequency also strongly correlated with the percentage of liver HCV RNA in double-stranded form (Fig 2F), consistent with published data showing that IFNα increases HCV RNA duplexes [22]. Analysis of clinical data revealed a significant inverse relationship between liver pDC frequency and blood platelet counts (p = 0.03, Fig 2G).
We also analyzed two additional DC subsets, BDCA1 and BDCA3 DCs. As determined by flow cytometry, BDCA1+ (classical) DCs were enriched in blood compared to liver (Fig 3A, left), while BDCA3+ (cross-presenting) DCs were enriched in liver (Fig 3B, left). The MFI of HLA-DR was higher on liver BDCA1+ DCs than on their counterparts in blood (Fig 3A, right), but HLA-DR did not differ between liver and blood BDCA3+ DCs (Fig 3B, right).
Single sample gene set enrichment analysis (ssGSEA) revealed a strong correlation between the frequency of intrahepatic BDCA1+ DCs and expression of “Hox cluster III”, R2 = 0.6 (Fig 3C and 3D) and “Cell movement, Adhesion & Platelet activation”, R2 = 0.52 (Fig 3C and 3E). Hox genes are critical for proliferation and differentiation of hematopoietic cells, especially T cells [23]. Liver BDCA3+ frequency strongly correlated with three pathways (Fig 3F): “Formyl peptide mediated neutrophil response” (Fig 3G), “Cell division stimulated CD4+ T cells” (Fig 3H), and “Enriched in B cells (IV).” Taken together, these data suggest that BDCA1+ and BDCA3+ DCs increase immune infiltration, migration, and induction of adaptive and innate immune responses.
CyTOF was used to definitively identify the IFNα-producing liver cells (Fig 1C); the CyTOF antibody panel is presented in S5 Table. LMCs were analyzed after incubation for 4 hours with media or R848, a TLR7/8 agonist, in the presence of brefeldin A (BFA) to block cytokine secretion. viSNE was employed to project the high-dimensional data onto two-dimensional space. Nine major subsets of CD45+ cells were identified based on canonical markers (Fig 4A; S1 Fig). pDCs comprised a distinct cluster in all three patients (Fig 4A, orange). The normalized mean signal intensity (nMSI) for IFNα (Fig 4B) was determined for all subsets and IFNα positivity was plotted for each population (Fig 4C). pDCs comprised the only population of IFNα+ cells (Fig 4B and 4C); on average 26% (13–40%) of the pDCs expressed IFNα following R848 stimulation (Fig 4C and 4D).
The quantity of IFNs secreted by liver pDCs and other LMCs was investigated (Fig 5A–5D) after incubation in media alone, with R848 or with Poly I:C, a TLR-3 agonist that is important for IFNλ production [24]. LMCs secreted 3-fold more IFNα than PBMCs in response to R848 stimulation, 345±207 pg/mL vs. 115±111 pg/mL, p = 0.0001 (Fig 5A).
The amount of IFNα secreted per liver pDC per hour was calculated by combining data from flow cytometry, Luminex, and CyTOF. Secretion assays contained a mean of 6000 pDCs, with ~26% (1560 pDCs) producing IFNα. Mean secretion was 86 pg of IFNα2a/2b/hour, which indicates that each IFNα-producing pDC was secreting over 1.7x106 molecules per hour. Actual secretion may exceed this number because the Luminex assay targets only IFNα2a/2b and there are 11 additional forms of human IFNα. Wimmers et al. showed that over the course of 12 hours, the small percentage of pDCs that initially produce IFNα later induce IFNα production in neighboring pDCs, in a local amplification loop [25]. Our calculation does not consider this amplification process because our incubations were only for 4 hours.
The amount of IFNα secreted by LMCs correlated strongly with the frequency of liver pDCs, p = 0.0002 (Fig 5E), consistent with CyTOF data indicating that pDCs are the only IFNα producers (Fig 4B). It also correlated with expression of the “Immune activation-generic cluster” in whole liver, p = 0.007 (Fig 5F), as well as expression of the KEGG pathway “Class I MHC Mediated Antigen Processing Presentation” (p = 0.007, Fig 5G) and “Enriched in B cells (IV)” (p = 0.026, Fig 5H). Minimal IFNα was secreted by LMCs or PBMCs incubated in media alone or with Poly I:C. Compared to IFNα, LMCs secreted far less IFNλ1 or 2/3. The greatest amount was 24±21 pg/mL of IFNλ1 (Fig 5B–5D), which is more than 14-fold lower than the greatest amount of IFNα. The quantities of IFNλ secreted by LMCs in response to TLR stimulation did not correlate with the frequency of any of the three DC subsets we examined.
RT/qPCR and microarrays were used to investigate gene expression in LMCs, PBMCs, and whole liver. Notably, IFNA1 mRNA was readily detected in ex vivo LMCs by RT/qPCR (Fig 6A), but neither IFNA1 mRNA, nor any of the other IFN mRNAs could be detected by RT/qPCR in whole liver: all 11 whole liver extracts had Ct values above 35. Ex vivo LMCs expressed higher levels of IFNA1, IFNB, and type III IFN mRNA (IL29 and IL28A/B) than ex vivo PBMCs (Fig 6A–6D). Consistent with the RT/qPCR results, microarray data showed that ex vivo LMCs had higher expression of the “Immune Activation–Generic Cluster” and higher “TLR and Inflammatory Signaling” than PBMCs (Fig 6E and S2B Fig, respectively).
IFN gene expression was also examined following four hours of incubation in media with and without TLR ligands. RT/qPCR analysis showed that the TLR-7/8 agonist, R848, increased expression of IFNA1 and IFNB in LMCs compared to ex vivo LMCs and compared to LMCs incubated in media (Fig 6A and 6B). R848 treatment also increased IL29 expression in LMCs relative to ex vivo LMCs, but it did not increase IL28A/B expression (Fig 6C and 6D). GSEA of microarray data of LMCs showed that R848 treatment up-regulated many IFNα genes (Fig 6F) and enhanced the “antiviral IFN signature” (S2A Fig). The TLR-3 agonist, Poly I:C, did not induce IFNA1 in LMCs, but it did induce IFNB, IL29, and IL28A/B (Fig 6A–6D). Poly I:C enhanced the “antiviral IFN signature” relative to ex vivo LMCs (S2C Fig), but not as intensely as R848 (S2D Fig). Compared to ex vivo or media, Poly I:C did not increase expression of any of the type I or type III IFN genes in PBMCs.
To explore the cytokine milieu more fully, TNFα, CXCL10, IL-6, and IL-10 secretion were examined by Luminex. LMCs made an average of 10-fold more TNFα (66±79 vs. 6±5 pg/mL), 3-fold more CXCL10 (278±236 vs. 98±110 pg/mL), 8-fold more IL-10 (16±12 vs. 2±2 pg/mL), and 10-fold more IL-6 (269±472 vs. 7±10 pg/mL) than PBMCs after incubation in media (without TLR stimulation), p≤0.05 for all comparisons (S3A–S3D Fig).
We used mass cytometry to measure the ability of individual pDCs to produce multiple factors, a capacity known as “polyfunctionality”. To test for polyfunctionality, we used a CyTOF panel with antibodies against 10 cytokines/chemokines, IFNα, TNFα, IL-8, IFNλ1 (IL-29), IL-6, IL-1β, IL-10, CCL3, CCL4, and CXCL10. Polyfunctionality was initially explored by selecting pDCs, BDCA1 DCs, or BDCA3 DCs of each patient by manual gating and then further gating on each combination of cytokines/chemokines (Fig 7A and 7B). IFNα+ pDCs were highly polyfunctional: >90% produced two or more additional factors. Remarkably, 5% of the IFNα-producing pDCs made five or more additional cytokines/chemokines (Fig 7A). IFNα- pDCs were less polyfunctional: 33% did not make any of the factors in our panel and most (58%) made only 2 or 3. Approximately 74% of the pDCs expressed TNFα and 73% expressed IL-8, more than the 26% that make the signature cytokine, IFNα (Fig 7D). BDCA1 DCs had a comparable level of polyfunctionality as pDCs, while BDCA3 DCs were mostly negative for the factors we analyzed (Fig 7B).
We used two additional analysis methods to characterize pDCs and to ensure that our findings were consistent regardless of which analytical method was applied. In the first approach, the pDCs of each patient were selected by manual gating and then Phenograph was used to identify subpopulations across all three patients (Fig 7C and Fig 7E, Fig 7F) [26]. Four metaclusters of R848-stimulated pDCs were identified (Fig 7F, blue box). Three (metaclusters 8, 17, and 7) had a high IFNα normalized mean signal intensity (nMSI) and also highly expressed TNFα and IL-8; they had variable expression of IL-6, CCL3, CCL4, and/or IFNλ1 (IL-29). In a final analysis, pDCs were first clustered using viSNE [27] followed by manually gating (Fig 7C–7G, and 7H). This process delineated eight clusters (Fig 7G). Similar to the metaclusters (Fig 7F), the viSNE clusters with high expression of IFNα (viSNE clusters 5, 6, 8) also had high levels of IL-8, TNFα, and they had variable expression of IL-6 and IL-29 (Fig 7H, left); nearly all the cells in cytokine-producing clusters came from cells stimulated by R848 (Fig 7H, right), consistent with manual gating in Fig 7I.
To obtain cells from a liver of a patient with no underlying liver disease, we turned to the buffer solution that is used to transport donor livers. This “perfusate” is a validated source of liver immune cells and has been used in previous studies [28]. We obtained two perfusates, one from a healthy liver donor and the other from a HCV-infected liver donor. Both livers were deemed healthy enough for organ donation. We assayed the pDCs for polyfunctionality. After a four hour stimulation with R848, 16% of pDCs from the healthy perfusate made IFNα (Fig 8A, top), while only 8% of the HCV+ perfusate’s pDCs made IFNα (Fig 8B, top). We used the same CyTOF panel with antibodies against 10 cytokines/chemokines to explore polyfunctionality (Fig 8A and 8B, bottom). IFNα+ pDCs were more polyfunctional than their IFNα- counterparts for both the healthy and HCV+ perfusates. Nonetheless, pDCs from a patient with no underlying liver disease were polyfunctional.
This is a detailed characterization of human liver pDCs from patients with a chronic hepatitis virus infection. It revealed that these rare innate immune cells are point sources of multiple immune activators and pro-inflammatory mediators, adding important new details about the fundamental capabilities of tissue-specific pDCs. Single cell mass cytometry (CyTOF) revealed that liver pDCs are the only IFNα producers among LMCs and revealed that only 15–40% synthesize it when stimulated with a TLR7/8 agonist (Figs 4D and 7D). Activated pDCs secreted large amounts of IFNα protein: ~2 million molecules per pDC per hour. While pDCs are traditionally known as “natural interferon producing cells,” our data revealed that they produce an array of bioactive molecules. Approximately 20% of pDCs make IFNα plus 4 of the other nine cytokines/chemokines we analyzed [TNFα, IL-8, IFNλ1 (IL-29), IL-6, IL-1β, IL-10, CCL3, CCL4, and CXCL10] and 13% make IFNα plus 5 or more (Fig 7A). Most IFNα+ pDCs expressed TNFα and IL-8, with variable amounts of IL-6, CCL3, CCL4, and IFNλ1 (IL-29). We did not determine the percentage of intrahepatic TNFα that is made by liver pDCs, but blood pDCs are major producers [29], suggesting that liver pDCs may be a significant source of this proinflammatory cytokine.
The perfusates from organ donors with livers healthy enough to transplant show that end-stage-liver disease is not a leading factor in whether or not pDCs are polyfunctional (Fig 8), nor does it seem to depend on HCV infection. The healthy perfusate pDCs were slightly more polyfunctional than the HCV-infected perfusate pDCs (Fig 8, bottom), suggesting that polyfunctionality is an innate characteristic of pDCs after stimulation through TLR7. It is worth noting that pDCs purified from liver tissue are more polyfunctional than the pDCs that did not pass through our lengthy extraction process (compare Fig 7A and Fig 8A, bottom). In future studies we hope to investigate pDCs from additional donors.
Four previous studies used flow cytometry to investigate pDC polyfunctionality; they demonstrated that individual pDCs can produce IFNα, TNFα, and IL-6 [30–33]. By using CyTOF, we were able to interrogate a larger number of factors than flow cytometry allows. We found that individual cells could express six or more cytokines and chemokines. It is likely that a broader CyTOF panel would reveal an even greater number of immune factors. One flow cytometric study demonstrated that gut pDCs of simian immunodeficiency virus-infected macaques secrete IFNα, TNFα, and MIP-1β [30]. Two showed that most activated human blood pDCs express two of the three factors analyzed [31, 32]. The fourth revealed that more than 95% of blood pDCs make two or fewer cytokines out of the four that were tested (IFNα, TNFα, IL-6, and IFNγ) after stimulation with R848 [33]. In addition, using a combination of single-cell RNA sequencing and single-cell cytokine analysis, Wimmers et al. recently reported that only a fraction of pDCs make IFNα, while most make TNFα, consistent with our data [25]. With our more comprehensive panel (10 cytokines and chemokines were analyzed) using CyTOF, we were able to show that liver pDCs, and similarly liver BDCA1 DCs, are highly polyfunctional for cytokine/chemokine production. Traditionally, polyfunctionality has been attributed to T cells and interpreted as an indicator of high functional capacity. Further studies are needed to understand the biological significance of having a single dendritic cell acting as a beacon of secreting multiple immune modulators. We postulate that polyfunctionality is important because it allows a single pDC to activate multiple signaling pathways on target cells, potentially raising the response to higher levels than could be achieved through the maximal activation of a single pathway. The exact composition of the cytokine/chemokine mix may also be important for eliciting appropriate responses.
Many recent studies provide novel information about the heterogeneity of pDCs. This heterogeneity may influence which pDCs acquire polyfunctionality. Alculumbre et al. demonstrated that when blood pDCs were stimulated with either influenza or CpG for 24 hours, the population matured into three distinct functional groups [34]. One subset produced IFNα, another stimulated T cells and a third had elements of both. At their 4 hour time point, the three subsets were not apparent. In addition, single cell RNA sequencing revealed that there are cells within the typical pDC gate that are not pDCs [20, 35]. Villani et al. showed that these pDC-like cells express AXL and SIGLEC1/6 but in fact function like conventional DCs by activating T cells [35]. Michea et al. showed that the microenvironment can increase pDC heterogeneity [36]. MacParland et al. demonstrated that the liver microenvironment changes the phenotype of resident macrophages [37] suggesting that the liver microenvironment may impact the phenotype of liver pDCs.
Our study sheds new light on the paradoxical absence of detectable type I IFN mRNA in the HCV-infected liver despite the central role IFNα plays in host viral defenses. RT/qPCR data revealed that while IFNα/β mRNAs were not detectable in whole liver RNA extracts, confirming published findings [15, 16], they were readily detected in extracts of isolated liver leukocytes, demonstrating that purifying LMCs prior to mRNA analysis improved the signal-to-noise ratio in the RT/qPCR assay.
Importantly, type I IFN mRNAs were detected in whole liver using microarrays, indicating that mRNA expression occurred in the whole liver and did not require cell isolation. Consistent with this, the frequency of liver pDCs strongly correlated with expression of BT module of the IFNα response and three monocyte-specific modules, indicating that pDCs activate surrounding immune cells. The liver pDC frequency also strongly correlated with the percentage of HCV RNA in double-stranded form, which provides additional evidence that IFNαs were produced in vivo; published data establish that IFNα increases the percentage of double-stranded HCV RNA [12].
R848-stimulation strongly induced IFNα genes in LMCs in vitro, as demonstrated by RT/qPCR and transcriptomic analysis. The amount of IFNα produced in response to R848 strongly correlated with the frequency of liver pDCs and with expression of the “Immune activation–generic cluster” in whole liver, suggesting that pDCs activate multiple antimicrobial, inflammatory, and immune response pathways in liver immune cells, as depicted in S4 Fig. pDCs retain the ability to respond to TLR ligands even in the face of HCV.
Our study revealed interesting differences between pDCs in the liver and blood during chronic HCV infection. Liver pDCs were more highly activated, as indicated by higher expression of HLA-DR and ex vivo LMCs had higher expression of the “Immune Activation–Generic Cluster” and “TLR and Inflammatory Signaling” genes than ex vivo PBMCs (Fig 6E and S2B Fig). LMCs secreted more IFNα than PBMCs. Some of the observed differences may reflect the different procedures used to prepare PBMCs and LMCs, the latter were exposed to collagenase/DNase and mechanical disruption, which may have activated the liver pDCs. However, single cell RNA sequencing studies have shown that the microenvironment plays an important role in shaping the phenotype and function of immune cells [36, 37]. Importantly, LMCs remained responsive to TLR agonists during ex vivo culture, indicating that whatever effect the extraction process might have had it did not render the cells refectory to further activation. Additionally, despite on-going exposure to viral proteins like pathogen-associated molecular patterns (PAMPs) and cellular debris (danger-associated molecular patterns), pDCs remained functional.
When stimulated for four hours, LMCs secreted minimal IFNλ, whereas liver BDCA3+ DCs produced abundant IFNλ3 in response to 24 hours of stimulation [38] indicating that our experimental conditions were not optimal for IFNλ production. This suggests that a four hour time course is not appropriate for analyzing type III IFN responses. While acknowledging the importance of type III IFNs, we consider it likely that pDCs play an important role in liver immune responses during chronic HCV infection because the antiviral signature in LMCs was more strongly induced by the TLR-7/8 ligand than by the TLR-3 ligand. Because pDCs express TLR-7 and not TLR-3 [24], this finding suggests that pDCs, and by extension IFNα, stimulate antiviral defenses in the HCV-infected liver (S4 Fig). Our results are consistent with evidence that blood pDCs make IFNα in response to cell culture-derived HCV [10], a TLR-7 agonist.
In addition to LMCs, hepatocytes, sinusoidal endothelial cells, and other liver cells can produce IFNs in response to stress, including HCV infection. Hepatocytes secrete greater amounts of IFNλs than IFNαs [19]. Liver endothelial cells make primarily IFNλs after HCV exposure [39]. Type I and type III IFNs have distinctive effects. IFNα is more effective at inhibiting HCV replication in vitro [39], but IFNλ induces a more prolonged ISG induction [40]. Moreover, HCV infection of primary human hepatocytes causes a down-modulation of IFNAR1[41]. This down-modulation, if it occurs during chronic HCV infection, could protect HCV from antiviral defenses and foster chronic inflammation.
After successful HCV treatment, expression of some IFNA genes may increase [17]. If expression continues into the post-cure period, it could drive persistent liver inflammation, while also helping to suppress HCV recrudescence. Liver injury and inflammation continues in up to 66% of patients cured of HCV [42] and the immunopathology pre- and post-cure may involve some of the same molecular pathways. Pathologists were recently warned that the histopathology of liver transplant patients cured of HCV so closely resembles that of chronic infection that the conditions can be easily mistaken for each other [43]. Inflammation increases cancer risk; the HCC risk in cured cirrhotic patients remains elevated, up to 5% annually [44].
A limitation of the study is that most of the samples came from HCV-infected patients with end-stage liver disease and/or hepatocellular carcinoma. Future experiments need to be done on liver pDCs from additional sources.
In summary: Our study provides important new details about primary human liver pDCs and their activity during chronic HCV infection. The investigation used a novel combination of CyTOF, molecular techniques, cytokine quantitation, and cell purification methods and provided evidence that activated liver pDCs produce large quantities of IFNα. The liver pDC response to stimulation was heterogeneous, as also reported by Wimmers et al. for blood pDCs [25]. A minority of liver pDCs produced IFNα and most IFNα+ pDCs also expressed 2 or more of the other nine cytokines/chemokines we examined. Liver BDCA1+ DCs were also highly polyfunctional. The circuits regulating gene expression in polyfunctional liver DC subsets merit investigation as the orchestrators of complex immune responses and as potential therapeutic targets.
This is a prospective study of specimens and medical records of 19 HCV-positive adults who received a liver transplant at the Mount Sinai Medical Center between 11/2013 and 8/2014 and who gave written informed consent. Blood for research and clinical testing was collected before surgery. Explants of three additional anonymous HCV-infected patients were also analyzed. Perfusates of two anonymous liver donors, one healthy and one HCV-infected, were collected and analyzed. The study was approved by Mount Sinai’s IRB in accordance with Helsinki guidelines. No explants were obtained from prisoners or other institutionalized persons.
Specimens were brought to the laboratory at a median of 45 min post-explantation. The liver capsule was removed. Tissue was minced, washed in Hank’s balanced salt solution (HBSS)/1% fetal calf serum (FCS), incubated in RPMI/5% FCS/0.1 mg/mL collagenase/50 μg/mL DNase at 37°C for 30 min, shaking every 5 min. Tissue was pressed through stainless steel mesh while washing with HBSS/1% FCS. Cells were washed and resuspended in HBSS/1% FCS and filtered through 100μm nylon mesh. Percoll gradients were used to purify LMCs from the filtrates [45, 46] and from PBMCs.
Perfusates were kept on ice during transportation and brought to the laboratory after anhepatic phase of liver transplantation was complete. Perfusates were spun down and resuspended in HBSS/1% FCS. Percoll gradients were used to purify PMCs from the perfusates.
The flow cytometry panel for DC subsets appears in S3 Table. Cells were stained with the surface stain panel, then fixed with 2% paraformaldehyde solution (Thermo Scientific) in PBS. Samples were run on an LSR Fortessa (BD) and analyzed using Flojo.
Without BFA: One million PBMCs or LMCs per 0.5mL media were incubated in RPMI with 10% FBS for four hours alone or with 1 μg/mL R848 or with 50 μg/mL Poly I:C at 37°C. Supernatants were collected for proteomic analysis. Cells were collected in Trizol (Life Technologies) for RT/qPCR and microarray analysis. With 1:1000 BFA (eBioscience): Up to ten million PBMCs, LMCs, or PMCs in 0.5mL media were incubated in RPMI with 10% FBS for four hours alone or with 1 μg/mL R848 at 37°C. Cells were collected for CyTOF antibody staining and acquisition.
RNAs were purified as before [47]. cDNA was made using SuperScriptIII First-Strand Synthesis (Invitrogen) and amplicons were quantified using the LightCycler480 SYBR Green II Master kit (Roche). Expression of genes was calculated using the ΔΔCt method normalized to RPS11 and to expression of the PBMC ex vivo sample. Primers for IFNA1, IFNB, IFNL1, IFNl2/3, RPS11, and TNFA were described previously [48]. Double stranded HCV RNA was quantified as described previously [47].
Luminex multiplex cytokine assays (Millipore) quantified IFNα2a/b, IFNλ1 (IL-29), IFNλ2 (IL-28A), IFNλ3 (IL-28B), interferon gamma-induced protein 10 (IP10 aka CXCL10), interleukin 6 (IL-6), IL-10, and TNFα.
Profiling data from Illumina Human-HT-Expression Beadchips were normalized using GenomeStudio’s quantile method. GenePattern was used for gene set enrichment analysis (GSEA) and single sample GSEA (ssGSEA) of immune pathways [49] using BT modules [50] and KEGG pathways. A false discovery rate (FDR) below 0.25 was considered statistically significant. LMCs/PBMCs: To obtain sufficient RNA, LMC samples of matched pairs of patients were pooled. Matching was based on age, gender, HCV genotype, baseline HCV RNA, natural MELD score, and HCC (yes/no). PBMCs were pooled similarly. Whole liver: Whole liver microarray data of 11 of the 19 patients consented for this study was used as before [47].
Panel presented in S5 Table. Samples were washed, fixed, and permeabilized (eBiosciences) then stained with intracellular antibodies. Samples were stored at 4°C in Ir intercalator (Fluidigm) in 2% formaldehyde until acquisition. Before acquisition, samples were mixed with EQ4 Element Beads (Fluidigm) and were acquired on a CyTOF2 (Fluidigm). Data were normalized using bead-based normalization in the CyTOF software and gated to exclude beads, dead cells, and doublets. Method 1: Gated pDCs were analyzed using an automated CyTOF data analysis pipeline at the Mt. Sinai HIMC, which uses an R-based implementation of Phenograph [26], an agonistic clustering method that utilizes the graph-based Louvain algorithm for community detection and identifies a hierarchical structure of distinct phenotypic communities. We utilized dynamic activation markers and intracellular cytokines as clustering parameters to resolve functional heterogeneity within the pDC population. Phenotypic clusters from 3 donors were meta-clustered identify consistent populations that could be reproducibly detected across individuals, thereby generating a consistent cluster structure across all samples in the dataset, while preserving the diversity and heterogeneity of all subpopulations. Method 2: viSNE was used to cluster the single-cell pDC data, creating t-distributed stochastic neighbor embedding (tSNE) plots in Cytobank [27]. viSNE uses a dimensionality-reducing algorithm to express multi-dimensional data in two dimensions. Canonical cell surface markers were then analyzed to identify cell populations overlaid on the viSNE map or manually identified clusters were gated on and overlaid on the viSNE map.
The Luminex-measured mean quantity of secreted IFNα was divided by the incubation time to determine production per hour. This quantity was divided by the molecular mass of IFNα and multiplied by Avogadro’s constant. The result (the molecules of IFNα secreted per hour) was divided by the mean number of IFNα-producing pDCs per reaction, which was determined by multiplying the number of LMCs per reaction by the frequency of IFNα-producing pDCs as determined by CyTOF.
GraphPad Prism was used. Paired and unpaired t-tests were performed. Pearson’s correlation coefficient was used for correlations.
|
10.1371/journal.pntd.0004219 | Loop-Mediated Isothermal Amplification for Laboratory Confirmation of Buruli Ulcer Disease—Towards a Point-of-Care Test | As the major burden of Buruli ulcer disease (BUD) occurs in remote rural areas, development of point-of-care (POC) tests is considered a research priority to bring diagnostic services closer to the patients. Loop-mediated isothermal amplification (LAMP), a simple, robust and cost-effective technology, has been selected as a promising POC test candidate. Three BUD-specific LAMP assays are available to date, but various technical challenges still hamper decentralized application. To overcome the requirement of cold-chains for transport and storage of reagents, the aim of this study was to establish a dry-reagent-based LAMP assay (DRB-LAMP) employing lyophilized reagents.
Following the design of an IS2404 based conventional LAMP (cLAMP) assay suitable to apply lyophilized reagents, a lyophylization protocol for the DRB-LAMP format was developed. Clinical performance of cLAMP was validated through testing of 140 clinical samples from 91 suspected BUD cases by routine assays, i.e. IS2404 dry-reagent-based (DRB) PCR, conventional IS2404 PCR (cPCR), IS2404 qPCR, compared to cLAMP. Whereas qPCR rendered an additional 10% of confirmed cases and samples respectively, case confirmation and positivity rates of DRB-PCR or cPCR (64.84% and 56.43%; 100% concordant results in both assays) and cLAMP (62.64% and 52.86%) were comparable and there was no significant difference between the sensitivity of the assays (DRB PCR and cPCR, 86.76%; cLAMP, 83.82%). Likewise, sensitivity of cLAMP (95.83%) and DRB-LAMP (91.67%) were comparable as determined on a set of 24 samples tested positive in all routine assays.
Both LAMP formats constitute equivalent alternatives to conventional PCR techniques. Provided the envisaged availability of field friendly DNA extraction formats, both assays are suitable for decentralized laboratory confirmation of BUD, whereby DRB-LAMP scores with the additional advantage of not requiring cold-chains. As validation of the assays was conducted in a third-level laboratory environment, field based evaluation trials are necessary to determine the clinical performance at peripheral health care level.
| Buruli ulcer disease (BUD) mainly occurs in remote rural areas of Sub-Saharan Africa, affects skin and soft tissue, and may lead to severe disabilities. Therefore, early diagnosis and treatment with antimycobacterial therapy are essential whereby the WHO recommends laboratory confirmation of 70% of the cases. As the current diagnostic gold standard (polymerase chain reaction [PCR]) is restricted to third-level laboratories, development of confirmatory point-of-care (POC) tests for BUD applicable at primary health care level has become a research priority to bring diagnosis closer to where the patients are. Loop-mediated isothermal amplification (LAMP) has been selected by the WHO as one of the promising candidate technologies for POC tests. The aim of this study was to establish and validate a LAMP assay applying lyophilized reagents which are stable at ambient temperature, thus avoiding the need for cold-chains. The results from this study suggest that the assay provides a valuable alternative to other PCR tests as currently used for laboratory confirmation of BUD.
| Buruli ulcer disease (BUD), caused by Mycobacterium ulcerans, is an infectious disease affecting skin, soft tissues and sometimes the bones. The major endemic foci occur in rural areas of Sub-Saharan Africa where BUD mainly affects children below the age of 15 years.
Antimycobacterial therapy can cure up to 80% of patients diagnosed in early stages of the disease. If treated in advanced stages or left untreated, extensive destruction of tissue followed by fibrous scarring and contractures may lead to severe sequelae such as functional limitation of affected joints, which occur in up to 25% of cases. In the absence of proven preventive strategies, early diagnosis and treatment are therefore crucial to avoid disease related disabilities [1–2].
The WHO recommends laboratory confirmation of at least 70% of clinically suspected BUD cases per country [3]. Application of the 100% M. ulcerans specific diagnostic reference standard for clinical samples, i.e. amplification of the multicopy insertion sequence (IS) 2404 by dry-reagent-based (DRB) PCR, conventional gel-based PCR (cPCR), or quantitative real-time PCR (qPCR) requires fully equipped molecular biology units with highly-skilled personnel and is thus mostly restricted to tertiary (reference) level laboratories or national research centres [4–9]. However, as the major burden of BUD exists in (remote) rural areas of endemic countries and up to one-third of BUD cases are diagnosed in advanced category III stages [10–12], molecular IS2404 detection formats applicable as point-of-care (POC) tests are urgently needed to bring diagnosis closer to where the patients live [13].
Behind this background, an expert group convened by the Foundation for New Innovative Diagnostics (FIND) and the WHO in November 2013 selected loop-mediated isothermal amplification (LAMP) as promising nucleic acid based candidate POC technology applicable for decentralized diagnosis at primary health care level [14].
The salient features of LAMP technology are attributable to the Bacillus stearothermophilus-derived Bst polymerase, which is characterized by strand-displacement activity (without 5’-3’ exonuclease activity), enzyme activity at constant temperature (~ 65 +/- 3°C) without the need of steps for denaturation of double-stranded DNA or primer annealing at different temperatures, high amplification efficiency (up to 1010 copies in 60 minutes) and low susceptibility to classical PCR inhibitors (e.g. melanin, collagen, humic acids). Furthermore, the ability to specifically amplify target sequences by the use of four distinct primers recognizing 6 distinct regions in a single step without the need for sophisticated laboratory equipment made this nucleic acid detection method promising as POC test. LAMP applications were thus established and validated for the diagnosis of various human pathogens such as (protozoan) parasites (e.g. Plasmodium falciparum, Leishmania spp., Trypanosoma brucei, Giardia duodenalis, Schistosoma mansoni/haematobium, Taenia solium), bacteria (e.g. Listeria monocytogenes, Staphylococcus aureus, Mycobacterium tuberculosis) as well as viruses and fungi in settings with limited resources [15–25].
To date, three different LAMP assays for laboratory confirmation of BUD were published. The assay described by de Souza et al. targets the enoyl reductase gene of the M. ulcerans virulence plasmid, technical validation of the assay however was conducted only on a limited number of samples [26]. Njiru et al. and Ablordey et al. reported two LAMP assays amplifying different regions of the IS2404 of M. ulcerans. Both assays underwent validation on various clinical and environmental samples of BUD patients and infected animals from Ghana and Australia, were 100% M. ulcerans specific (without any false positive result) and revealed analytical sensitivities of 20 [27] as well as 30–300 [28] copies of the respective IS2404 target sequence, which equals 0.1 to 1.5 genome equivalents of M. ulcerans, respectively. These analytical sensitivities approach that of cPCR [6, 27–28], but not that of qPCR [7, 29]. However, both assays were evaluated under optimal laboratory conditions applying high-standard DNA extraction and purification procedures in third level laboratories or national research centers, which may not be practicable at primary health care level. To simulate technical feasibility under field conditions, crude (i.e. boiled) DNA extracts were used without further purification for LAMP testing of clinical samples and led to a significant decrease in sensitivity [28]. Moreover, all LAMP assays described so far require unlimited cold-chains as well as shipment of reagents on dry-ice, which is a major cost factor for endemic settings and not always feasible at decentralised facilities. Therefore, technical advancement of LAMP technology and DNA extraction into utterly field friendly formats is unanimously recommended [27–28].
Against this background, the aim of this study was to establish an IS2404 detection based LAMP assay employing lyophilized reagents (dry-reagent-based [DRB] LAMP) which provides significant benefit for application under tropical climate conditions, to validate the assay on clinical samples including fine needle aspirates (FNA) which were largely omitted in previous studies, and to provide a prototype assay for future large-scale field testing.
The study was approved by the Ghanaian KNUST (CHRPE/91/10) and the national Togolese (14/2010/CRBS) ethics committees. All samples analyzed in this study were collected for diagnostic purposes. Written informed consent was obtained from all study participants and/or their legal representative, if aged below 18 years.
Clinically suspected BUD patients were recruited from two study sites in Ghana (Agogo Presbyterian Hospital, Asante Akim North District, n = 12; Tepa Government Hospital, Ahafo Ano North District, n = 20) and one study site in Togo (“Centre Hospitalier Régional de Tsévié”, region “Maritime”, n = 59) and 140 diagnostic samples (FNA, n = 66; swab samples, n = 32; punch biopsy samples, n = 42) were collected according to standardized procedures. Briefly, swabs were taken by circling the undermined edges of ulcerative lesions, and FNA or 3mm punch biopsies were obtained from the center of non-ulcerative lesions. Samples were transported to the Kumasi Center for Collaborative Research in Tropical Medicine (KCCR, Kumasi, Ghana) or the “Institut National d’Hygiène” (INH, Lomé, Togo) in 2 ml screw cap tubes containing 700 μl (swab and punch biopsy samples) or 300 μl (FNA samples) cell lysis solution (CLS; Qiagen, Hilden, Germany) within one day at ambient temperature [10, 30–33].
Clinical, epidemiological and routine laboratory data were collected by means of WHO BU 01.N forms [34] and standardized project specific laboratory data entry forms, and were entered in a web-based database as previously described [10].
Whole genome DNA was extracted from clinical samples in CLS at KCCR or INH using the Gentra Puregene DNA extraction kit (Qiagen) with minor modifications of the manufacturer’s instructions as described in S1 Protocol [6]. DNA extracts were stored at 4–8°C (up to one week) or -18°C (long-term storage).
For routine on-site laboratory confirmation DNA extracts were subjected to IS2404 DRB PCR at KCCR and INH as previously described [6, 10, 35]. For comparative testing in the context of external quality assurance programs with conventional, gel-based IS2404 PCR (cPCR) [4–5, 33, 35, 36] and a recently described modified IS2404 qPCR based on the assay published by Fyfe et al. [7, 29] aliquots of DNA extracts were shipped to the Department of Infectious Diseases and Tropical Medicine (DITM), Munich, Germany by courier service at ambient temperature [10].
The development and validation of the LAMP assay was conducted in the laboratories of DITM.
The study was observational and transversal (cross-sectional study design). An approximative test (McNemar chi-square test for matched pairs of samples with categorical test results) and estimation of standard error of proportion (to calculate 95 percent confidence intervals [95%-CI] of categorical test results) were conducted. Significant differences were defined as P-values below 0.05 or as not overlapping of 95%-CI of proportions.
In silico analysis of the novel IS2404 LAMP primers and testing of 17 DNA extracts from mycobacterial cultures (M. ulcerans, n = 5; other mycobacteria, n = 12) by cLAMP and DRB LAMP revealed 100% specificity of both assays for M. ulcerans. The LODs were 50 (DRB PCR and cPCR), 3 (qPCR) and 100 (cLAMP) copies of the target sequence IS2404, corresponding to 0.2, 0.01, and 0.5 M. ulcerans genome equivalents, respectively.
Out of the 91 patients with suspected BUD, 68 were laboratory confirmed as BUD patients (74.73%) by routine methods. Among 68 confirmed BUD patients, 40 patients (58.82%) were in age group 5–14 years (age range 5–56 years, mean 14 years, median 11 years), 33 patients (48.53%) were male, and 36 patients (52.94%) presented with non-ulcerative lesions.
DRB PCR (on-site at KCCR or INH) and cPCR (DITM, 100% concordance between DRB and cPCR results) confirmed 59/91 (64.84%; 95%-CI: 55.02%-74.65%) of the suspected BUD cases, the qPCR confirmed 68/91 (74.73%; 95%-CI: 65.80%-83.65%), thus added an additional diagnostic yield of 9.89%. The confirmation rate for cLAMP was 62.64% (95%-CI: 52.70%-72.58%; n = 57). Neither DRB PCR nor cPCR or cLAMP had false positive results compared with qPCR, and confirmation rates were not significantly different. According to McNemar test, there was no significant difference between DRB PCR or cPCR (100% concordant results) compared with qPCR (ORcrude = 1.60; 95%-CI: 0.81–3.20; P-value = 0.15), between cLAMP compared with qPCR (ORcrude = 1.76; 95%-CI: 0.89–3.50; P-value = 0.08) and between DRB PCR or cPCR compared with cLAMP (ORcrude = 1.10; 95%-CI: 0.57–2.11; P-value = 0.76).
Among the 68 BUD cases confirmed by qPCR, the sensitivity was 86.76% (95%-CI: 78.71%-94.82%; n = 59) for DRB PCR and cPCR and 83.82% (95%-CI: 75.07%-92.58%; n = 57) for cLAMP. According to McNemar test, there was no significant difference between DRB PCR and cPCR compared with cLAMP (ORcrude = 1.27; 95%-CI: 0.44–3.63; P-value = 0.63).
Among the 140 samples from 91 clinically suspected BUD cases, the positivity rate was 56.43% (95%-CI: 48.21%-64.64%; n = 79) for DRB PCR and cPCR, 67.14% (95%-CI: 59.36%-74.92%; n = 94) for qPCR, and 52.86% (95%-CI: 44.59%-61.13%; n = 74) for cLAMP. Neither DRB PCR nor cPCR or cLAMP revealed false positive results compared with qPCR. According to McNemar test, there was no significant difference between DRB PCR or cPCR compared with qPCR (ORcrude = 1.58; 95%-CI: 0.94–2.64; P-value = 0.07) and between DRB PCR or cPCR compared with cLAMP (ORcrude = 1.16; 95%-CI: 0.70–1.90; P-value = 0.55), whereas the difference between cLAMP compared with qPCR was significant (ORcrude = 1.82; 95%-CI: 1.09–3.05; P-value = 0.02).
Stratification into sample types did not reveal significant differences in positivity rates of DRB PCR, cPCR, qPCR and cLAMP among FNA, swab or punch biopsy samples.
Among the 94 samples from 68 BUD cases confirmed by qPCR, the sensitivity was 84.04% (95%-CI: 76.64%-91.45%; n = 79) for DRB PCR and cPCR, and 78.72% (95%-CI: 44.59%-61.13%; n = 74) for cLAMP. According to McNemar test, there was no significant difference between DRB PCR or cPCR compared with cLAMP (ORcrude = 1.42; 95%-CI: 0.64–3.18); P-value = 0.35).
DRB LAMP revealed the same performance characteristics as determined for cLAMP (i.e. 100% M. ulcerans specificity and a LOD of 0.5 M. ulcerans genome equivalents).
To compare DRB LAMP with DRB PCR, cPCR, qPCR and cLAMP, 32 samples (FNA and swab samples, n = 16, respectively) from 32 suspected BUD cases were subjected to the assays. The positivity rate was 75.0% (95%-CI: 62.41%-87.59%; n = 24) for DRB PCR, cPCR and qPCR, 71.88% (95%-CI: 87.84%-100%; n = 23) for cLAMP, and 68.75% (95%-CI: 80.61%-100%; n = 22) for DRB LAMP. Neither cLAMP nor DRB LAMP revealed false positive results compared with DRB PCR, cPCR and qPCR.
According to McNemar test, there was no significant difference neither between DRB PCR, cPCR or qPCR compared with cLAMP (ORcrude = 1.17; 95%-CI: 0.34–4.10; P-value = 0.78), nor between DRB PCR, cPCR or qPCR compared with DRB LAMP (ORcrude = 1.36; 95%-CI: 0.40–4.49; P-value = 0.58), nor between cLAMP compared with DRB LAMP (ORcrude = 0.86; 95%-CI: 0.26–2.87; P-value = 0.79).
Among the 24 samples from 24 BUD patients confirmed by DRB PCR, cPCR and qPCR, the sensitivity was 95.83% (95%-CI: 87.84%-100%; n = 23) for cLAMP and 91.67% (95%-CI: 80.61%-100%; n = 22) for DRB LAMP. According to McNemar test, there was no significant difference between cLAMP compared with DRB LAMP (ORcrude = 0.48; 95%-CI: 0.02–7.54); P-value = 0.56).
Out of 74 amplicons derived from cLAMP reactions, 74/74 (100%) were judged positive by gel-electrophoresis and 73/74 (98.65%) by SYBR Green I staining. All products derived from DRB LAMP were likewise analyzed and the concordance rate was 22/22 (100%) between both detection methods.
Table 2 shows confirmation rates, sensitivity, specificity and significance of the applied molecular tests.
BUD belongs to the currently five neglected diseases in line for the IDM (innovative and intensified disease management) approach demanding a major scaling up of active detection, treatment, monitoring and surveillance. Development of diagnostic tests that bring health services closer to where NTDs are is considered a research priority. LAMP, a technology that features cost effectiveness, robustness and modest needs in terms of equipment, has recently been selected by the WHO as one of the promising tools for decentralized diagnostics [13–14, 41].
Several investigators recently succeeded in developing M. ulcerans specific LAMP assays which showed performance characteristics comparable to conventional PCR formats [26–28]. Based on longstanding experience with a DRB PCR format for laboratory confirmation of BUD in Ghana and Togo [6, 10, 30, 33, 36, 42], the development of a DRB LAMP assay applicable under tropical climate conditions at primary health care level was envisaged in this study. Thermodynamic reasons (i.e. leaving out an initial denaturation step for annealing of primers) required the design of modified primers, therefore as a first step a new cLAMP assay was established that constituted the basis for the DRB format. During development of the DRB assay lyophilization of the reaction mix initially constituted a major challenge. Due to the glycerol content of Bst polymerase and reaction buffer as employed in previous cLAMP formats (including our own), customized glycerol-free reagents had to be obtained and adequate lyophilization protocols had to be established.
The comparable performance of cLAMP, DRB LAMP and DRB PCR as well as cPCR suggests that both LAMP formats constitute a reliable alternative to conventional routine assays. Our data also show that the LAMP assays and the DRB PCR as well as cPCR have equal sensitivity for FNA samples. Both LAMP formats are applicable at primary health care level, the DRB format however provides significant advantages such as a simplified test layout and the possibility of storage of reagents at ambient temperature. Decentralized utilization of LAMP technology furthermore would lead to cost saving due to reduced expenditures for transportation of samples to a reference center as well as reduced test costs, i.e. US$ 1–2 per LAMP reaction as compared to US$ 8–10 per DRB PCR or cPCR reaction.
In this study it was not possible to assess long-term storage of DRB-LAMP reaction tubes under tropical conditions. Long-term storability of DRB PCR reaction tubes was however previously proven [6, 36] which allows the conclusion that maximum storage periods of up to 12 months also apply for LAMP reagents.
Although in our study routine PCR and LAMP assays for the most part did not perform significantly different from qPCR, it must be assumed that qPCR renders an additional diagnostic yield of approximately 10% [10]. Therefore, regardless of the method used, confirmation of negative samples by qPCR e.g. through the global network of laboratories for confirming M. ulcerans infection [43] should be attempted. Likewise, participation of laboratories in external quality assurance programs as implemented by Eddyani et al. in collaboration with the WHO is strongly recommended [44].
While the amplification procedure of LAMP technology especially in the DRB format can be considered field friendly without restriction, current DNA extraction procedures are not yet entirely appropriate for POC testing and need optimization. As shown by Ablordey et al. the use of boiled crude DNA extracts led to a significant decrease in sensitivity [28]. Other options such as one-tube silica-membrane based extraction protocols [45] or one-tube enzyme-based lyophilized reactions are yet to be evaluated. A field friendly approach to storage of DNA extracts for purposes of quality assurance could be the filter paper technology as successfully applied for TBC [46].
In conclusion, the cLAMP and DRB LAMP formats evaluated in this study are equivalent alternatives to conventional PCR techniques and, provided the availability of field friendly DNA extraction formats, constitute valuable tools for decentralized laboratory confirmation of BUD. As in the case of other investigators who previously developed BUD specific LAMP assays, the validation of the LAMP assays presented in this study was conducted in a third-level laboratory environment, therefore field based evaluation trials are necessary to determine the clinical performance at peripheral health care level.
|
10.1371/journal.pbio.1000153 | Emergence of a Stable Cortical Map for Neuroprosthetic Control | Cortical control of neuroprosthetic devices is known to require neuronal adaptations. It remains unclear whether a stable cortical representation for prosthetic function can be stored and recalled in a manner that mimics our natural recall of motor skills. Especially in light of the mixed evidence for a stationary neuron-behavior relationship in cortical motor areas, understanding this relationship during long-term neuroprosthetic control can elucidate principles of neural plasticity as well as improve prosthetic function. Here, we paired stable recordings from ensembles of primary motor cortex neurons in macaque monkeys with a constant decoder that transforms neural activity to prosthetic movements. Proficient control was closely linked to the emergence of a surprisingly stable pattern of ensemble activity, indicating that the motor cortex can consolidate a neural representation for prosthetic control in the presence of a constant decoder. The importance of such a cortical map was evident in that small perturbations to either the size of the neural ensemble or to the decoder could reversibly disrupt function. Moreover, once a cortical map became consolidated, a second map could be learned and stored. Thus, long-term use of a neuroprosthetic device is associated with the formation of a cortical map for prosthetic function that is stable across time, readily recalled, resistant to interference, and resembles a putative memory engram.
| Brain–machine interfaces (BMIs) have the potential to revolutionize the care of neurologically impaired patients. Numerous studies have now shown the feasibility of direct “brain control” of a neuroprosthetic device, yet it remains unclear whether the neural representation for prosthetic control can become consolidated and remain stable over time. This question is especially intriguing given the evidence demonstrating that the neural representation for natural movements can be unstable: BMIs provide a window into the plasticity of cortical circuits in awake-behaving subjects. Here, we show that long-term neuroprosthetic control leads to the formation of a remarkably stable cortical map. Interestingly, this map has the putative attributes of a memory trace, namely, it is stable across time, readily recalled, and resistant to the storage of a second map. The demonstration of such a cortical map for prosthetic control indicates that neuroprosthetic devices could eventually be controlled through the effortless recall of motor memory in a manner that mimics natural skill acquisition and motor control.
| Research into the development of brain–machine interfaces (BMIs) [1] has flourished in the last decade, with impressive demonstrations of rodents, nonhuman primates, and humans controlling robots or computer cursors in real time [2]–[17]. Studies of closed-loop cortical BMIs have further demonstrated that improvements in performance require learning [3]–[7],[11],[12],[14]–[17]. Basic research into the neural basis of such adaptations has indicated that changes in the directional tuning properties of neurons are associated with the process of learning [5],[6],[14],[17].
However, the neural plasticity and the cortical dynamics associated with long-term BMI use remains unclear. Studies into the neural plasticity associated with BMI use typically incorporated variable ensembles of neurons from day to day [3]–[7],[11],[12],[14]–[17]. In addition, the transform of cortical activity into a prosthetic motor output (i.e., the decoder) was modified at the start of each daily session. Under such conditions, it is likely that novel neural adaptations were required each day to learn the new transform between neural activity and neuroprosthetic control [5],[6],[12],[16]–[18]. Thus, it remains unclear whether a neural representation for prosthetic function can be stabilized and recalled in a manner that mimics our natural ability to recall motor skills.
A better understanding of the cortical dynamics during long-term neuroprosthetic use is important, both from a basic neuroscience point of view as well as from the perspective of neuroprosthetics. Past studies of the neural basis of natural motor control have presented conflicting evidence for a stable neuron-behavior relationship in motor areas [19]–[25]. For example, whereas some studies have found that the neuron-behavior relationship in primary motor cortex (M1) is constant during stereotyped movements [24], others have shown that this relationship can be nonstationary [20],[23]. Specifically, it remains unclear whether the directional tuning properties of M1 neurons are truly stable across time. It would also be valuable to understand these dynamics during long-term neuroprosthetic control. For example, in the scenario that the neural code for prosthetic control is inherently unstable across time, sophisticated adaptive algorithms may be necessary for long-term reliable performance [21],[25].
To fully delineate the ensemble cortical dynamics during the process of learning and reliably using a BMI, we specifically paired a fixed decoder with stable recordings from ensembles of neurons in two macaque monkeys across a period of up to 19 d. The incorporation of a stable ensemble of putative single neurons across days allows us to track specific changes in neural properties over time. Moreover, as we are primarily interested in understanding the long-term neural adaptations to a fixed transform of neural activity into cursor movements, the decoder was held constant over the time period of each experiment. Using such conditions, we demonstrate for the first time, to our knowledge, the long-term reorganization of motor cortex activity associated with daily practice of a center-out task under brain control. We found that the motor cortex is able to form and consolidate an ensemble cortical map for prosthetic control. This neural representation was found to be remarkably stable across time and could be readily recalled at the start of a daily session.
Two macaque monkeys were first trained to manually perform delayed center-out reaching movements using a robotic exoskeleton that limited movements to the horizontal plane (i.e., manual control, MC). This commercially available robotic system allows precise and accurate measurement of kinematic parameters [26]. Following implantation of microelectrode arrays in bilateral primary motor cortex (M1) (128 microelectrodes in each of the two monkeys), each animal was trained to perform the same center-out task in brain control (BC), in which the neural activity directly controlled the position of the cursor (Figure 1A). In each animal, we could record approximately 75–100 well-isolated units during each daily session. However, consistent with reports in the literature [15],[19],[24],[27]–[31], several months postimplantation, a small ensemble of units were found to be extremely stable across a period of days to weeks. Past studies have demonstrated that ensembles of a similar size can be successfully used for two- or three-dimensional control of neuroprosthetic devices [4],[5].
In the specific experiments presented here, we ensured that the ensemble of neurons used for BC were stable over the time frame of the experiment (hereafter referred to as a “stable neural ensemble”). Stability of well-isolated units across days was first assessed by the stationarity and quality of waveforms (Figure 1B). In order to also quantify the stability of waveforms, we compared waveform characteristics across multiple days using principal components analysis (see Figure S1). Recent studies have indicated that this is a valid metric of waveform stability across days [27]–[31]. As an additional measure, we also ensured that the firing statistics (i.e., interspike interval [ISI] distribution) of each putatively stable single unit did not significantly change from day to day [31]. Figure 1C shows three representative ISI distributions for three single units for two separate days. There were no significant changes in the distributions (p>0.05, Kolmogrov-Smirnov Test). Finally, as a measure of ensemble stability across time, we periodically measured the directional tuning of each unit during daily MC sessions. As shown in Figure S2, the ensemble tuning properties were also stable across time.
In this study, we were primarily interested in understanding the long-term neural adaptations to a fixed transform of neural activity into cursor movements (i.e., a fixed decoder across days). As in previous closed-loop BMI studies [4],[6],[11],[14], we used a linear decoder optimized for physical movements of the upper limb. The linear decoder [6],[21],[24] remains a straightforward and transparent method to transform neural activity into a control signal for closed-loop BMI experiments. As shown in Figure 1A, while the animal physically performed center-out movements during MC, the recorded M1 spike activity was regressed against the elbow and shoulder angular positions to generate correlations for each variable. We will use the term decoder to refer to the combined transforms for both shoulder and elbow position. In BC mode (Figure 1A), this decoder allowed neural activity to control the computer cursor. For the initial set of experiments, BC performance was measured in the setting of (1) recordings from a stable ensemble of primary motor cortex (M1) neurons over days, and( 2) a linear decoder that was held constant after training during the MC session on day 1 (hereafter referred to as “fixed decoder”).
Figure 2A quantifies the daily performance of the center-out task in BC for two animals with a fixed decoder. Previous studies have used a variety of tasks to study BC. Because these tasks range from discrete to continuous control, it is difficult to directly compare task performance across studies [3]–[17]. In this study, the cursor was under constant neural control, and the subject was required to perform multiple steps for a correct trial (including initiation by movement to the center followed by a brief hold period). Previous studies suggest that such continuous-control, multistep tasks are significantly more difficult than single-step tasks [6],[12]. Accordingly, longer periods of practice were initially required to learn this multistep task in BC. For the experiments from Monkey “P” and “R” shown in Figure 2, ensembles of 15 units and ten units were used, respectively. For both subjects, with daily practice with a fixed decoder, there was a monotonic increase in BC performance and accuracy (Figure 2A).
As also evident in Figure 2A, there was a similar monotonic decrease in the mean time to reach targets. Whereas the initial cursor trajectories meandered, they became more direct with practice (Figure 2D, comparison of representative trajectories from day 3 and day 13 for Monkey P). It is important to note that the subjects were not required to follow a straight path from the center to each target. Interestingly, the mean trajectory to each target became increasingly stereotyped over time, suggesting that a relatively stable solution emerged for the path to each target. We quantified the similarity between each set of daily mean trajectories by performing pairwise correlations (see Materials and Methods). As illustrated by the color map in Figure 2D, the correlation between the mean paths for each day initially increased and then stabilized. Similar results were obtained for Monkey R (see Figure S3)
We conducted a detailed examination of the performance during each daily session to identify whether BC “skill” could be transferred from one day to the next with practice under these conditions. Past studies have typically presented performance characteristics for an entire session [4]–[7]. As evident in Figure 2B, with practice, subjects could attain accurate performance at the very start of each daily session. Closer examination of the first 5 min of performance each day produced striking evidence of this accuracy at the start of a session (Figure 2C). As expected, there was also a marked reduction in the variability of performance each day under these conditions. Identical levels of performance were also evident in a second animal (Monkey R). Thus, with daily practice in the setting of a stable neural ensemble and a fixed decoder, subjects developed a level of BC skill that could be readily recalled at the start of a session.
We subsequently characterized the changes in M1 neural activity accompanying the sustained improvements in task performance. For the 19-d experiment shown in Figure 2A, a stable level of performance was evident after day 8. We first examined the neuron-behavior relationship during that period (i.e., days 9 through 19) by calculating the directional modulation of neural activity during BC [32]. The directional modulation of neural activity was initially measured with respect to the intended target. Interestingly, we found that a remarkably stable neuron-behavior relationship was associated with proficient task performance. Figure 3A and 3B illustrate the directional modulation of two representative single units during a single BC session. The insets in Figure 3A and 3B illustrate the stability of this directional tuning relationship for BC across a period of 10 d (no significant changes in preferred direction [PD], bootstrap method, false detection rate [FDR] corrected for multiple comparisons). Overall, 14 of the 15 units did not experience a significant change in PD (bootstrap method, FDR corrected for multiple comparisons). We also evaluated whether this was evident at the level of the neural ensemble. As illustrated by the series of color maps in Figure 3C, we again calculated the daily directional tuning relationship for all units within the ensemble during BC. To compare each daily “ensemble map,” we performed pairwise correlations among each daily set of ensemble tuning properties [6]. The similarity among daily ensemble maps initially increased and then stabilized (Figure 3C).
To compare the temporal course of skill acquisition with the process of map stabilization, we calculated a measure of map similarity across days. Thus, for each day, we calculated the mean correlation for comparisons between a given daily map and all other maps (i.e., mean of each column in the right panel of Figure 3C with exclusion of comparison to self). Remarkably, changes in map similarity closely tracked improvements in task performance for both animals (Figure 3D). Thus, stable task performance was strongly associated with the consolidation of an ensemble activation pattern (a “prosthetic motor map”).
We next examined in greater detail the temporal windows during “movement execution.” For instance, cursor control from the center to each target likely has an initial feedforward stage followed by a period in which visual feedback can lead to path corrections. We thus tested whether a similar stable map emerged when only taking into account the initial stages of execution. As shown in Figure 3D (dotted lines), a similar process of map stabilization also occurred for the first 200 ms of neural activity.
We also performed an additional set of analyses to exclude a potential confounder. As evident in Figure 2D, there was considerable variability in the path taken from the center to each of the targets. It is possible that the apparent evolution of ensemble tuning properties reflects changes in the path as opposed to changes in intrinsic neuronal properties. We thus took into account moment-to-moment changes in the cursor trajectories (i.e., 100-ms steps, see Materials and Methods) when calculating the directional modulation of neural activity (Figure 3D). Unlike the previous analysis based on the intended target, this measure accounts for changes in tuning solely resulting from a modified cursor path. This analysis revealed that the tuning properties of neurons evolved during the period of learning independent of any changes in the actual cursor path.
The analysis described above focused on changes in preferred direction during learning and long-term use of a neuroprosthetic device. However, past studies have also indicated that other changes in neural properties can also be present [6],[17]. We thus examined the daily changes in the mean firing rate and the depth of modulation of the neural tuning curves. We first compared the mean changes in firing with practice. For Monkey P, eight of 15 units were found to experience long-term changes in the mean firing rate with practice (p<0.05. t-test comparing days 1–5 with days 15–19, FDR correction for multiple comparisons). Of the eight neurons, seven experienced a net increase, and one demonstrated a slight but significant increase. For Monkey R, six of the ten neurons experienced a significant increase in the mean firing rate with time.
We next evaluated for systematic changes in the depth of modulation associated with long-term neuroprosthetic use. Figure 4A and 4B illustrate representative examples of units with a persistent increase in the depth of modulation (p<0.05. t-test, FDR correction for multiple comparisons). For Monkeys P and R, respectively, seven of 15 and five of ten units demonstrated similar persistent increases in the depth of modulation. The remaining units did not experience significant changes in the depth of modulation. Taken together, these results further highlight the long-term stability of changes in neural properties that tracked improvements in task performance for both animals.
Our results thus far suggest that a stable pattern of neural activity is associated with stable BC performance. We next examined whether the entire ensemble is actually involved in BC. For instance, it is possible that only a small fraction of neurons are actually being used for closed-loop BC. We thus generated an “online” neuron dropping curve to quantify the effects of ensemble size on BC performance. After a session in which BC performance was demonstrably accurate (>95% accuracy), a random number of neurons were excluded during subsequent closed-loop BC. Each of these sessions lasted 10 min. We subsequently confirmed that the level of performance returned to the previous baseline. These experiments were performed for both the ten- and the 15-neuron ensembles. As shown in Figure 5, removal of three neurons (i.e., 20% vs. 30% of neurons, depending on the ensemble size) resulted in a greater than 50% drop in accuracy. Moreover, for correct trials under such conditions, it took significantly longer to reach each target (mean time to target of 2.5 s vs. 5.3 s, p<0.05, t-test). These results indicate that once a neural representation for neuroprosthetic control is consolidated, the entire ensemble map appears to be actively involved in BC.
Our results suggest that an ensemble of motor cortex neurons can settle upon a remarkably stable activation pattern for prosthetic control in response to a constant decoder. We tested the limits of this conclusion by evaluating whether ensembles of neurons can learn an arbitrary, fixed transform. We thus applied a “shuffled” version of a decoder trained during a MC session. In comparison to the reliable predictions of the actual decoder shown in Figure 6A, the “shuffled decoder” could not reliably predict limb position across time as expected (new ensemble in Monkey P, n = 41 neurons). Surprisingly, accurate prosthetic control was achieved after several days of BC practice in the presence of the shuffled decoder (days 3–8: correct trials = 94±1%, mean±standard deviation [SD]; mean time to target = 2.5±0.3 s, mean±SD). Moreover, a stable prosthetic motor map also emerged under these conditions (Figure 6B). In addition to suggesting that a decoder unrelated to arm movements (i.e., a nonbiomimetic decoder) can be learned, this experiment further supports the notion that a stable decoder is crucial for the formation of a stable cortical representation for prosthetic control.
We subsequently tested the specificity of neural adaptations to the initial fixed decoder. Although many options are available to perturb the transform of neural activity to cursor movements [4],[17], we chose to retrain the linear decoder prior to select BC sessions. The linear decoder was created using multivariate linear regression techniques [33]. It is well known that multivariate linear regression can result in variable model parameters when multiple colinearity is present in the dataset [22],[33]. Thus, two models can be equally effective in predicting a parameter but have different model structures. For prediction of movement parameters from neural data, this can result in slightly different decoder structures (i.e., weight given to each neuron) even while the overall movement prediction is stable [22],[33]. Such variability in the weights can occur for sequential datasets from the same recording session [21],[22],[33]. As shown in Figure 7A (upper panel), similar findings were also evident when two decoders were trained on different days. We thus used daily retraining of the decoder as a means to perturb the transform of neural activity to cursor movements.
Interestingly, substitution of the learned decoder (DecoderOLD in Figure 7A, black bar in upper panel) with a newly trained decoder (DecoderNEW, green bar) caused a drop in BC performance. However, the animal could rapidly resume accurate BC upon reinstatement of the well-learned decoder. A significant drop in overall performance was evident for multiple experiments conducted on different days for both animals (Figure 7B). These results suggest that small but significant changes in the model weights are sufficient to prevent an established cortical map from being transformed into a reliable control signal.
We subsequently tested whether a stable prosthetic motor map can emerge in the presence of variability in the decoder. For example, the brain may settle upon a solution that takes into account the inherent variability of the neuron–cursor relationship. We again specifically made use of the variability in the model parameters present with retraining the decoder each day. Under such conditions, more variable daily performance was observed, likely the result of having to relearn the relationship for cursor control each day (see Figure S4A). Moreover, there was no similar trend of cortical map stabilization within the timeframe of the experiment (see Figure S4B). Thus, variability in the decoder impedes the emergence of a stable cortical map for prosthetic control.
The results presented above further indicate that the formation of a stable and readily recalled prosthetic map is closely associated with stable task performance. Once stabilized, is a specific prosthetic motor map resistant to interference from learning a second map? To address this question, we examined whether an animal could simultaneously learn and recall cursor control for two distinct biomimetic decoders using the same set of neurons. As shown by our results, a retrained decoder can prevent accurate transformation of neural activity (Figure 7). We thus allowed a subject to practice BC each day using both a “new” biomimetic decoder and a well-consolidated (“old”) biomimetic decoder (Figure 8A). The new decoder was trained during a MC session on day 1. In comparison to the old decoder, there were significant changes in four of the 15 weights (p<0.05. t-test, FDR correction for multiple comparisons) for the elbow decoder, and seven of the 15 weights for the shoulder decoder (p<0.05. t-test, FDR correction for multiple comparisons).
As expected, introduction of the newly trained decoder reduced task performance (Figure 8A, day 1). Reintroduction of the consolidated decoder, however, rapidly restored BC performance. Over the course of several days, the subject demonstrated skilled performance with each of the two decoders (day 4, 97.5% vs. 99% trials correct, mean time to target of 2.3 vs. 2.4 s). Surprisingly, the prosthetic motor map was distinct for each of the two decoders. Figure 8B shows examples of changes in directional tuning during BC under each condition (insets i and iii). Nine of 15 units exhibited significant changes in directional tuning (bootstrap, p<0.05, FDR corrected). Moreover, although the previously consolidated map remained stable (n = 6 comparisons, R = 0.86±0.03, mean±SD), the new prosthetic motor map was less similar to previous maps (n = 6, R = 0.3±0.05, mean±SD).
As suggested previously, these changes in directional tuning could be the result of a change in the cursor path. As the subjects were not required to reach the targets with a straight path, there was some variation between the cursor paths for trials under each of the two decoders (See Figure S5). We next tested whether changes in the path could account for the observed change in directional tuning. We again computed the directional modulation of neural activity with respect to the actual cursor path during the first 200 ms (as opposed to direction of intended movement to a given target). Using this measurement, the calculated PDs were somewhat different for each neuron (compare tuning curves in Figure 8B with those in Figure S6). This likely reflects the difference between the actual curved paths taken in comparison to an idealized straight path (i.e., directional modulation based on the intended direction). As such, there was a systematic shift in the respective PDs for each neuron (e.g., Figure 8B vs. Figure S6: [i] PDnew decoder = 29° vs. 96°; [ii] PDnew decoder = 352° vs. 74°). Most importantly, even after taking into account the variations in the actual path of the cursor, significant changes in neural tuning were evident during BC with each of the two decoders (see Figure S6).
In summary, this study demonstrates that the motor cortex can form a stable neural representation for neuroprosthetic control. The stability of the emergent cortical map across days is remarkable given that these neurons also participate in the control of natural arm movements for a greater part of the day (in comparison to the approximately 2 h of BC each day). Our results further suggest that the stationarity of this relationship relies upon the constancy of the decoder that transforms neural activity into cursor movements. Interestingly, under such conditions, even nonbiomimetic shuffled decoders allowed the formation of a cortical map that is readily transformed into cursor movements and reliable task performance.
Our analysis of the neural activity during the period of learning indicates that the neuronal tuning functions (i.e., PDs, mean firing rates, and the depth of modulation) appear to undergo a period of modifications after which a stable ensemble activity pattern emerges. These tuning functions were estimated using either the intended direction of motion (i.e., idealized straight path to the target) or the actual motion of the cursor. For natural motor control, different brain areas may represent each of these aspects [34]. During BC, although these two methods can result in different estimations of neuronal tuning properties (depending of cursor path), they provide complementary estimates of the neuron-behavior relationship during prosthetic control [5],[6]. Together, they indicate that a truly stable neuron-behavior relationship emerges with practice.
The stability of neuronal properties at the level of the ensemble further suggests that a functional cell assembly may have formed during the process of learning and continued daily practice [35]. Accordingly, it is possible that systematic alterations in the dynamics of interneuronal correlations also accompany the long-term modifications of single neuronal tuning properties [36].
A topic for future research is the relationship between feedforward “internal models” and feedback during active neuroprosthetic control [37]. The emergence of stable ensemble activity patterns during the early part of each trial (e.g., the first 200 ms of cursor movement) suggests that BC practice leads to the formation of an internal model for cursor control. It is less likely that visual feedback is responsible for shaping these early time periods [16],[38].
A better understanding of these two factors will elucidate principles of trajectory formation during BC. Interestingly, the emergence of stereotyped trajectories that are not necessarily straight is consistent with a recent study suggesting that the process of motor learning balances the acquisition of reward states with the costs of movement [39]. Under this formulation, optimal paths do not necessarily follow a straight trajectory. Consistent with this concept was also the finding that cursor trajectories under each of the two decoders were both curved and somewhat different for each set of trials.
Our findings add to the recent debate on the stability of the neuron-behavioral relationship for both natural motor control and for neuroprosthetic control. Studies have presented conflicting evidence for the notion of a stable neuron-behavior relationship for stereotyped and free-arm movements [19]–[25]. Possible reasons for the apparent variability include the process of learning a novel motor relationship [40]–[42], postural changes, and subtle changes in the pattern of muscle activation [20],[23]. Our experimental setup allowed us to address this in a setting in which the output of an ensemble of neurons can be controlled. Thus, neural activations for a purely disembodied BC task can achieve a stable neuron-behavior relationship after an initial period of instability during learning.
Past studies have presented evidence of long-term improvements in neuroprosthetic control with practice [3]–[7],[11],[12],[14]–[17]. As indicated by our results, however, there are at least two distinct mechanisms for such long-term improvements in performance. There are improvements in learning as a result of the formation and consolidation of a neural representation for prosthetic control. Alternatively, long-term improvements in performance can be the result of daily relearning and the formation of a novel neural representation. Our experiments indicate that incorporation of stable neural ensembles and a fixed transform of neural activity allows for monotonic and reliable improvement in performance. That consolidation of a cortical representation was important for these improvements is suggested by (1) evidence for rapid recall of performance at the start of each daily session, and (2) stabilization of neural tuning functions.
Our primary interest in this study was to characterize the long-term dynamics of the neuron-behavior relationship for direct cortical control of a cursor. This was best achieved by applying a constant decoder across time while observing the changes in neural activity. Although many decoders are likely to be useful for this purpose, the linear decoder has proven to be effective and offers a ready comparison to past successful BMI studies [4]–[7]. Moreover, our results suggest that cortical map formation can be truly independent of the exact decoder used (e.g., Figure 5 shows learning across days with a shuffled decoder).
One implication of our findings is that cortical control of a prosthetic device depends on specific neural adaptations to the applied decoder [1],[43]. Whereas two decoders may both predict MC movement parameters equally well, there may be significant variability in the parameters assigned to a specific neuron [33]. As shown by our results, this variability prevents the formation of a stable neural representation. Minimizing decoder variability would be less important if an entirely new set of neurons are recorded each day. However, in the more likely scenario where subsets of neurons are stable across time [15],[19],[24],[27]–[31], it will be important to consider parameters assigned to stable units. Taking into account such information could allow “graceful degradation” of function, where the loss of a subset of units would not be catastrophic. This may also minimize the extent of required relearning with changes in the recorded ensemble.
Our results further indicate that the formation and stabilization of a cortical map for prosthetic function is closely linked to the process of long-term neuroprosthetic skill acquisition. Strikingly, the features of this map (i.e., readily recalled, stable, and resistant to interference) resemble properties often attributed to a putative long-term memory engram [44]. It is easy to imagine that in real-world situations, complicated neuroprosthetic control will require consolidation of an analogous “prosthetic motor memory” for long-term retention of skilled function [45]. With continued improvements in technology [46],[47], neuroprosthetic devices could be controlled through effortless recall of such a motor memory in a manner that mimics the natural process of skill acquisition and motor control.
Two adult male rhesus monkeys (Macaca mulatta) were chronically implanted in the brain with arrays of 64 Teflon-coated tungsten microelectrodes (35 µm in diameter, 500-µm separation between microwires) in an 8×8 array configuration (CD Neural Engineering). Monkey P was implanted in the arm area of primary motor cortex (M1) and the arm area of dorsal premotor cortex (PMd), both in the left hemisphere, and the arm area of M1 of the right hemisphere, with a total number of 192 microwires across three implants. Monkey R was implanted bilaterally in the arm area of M1 and PMd (256 microwires across four implants). Localization of target areas was performed using stereotactic coordinates from a neuroanatomical atlas of the rhesus brain [48]. Implants were targeted for pyramidal tract neurons in layer 5, and were typically positioned at a depth of 3 mm in M1 and 2.5 mm in PMd. Depth of electrode placement was guided by intraoperative monitoring of spike activity. All procedures were conducted in compliance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the University of California at Berkeley Institutional Animal Care and Use Committee.
Unit activity was recorded using the MAP system (Plexon). For this study, only units from primary motor cortex were used. Only single units that had a clearly identified waveform with a signal-to-noise ratio of at least 4∶1 were used. Activity was sorted prior to recording sessions using an on-line spike-sorting application (Sort-Client; Plexon). Large populations of well-isolated units (∼75–100) were recorded during each daily session in both monkeys (typical number of units was defined by waveform quality and ISI distributions). Consistent with reports in the literature [24],[27]–[31], several months postsurgery, we found a subset of stable units whose waveform shape, amplitude, and relationship to other units on a channel varied little from day to day (i.e., the sorting template in the Sort-Client required no or very minor daily modifications). The stationarity of such properties was the first criterion for a putative stable unit. We also examined the properties of the ISI distribution and the presence of an absolute refractory period to confirm the presence of a stable single unit. We also confirmed the stability of the waveforms using commercially available software (Wavetracker; Plexon). Specifically, we utilized the features that allow mapping of waveform characteristics into a two- and three-dimensional principal components space. Stability of waveforms could be assessed by comparison of the stability of the projections across time (please see Figure S1 for examples). Multivariate ANOVA tests allowed statistical comparison.
Moreover, we also estimated the PD in MC of select ensembles of putative stable units. For these subsets of ensembles, MC sessions were performed each day to estimate the directional tuning curves (e.g., Figure S2 shows the similarity of the tuning curves within an ensemble across days). Moreover, the precise number of units per experiment was determined by examining all recorded units over a period of several days to ascertain units with stationary properties. Our conclusions did not appear to depend on ensemble size.
Monkeys were trained to perform a center-out delayed reaching task using a Kinarm (BKIN Technologies) exoskeleton. In this device, the shoulder and elbow are restricted to move in the horizontal plane, giving two degrees of freedom (flexion/extension). During training and recording, animals sat in a primate chair that permits limb movements and postural adjustments. Head restraint consisted of the animal's headpost fixated to a primate chair. Recording sessions typically lasted 2–3 h per day. Kinematic variables (position, velocity, and acceleration) were continuously monitored and recorded.
The behavioral task consisted of hand movements from a center target to one of eight peripheral targets (i.e., center-out task) distributed over a 14-cm diameter circle. Target radius was typically 0.75 cm. Trials were initiated by entering the center target and holding for a variable time period of 500–1,000 ms. The “GO” cue (center changed color) was provided after the hold period. A liquid reward was provided after a successful reach to each target and a peripheral hold period (200–500 ms). Visual feedback of hand position was provided by a cursor precisely colocated with the center of the hand (cursor radius = 0.5 cm). During the task, the nontask arm was immobilized in a padded splint.
In BC, the cursor was continuously controlled by neural activity, and each animal received visual feedback of cursor movements. The task-related hand (right) was removed from the exoskeleton and restrained during BC. The cursor was under continuous volitional control throughout the experiment. The subjects were required to self-initiate each trial by bringing the cursor to the center. As mentioned below, the hold period for BC was optimized in order to minimize false-positive activations.
Typical BC trials required a fixed center hold period of 250–300 ms. As in other studies [6], subjects experienced difficulty completely stopping the cursor. During typical hold periods, the cursor slowed down enough to trigger the GO cue. However, with practice (e.g., after >6–7 d for a given set of neurons and a fixed decoder), animals could perform tasks that required longer hold periods (e.g., 1,000 ms) as well as variable hold periods. During these trials, the cursor appeared to be actively held in place. Moreover, reward was provided when the cursor was inside of the peripheral target for >100 ms. Typically, a reduction in velocity was sufficient to accomplish this.
A trial was considered incorrect if the cursor failed to reach the target within 10 s after a GO cue. During selected sessions, we concurrently performed video and surface electromyelogram (EMG) recordings from proximal muscle groups. As in past studies, neither animal moved their upper extremity during BC [5],[6]. The observation that movement was not critical for BMI performance is further highlighted by the fact that a shuffled decoder with no relation to actual movements could be learned.
During experiments in which new decoders were introduced (e.g., see Figure 7), no cues were given. These blocks occurred in a randomized, unpredictable manner. Moreover, these trials were brief (∼20 min). However, for experiments in which two decoders needed to be learned, two different color-coding schemes were used to indicate differences between BC sessions involving the two decoders (e.g., data shown in Figure 8). For these experiments, the color of peripheral targets was different for trials using either the old decoder (blue) or the new decoder (yellow). The respective color schemes for the center target (green) and the GO cue (change from green to red) remained constant. In experiments requiring relearning of a daily decoder (i.e., data shown in Figure 4), animals were given longer sessions (1–2 h) in order to adapt to the changes.
Finally, there was evidence of generalization of prosthetic control beyond the stereotyped structure of the center-out task. In selected experimental blocks, animals were able to generate novel cursor trajectories in order to reach the targets (see Figure S7).
Previous analyses [2],[6],[21] have demonstrated that hand position and velocity can be accurately predicted with a linear regression model. In this model (Equation 1), the inputs, X(t), were a matrix with each column corresponding to the discharges of individual neurons, and each row representing one time bin. The output Y(t), was a matrix with one column per motor parameter. The linear relationship between neuronal discharges in X(t), and behavior (elbow and shoulder joint positions) in Y(t) was expressed as(1)where a and b are constants, calculated to fit the model optimally. First, a(u) are the impulse response functions required for fitting X(t) to Y(t) as a function of time lag u between the inputs and the outputs. Ten time lags were used during these experiments. Second, b represents the Y-intercept in the regression. The final term in the equation, ε(t), represents residual errors. The linear filter was generated using the techniques described above and neural (spike activity from a select group of neurons binned into 100-ms bins) and kinematic data (continuous recordings of the elbow flexion/extension and shoulder flexion/extension angles) recorded from a 10-min session of MC (while performing the center-out task). Past studies have shown that a bin size of 100 ms is optimal [2],[6],[21]. A new decoder was trained by repeating the algorithm outlined above during a MC session on subsequent day.
|
10.1371/journal.pbio.2005189 | Tissue-specific degradation of essential centrosome components reveals distinct microtubule populations at microtubule organizing centers | Non-centrosomal microtubule organizing centers (ncMTOCs) are found in most differentiated cells, but how these structures regulate microtubule organization and dynamics is largely unknown. We optimized a tissue-specific degradation system to test the role of the essential centrosomal microtubule nucleators γ-tubulin ring complex (γ-TuRC) and AIR-1/Aurora A at the apical ncMTOC, where they both localize in Caenorhabditis elegans embryonic intestinal epithelial cells. As at the centrosome, the core γ-TuRC component GIP-1/GCP3 is required to recruit other γ-TuRC components to the apical ncMTOC, including MZT-1/MZT1, characterized here for the first time in animal development. In contrast, AIR-1 and MZT-1 were specifically required to recruit γ-TuRC to the centrosome, but not to centrioles or to the apical ncMTOC. Surprisingly, microtubules remain robustly organized at the apical ncMTOC upon γ-TuRC and AIR-1 co-depletion, and upon depletion of other known microtubule regulators, including TPXL-1/TPX2, ZYG-9/ch-TOG, PTRN-1/CAMSAP, and NOCA-1/Ninein. However, loss of GIP-1 removed a subset of dynamic EBP-2/EB1–marked microtubules, and the remaining dynamic microtubules grew faster. Together, these results suggest that different microtubule organizing centers (MTOCs) use discrete proteins for their function, and that the apical ncMTOC is composed of distinct populations of γ-TuRC-dependent and -independent microtubules that compete for a limited pool of resources.
| Eukaryotic cells require specific arrangements of microtubules to carry out diverse functions, including cell division and intracellular transport. In dividing animal cells, microtubules are arranged radially around two centrosomes, promoting the correct distribution of DNA into daughter cells. As cells differentiate, microtubules become organized at non-centrosomal sites, often yielding decentralized microtubule arrays. Although a large body of work has focused on understanding how microtubules are grown from and organized by the centrosome, very little is known about how these activities are performed at non-centrosomal sites. We optimized a technique to degrade proteins in C. elegans differentiated cells, specifically testing the role of essential centrosome proteins in building non-centrosomal microtubules, which emanate from the apical membrane in embryonic intestinal epithelial cells. Surprisingly, we found fundamental differences between the centrosome and membrane, both in the mechanisms that recruit microtubule regulators to each site, and in the proteins that are required to build and organize microtubules. In addition, we found that when fewer microtubules grew from membranes, the remaining microtubules grew faster, suggesting competition for limited materials. Together, our study highlights differences in how centrosomes and membranes grow and organize microtubules, and that multiple pathways contribute membrane-organized microtubules.
| Described nearly 50 years ago, microtubule organizing centers (MTOCs) generate specific spatial patterns of microtubules as needed for cell function [1]. The best-studied MTOC is the centrosome, a non-membrane bound organelle that organizes microtubules into a radial array from its pericentriolar material (PCM) or from subdistal appendages attached to the mother centriole. However, in many types of differentiated cells, microtubules are organized at non-centrosomal sites to accommodate diverse cell functions. In animal cells, these non-centrosomal MTOCs (ncMTOCs) can be found in the axons and dendrites of neurons, around the nuclear envelope of skeletal muscle cells, at the apical surface of epithelial cells, and at the Golgi complex [2–8]. ncMTOCs can promote nonradial arrangements of microtubules, such as the linear arrays of microtubules present along the apicobasal axis in epithelial cells. How ncMTOCs are established and whether they are composed of the same proteins that impart MTOC activity at the centrosome are largely unknown.
In general, MTOCs can be defined as cellular sites that nucleate, anchor, and stabilize microtubules; however, the molecular basis for these functions has been elusive [9,10]. Because of the inherent structural and chemical polarity of microtubule polymers, microtubules are nucleated and anchored at their minus ends. Thus, a defining feature of MTOCs is that they interact with microtubule minus ends. The first microtubule minus-end protein described was γ-tubulin, which, together with GCP2 and GCP3, forms the γ-tubulin small complex (γ-TuSC) [11]. In some organisms, additional γ-tubulin complex proteins (GCPs) combine with the γ-TuSC to form the larger γ-tubulin ring complex (γ-TuRC); organisms lacking these additional GCPs are thought to oligomerize γ-TuSCs into similar ring complexes [12,13]. To date, only the γ-TuSC components TBG-1/γ-tubulin, GIP-1/GCP3, and GIP-2/GCP2 have been identified in C. elegans, suggesting that C. elegans γ-TuRC may share the yeast γ-TuRC composition. As our experiments do not distinguish between γ-TuSC and γ-TuRC, we will use the term γ-TuRC for simplicity. A putative C. elegans ortholog of mitotic spindle-organizing protein associated with a ring of γ-tubulin (MOZART1), MZT1, a γ-TuRC-interacting protein and proposed γ-TuRC component, was identified based on sequence homology to the Arabidopsis ortholog, but its function has not been investigated [14]. γ-TuRC has microtubule nucleation capacity and can also cap the minus ends of microtubules, preventing minus-end growth or depolymerization [15,16]. Whether γ-TuRC predominantly functions as a nucleator or as a minus-end cap or anchor in vivo is a matter of debate.
Although γ-TuRC is essential in organisms ranging from yeast to humans, γ-TuRC depletion does not result in the elimination of all microtubules from the cell, suggesting that other mechanisms exist to grow and anchor microtubules at MTOCs. γ-TuRC removal in vivo has severely deleterious effects on mitosis, but microtubules are still present [17–19]. The presence of microtubules in dividing C. elegans embryonic cells appears to rely on both γ-TuRC function and the mitotic kinase AIR-1/Aurora A [20,21], as only depletion of both TBG-1 and AIR-1 from dividing cells results in the elimination of centrosomal microtubules. Whether γ-TuRC and AIR-1, or other essential centrosomal MTOC proteins, function redundantly to build microtubules at ncMTOCs in animal cells is unknown, most notably because of the early requirement of these proteins in mitosis that prohibits an assessment of any later roles during differentiation.
In C. elegans embryonic intestinal cells, MTOC function is reassigned from the centrosome during mitosis to the apical surface as cells begin to polarize [6], thereby establishing an apical ncMTOC in each cell. Intestinal cells all derive from the “E” blastomere, undergoing four rounds of division before polarizing at the “E16” stage, when the intestinal primordium is comprised of 16 epithelial cells (S1A Fig, [22]). Shortly after the E8–E16 division, the E16 cells follow a stereotypical pattern of polarization and establish their apical surfaces facing a common midline [6], the eventual site of the lumen of the epithelial tube. ncMTOCs positioned along these apical surfaces nucleate and organize microtubules into fountain-like arrays emanating away from the midline on either side [6,23,24]. Intriguingly, many centrosomal MTOC proteins, including AIR-1 and GIP-1, also localize to this ncMTOC [6]. Because the earliest stages of ncMTOC formation can be easily tracked, the embryonic intestinal primordium provides an ideal system in which to test the role of specific proteins in ncMTOC establishment in vivo.
Here, we test the hypothesis that the proteins required to build microtubules at the centrosome play a similar role at ncMTOCs. To do this, we optimized an existing tissue-specific degradation strategy to test the role of GIP-1 and AIR-1 at ncMTOCs in C. elegans embryonic intestinal cells. As at the centrosome, we find that GIP-1 is required to localize the other γ-TuRC members, TBG-1 and GIP-2. Additionally, we show that a predicted ortholog of the γ-TuRC protein MZT1 is essential in C. elegans and colocalizes with γ-TuRC in all contexts but is uniquely required for localization of γ-TuRC to the PCM, and not to the centriole or to the apical ncMTOC. This differential requirement for proteins at the centrosome versus the apical ncMTOC was a common trend, as AIR-1 was also only required to localize GIP-1 and TAC-1 to the PCM, but not to the apical ncMTOC. In addition to GIP-1 and AIR-1, we assessed the requirement of other known microtubule regulators, including ZYG-9/chTOG, PTRN-1/CAMSAP, NOCA-1/Ninein, and TPXL-1/TPX2. Surprisingly, we found that, overall, the depletion of these proteins did not disrupt apical microtubule organization. Furthermore, removal of GIP-1 had only a minor effect on microtubule dynamics at the apical ncMTOC; a subset of microtubules was perturbed, as indicated by a change in EBP-2/EB1 localization and dynamics. These results highlight the differences between the centrosome and other MTOCs and suggest that ncMTOCs are composed of at least two populations of microtubules, γ-TuRC-dependent and γ-TuRC-independent.
To test the role of γ-TuRC and AIR-1 at ncMTOCs, we needed a strategy to deplete proteins essential in the early embryo (early essential proteins) at later stages of development. For example, we had previously been unable to assess the function of γ-TuRC components in differentiated cells in vivo, as their depletion causes severe mitotic defects that result in early embryonic lethality. Tissue-specific degradation strategies have provided a means to deplete such early essential proteins [25,26]. We therefore optimized an existing tissue-specific degradation system (Fig 1A and 1B). The germline cell fate determinant PIE-1 is degraded in somatic cells in the early embryo ([27], Fig 1C). This degradation requires a zinc finger domain 1 (ZF) on PIE-1 and the SOCS-box protein ZIF-1, which targets PIE-1 for degradation by an E3 ubiquitin ligase [27]. Previous reports found that the ZF domain could be added to any protein of interest and that protein is degraded in somatic cells by ZIF-1 ([28], Fig 1E and 1G). However, endogenous ZIF-1 is only expressed in the early embryo, so degradation of targets later in development requires exogenous expression of ZIF-1 [29]. The major drawback of this system is that degradation of early essential proteins by endogenous ZIF-1 leads to an early arrest. We found that a zif-1 deletion mutant is viable (92% ± 8.4% embryonic viability in zif-1(gk117) worms compared to 99% ± 1.9% in N2 worms) despite the apparent loss of ZIF-1 activity (Fig 1C and 1D). Using a zif-1 mutant background (“zif-1(−)”), we can tag any gene with the ZF domain using CRISPR/Cas9 and the resulting ZF-tagged protein is not degraded (Fig 1F, 1H, 1I, 1L and 1O), thus allowing for normal development. We then express ZIF-1 under the control of a tissue-specific promoter to degrade ZF-tagged targets.
Using this strategy, we tagged the γ-TuRC component GIP-1/GCP3, the predicted MZT1 ortholog W03G9.8, which we hereafter refer to as MZT-1 (see below), and the mitotic kinase AIR-1/Aurora A with ZF::GFP, allowing us to monitor protein expression and localization and to degrade each protein with exogenous ZIF-1 expression (Fig 1I–1Q). As expected, GIP-1 localized to the apical ncMTOC in intestinal cells and AIR-1 decorated microtubules (Fig 1I and 1O, [6]). ZIF-1 was then expressed using the promoter for the elt-2 gene, which is expressed exclusively in the intestine starting at intestinal stage E2 (S1A Fig). ZIF-1 expression led to intestine-specific removal of GIP-1, MZT-1, or AIR-1 (“GIP-1gut(−),” “MZT-1gut(−),” “AIR-1gut(−)”), as demonstrated by the loss of both apical and cytoplasmic GFP signal (Fig 1I–1Q, S1B–S1E Fig, S2A–S2D’ Fig). We quantified this intestine-specific depletion in two ways. First, we measured the total amount of reduction of GFP signal in the intestinal primordium of “gut(−)” embryos as compared to “gut(+)” siblings that lacked the ZIF-1-expressing array (percent GFP depletion: GIP-1gut(−), 93.1%; MZT-1gut(−), 92.1%; AIR-1gut(−), 82.1%; S3A Fig). This is likely an underestimate that is due to the out-of-focus light contributed by non-degraded ZF::GFP in non-intestinal cells that complicates this assessment, especially when highly expressed genes like air-1 were tagged. We also took line scans across the midline of the intestinal primordium in both gut(−) and gut(+) embryos to measure apical enrichment (Fig 1K, 1N and 1Q, S3B and S3D Fig). We see no significant apical enrichment of GFP signal in GIP-1gut(−), MZT-1gut(−), or AIR-1gut(−) embryos and a significant reduction in both apical and cytoplasmic GFP intensity, as compared to gut(+) control embryos. This degradation strategy can also be used to co-deplete AIR-1 and GIP-1 ([GIP-1;AIR-1]gut(−), S1D Fig, S2A–S2D’ Fig, Materials and methods). Thus, the zif-1 mutant coupled with the ZIF-1/ZF degradation system provides an effective tool for depleting early essential proteins in a tissue-specific manner.
To test the role of γ-TuRC and AIR-1 in establishing the apical ncMTOC, we needed to effectively remove these proteins from intestinal cells prior to the E16 stage, when cells reassign MTOC function to the apical membrane. Intestinal differentiation in C. elegans proceeds in the absence of cell division [30], suggesting that mitotic defects in intestinal cells per se would not affect their ability to build an ncMTOC. Thus, we could begin degradation of the desired targets during the intestinal divisions to ensure they would be effectively cleared by the time cells began to polarize and establish the apical ncMTOC.
In C. elegans, loss of maternal AIR-1 results in severe mitotic defects, including multinucleate cells, polyploidy, disorganized microtubules, and failed centrosome separation [31,32]; the additional removal of γ-TuRC components results in monopolar spindles and loss of centrosomal microtubules in the first cell division [20,21]. We thus used mitotic defects in intestinal cells as a phenotypic readout for effective degradation of GIP-1 and AIR-1 prior to polarization. The E blastomere undergoes four rounds of division to generate the polarized 16-cell primordium (E16, Fig 2A). As ZIF-1 was expressed from an early promoter (elt-2p) that is active beginning around E2–E4, we expected that successful removal of γ-TuRC and AIR-1 should result in polarized intestinal primordia with between 2 and 16 cells. Indeed, we found that degradation of GIP-1, MZT-1, or AIR-1 resulted in embryos with 8.6 ± 2.3, 7.6 ± 1.8, or 9.2 ± 1.5 intestinal nuclei, respectively, and that degradation of both AIR-1 and GIP-1 resulted in embryos with 4.0 ± 0.0 intestinal nuclei (Fig 2A and 2B). Expression of ZIF-1 from a promoter that is active around E8 (ifb-2p) also significantly reduced the number of intestinal nuclei, although to a lesser extent (Fig 2A and 2B, S1A Fig). As further proof that we were effectively depleting the desired targets, we found that embryos with decreased or no zygotic air-1, and with only a maternal supply of ZF::GFP-tagged AIR-1 (“AIR-1*”), had intestinal nuclear numbers indistinguishable from AIR-1gut(−) embryos (see Materials and methods, 9.1 ± 2.2, Fig 2B). We frequently observed mitotic defects in AIR-1gut(−) and GIP-1gut(−) embryos, such as scattered condensed chromosomes, binucleate cells, and abnormal mitotic spindles (S2I Fig, S1 Movie). These results are consistent with the reported role for γ-TuRC and AIR-1 in mitosis and suggest that GIP-1 and AIR-1 are effectively depleted from intestinal cells beginning at approximately E4.
We next tested whether intestinal cells can polarize and differentiate in the absence of GIP-1 or AIR-1. In intestinal cells depleted of both GIP-1 and AIR-1, the apical polarity protein PAR-3, whose localization to the apical surface is a hallmark of apicobasal polarity [33], was unperturbed (S2E–S2H Fig). Similarly, we observed intestine-specific lysosome-like organelles known as gut granules (Fig 3D), which are hallmarks of intestinal differentiation [23,30]. Together, these results confirmed that we could use tissue-specific degradation to deplete GIP-1 and AIR-1 prior to apical ncMTOC formation without any dramatic effects on intestinal differentiation.
MZT proteins have been shown to recruit γ-TuRC to spindle poles in plants, fungi, and human cell culture lines but have not been characterized in vivo in animal cells [14,34–38]. A previous study of the Arabidopsis homolog of the small protein MZT1 (GIP1) identified the uncharacterized gene W03G9.8 (hereafter called mzt-1) by sequence homology as a potential MZT1 ortholog in C. elegans (Fig 3A, [14]). We tagged the endogenous locus of mzt-1 with ZF::GFP to monitor endogenous MZT-1 localization and assess its function. In all embryonic, larval, and adult tissues examined, we observed consistent colocalization of ZF::GFP::MZT-1 with tagRFP::GIP-1, including at centrosomes in the early embryo, the apical membrane of the intestinal and pharynx primordia, the cell junctions of seam cells, and the plasma membrane and centrioles of the germline (Fig 3B). In the intestinal primordium, MZT-1 localized to centrioles and the apical ncMTOC in an identical pattern to all three γ-TuRC components GIP-1/GCP3, GIP-2/GCP2, and TBG-1/γ-tubulin (Fig 3C, 3F, 3I and 3L). Like other γ-TuRC members tbg-1 and gip-1 [17,39], we find that mzt-1 is also required for embryonic viability. In a zif-1(+) background, ZF::GFP::MZT-1 is degraded ubiquitously in somatic cells by endogenous ZIF-1, and we observed fully penetrant maternal effect lethality (0/420 embryos hatched). By contrast, in control embryos (zif-1(+); GFP::MZT-1), endogenous GFP::MZT-1 is not degraded by ZIF-1 and the embryos are viable (340/344 hatched and grew to adulthood). Based on sequence homology, colocalization with GIP-1, and embryonic lethality, our results suggest that W03G9.8/mzt-1 is the C. elegans ortholog of MZT1.
Our ability to deplete GIP-1 (Fig 3G) and MZT-1 (Fig 3E) in intestinal cells afforded us the opportunity to test their role in the recruitment of each other and of other γ-TuRC components to the apical ncMTOC. At the centrosome, γ-TuRC components exhibit interdependent localization [17]. We see a similar requirement for GIP-1 in localizing MZT-1, GIP-2, and TBG-1 to the apical membrane (compare Fig 3C, 3I, 3L to 3D, 3J and 3M), suggesting that, as at the centrosome, these proteins localize to the apical ncMTOC as a complex. Additionally, the disrupted localization of other γ-TuRC components upon GIP-1 depletion further confirmed that we had successfully perturbed GIP-1 function. In contrast to GIP-1gut(−) embryos, GIP-1, GIP-2, and TBG-1 all localized apically in MZT-1gut(−) embryos (Fig 3H, 3K and 3N), suggesting that MZT-1 is not required to recruit γ-TuRC to the apical ncMTOC.
The surprising finding that γ-TuRC still localizes to the apical ncMTOC upon MZT-1 depletion led us to further investigate the role of MZT-1 in γ-TuRC recruitment to different sites. γ-TuRC localizes to three distinct locations around the time of polarization. During the divisions that precede intestinal polarization, γ-TuRC accumulation at the PCM is coupled with the cell cycle. GIP-1 is recruited to the PCM of the centrosome as cells enter mitosis (Fig 4A). Following mitosis, the centrosome migrates to the lateral membrane and the PCM is greatly reduced. In this lateral configuration, we frequently observe two γ-TuRC positive puncta per cell, which, by electron microscopy, are centrioles associated with a small shell of PCM (“paired centrosomes” [6], Fig 4B, “interphase” inset, blue double arrows). The paired centrosomes then migrate to the apical surface, where they appear as naked centrioles that completely lack PCM by electron microscopy [6]. At this stage, GIP-1 localizes to both the centrioles and the apical ncMTOC (Fig 4C). By contrast, MZT-1gut(−) embryos localized GIP-1 to their centrioles but failed to recruit GIP-1 to the PCM during mitosis (compare Fig 4A to 4D). We frequently saw inappropriate numbers of centrioles in MZT-1gut(−) embryos (number of GIP-1 positive foci in control: 1.8 ± 0.5, n = 64 cells; MZT-1gut(−): 3.0 ± 1.0, n = 134 cells; two-tailed t test: p = 4.4 × 10−24), which may be an indirect consequence of earlier mitotic defects and which appeared to result in multipolar spindles (Fig 4D). Following mitosis in MZT-1gut(−) embryos, GIP-1 remained associated with centrioles and accumulated at the apical ncMTOC, as in control embryos (Figs 3H, 4B, 4C, 4E and 4F). To our knowledge, these results provide the first in vivo role for MZT-1/MZT1 in animal cells, suggesting that MZT-1 is a PCM-specific linker for γ-TuRC in dividing cells. Consistent with these findings, human MZT1 appears to promote the targeting and activation of an intact γ-TuRC to the centrosome in human tissue culture cells [38]. In C. elegans, MZT-1 localization tracks with γ-TuRC localization at the centriole, the PCM, and the apical ncMTOC. However, MZT-1 is not required to target γ-TuRC to the apical ncMTOC but requires GIP-1 to localize there, suggesting that MZT-1 is stably associated with the complex even when not playing a targeting role.
Many microtubule minus-end proteins are found at MTOCs, reflecting the defining ability of MTOCs to nucleate and stabilize microtubule minus ends. To date, only a handful of these proteins have been identified [5], including γ-TuRC and the microtubule-stabilizing protein PTRN-1/Patronin/CAMSAP [11,15,40–43]. Additionally, NOCA-1/Ninein is often found to colocalize with the minus ends of microtubules, although it has never been shown to directly bind to minus ends [44]. A previous report found that γ-TuRC and NOCA-1 function together in parallel with PTRN-1 to maintain non-centrosomal microtubule arrays in C. elegans larval and adult skin [45]. Furthermore, the NOCA-1 h-isoform appears to localize to the membrane ncMTOC in the C. elegans adult germline using a palmitoylation site [45]. In the absence of this site, γ-TuRC is required to target NOCA-1 to the ncMTOC. We found that both PTRN-1 and NOCA-1 localize to the apical ncMTOC in wild-type embryonic intestinal cells (S4A, S4B and S4E Fig). We therefore tested the roles of GIP-1 and MZT-1 in the apical localization of PTRN-1 and NOCA-1. PTRN-1 and NOCA-1 (d- and e-isoforms) appeared to localize normally in GIP-1gut(−) or MZT-1gut(−) embryos (S4C, S4D, S4F and S4G Fig). The NOCA-1 d- and e-isoforms lack the NOCA-1h region containing the characterized palmitoylation site; however, we did not rule out the use of alternative palmitoylation sites, which the d-isoform is predicted to contain [46].
AIR-1/Aurora A is a mitotic kinase that helps activate MTOC function at the centrosome, in part by driving PCM accumulation of targets required for microtubule nucleation and polymerization, such as γ-TuRC and TAC-1/TACC3 [32,47–50]. The phosphorylated, kinase-active form of AIR-1 localizes to the apical ncMTOC along with GIP-1 and TAC-1 [6], suggesting that AIR-1 could similarly regulate the accumulation of these targets at the apical ncMTOC. We first asked whether AIR-1 is required for normal GIP-1 and TAC-1 accumulation at the mitotic centrosome of dividing E8 cells; as predicted from previous studies in other cell types and organisms, average GFP::GIP-1 centrosomal fluorescence was significantly reduced in AIR-1gut(−) cells compared with control cells (532.7 ± 172.8 versus 1,194.6 ± 329.9, p = 1.35 × 10−7, Fig 5A–5C, see Materials and methods). We note that centriolar GIP-1 signal appeared to remain upon AIR-1 depletion (Fig 5B). We observed a similar reduction in GFP::TAC-1 levels at the mitotic centrosome of AIR-1gut(−) cells compared to control cells (501.9 ± 91.7 versus 972.0 ± 463.6, p = 2.88 × 10−6, Fig 5D–5F), suggesting that AIR-1 can be efficiently depleted prior to ncMTOC formation. In contrast to the E8 mitotic centrosomes, GFP::GIP-1 and GFP::TAC-1 were still recruited to the apical ncMTOC in control and AIR-1gut(−) embryos (Fig 5G–5J). These results indicate that AIR-1 is required for GIP-1 and TAC-1 recruitment to intestinal mitotic centrosomes, as is known in other cell types and organisms [48,51], but that AIR-1 is not required for their localization to the apical ncMTOC, further distinguishing the centrosome from the apical ncMTOC.
γ-TuRC and AIR-1 are required to nucleate microtubules at the centrosome in C. elegans [20]. We used GIP-1gut(−), MZT-1gut(−), AIR-1gut(−), and [GIP-1; AIR-1]gut(−) embryos to test whether these proteins are similarly required to build microtubules at the apical ncMTOC. To monitor microtubules upon depletion, we simultaneously expressed intestine-specific mCherry::TBA-1/α-tubulin along with “early” ZIF-1 (Fig 6A and 6B) or we expressed “late” ZIF-1 and labeled microtubules ubiquitously with mCherry::TBA-1 from a maternally expressed integrated transgene (Fig 6E). In control embryos, microtubules emanate from the apical surfaces of intestinal cells (Fig 6B and 6E), an observation that is visualized by plotting mCherry::TBA-1 intensity along a line drawn across the apical midline and quantified by measuring apical enrichment (Fig 6C–6G, S3 Fig). Despite severe mitotic defects in GIP-1gut(−) cells revealed by live imaging, microtubules still localize robustly to the apical ncMTOC (Fig 6A, S1 Movie). We see apical microtubule enrichment in MZT-1gut(−), GIP-1gut(−), AIR-1gut(−), and [GIP-1; AIR-1]gut(−) embryos that is not significantly different (Fig 6A–6F), and in one case significantly higher (Fig 6G), than in control embryos.
The surprising finding that known microtubule nucleators are not required to build the majority of microtubules at the apical ncMTOC indicates that other mechanisms or molecular players are required to perform this task. We investigated other known microtubule regulators—the anchoring protein NOCA-1/Ninein, stabilizers PTRN-1/CAMPSAP and TPXL-1/TPX2, and the polymerase ZYG-9/chTOG [18,52,53]—to determine if they are required to organize microtubules apically. We found that apical microtubule enrichment was slightly but significantly decreased only upon depletion of ZYG-9, but that, even then, microtubules remained apically enriched (Fig 6B, 6D and 6F). We were particularly surprised to see grossly normal apical microtubule organization in [GIP-1; NOCA-1]gut(−); ptrn-1(0) triple mutant embryos, as noca-1 and ptrn-1 are required in parallel to maintain the organization of non-centrosomal microtubule arrays in hypodermal epithelial cells [45]. These results suggest that different MTOCs, and even different ncMTOCs, have distinct molecular and genetic requirements to generate specific microtubule arrays, and that more mechanisms remain to be identified.
While we found that overall apical organization of microtubules was intact, we explored whether microtubule dynamics were altered upon depletion of microtubule regulators. One possibility is that a majority of microtubules at the apical ncMTOC are stable, persisting from the mitotic divisions prior to polarization. We tested this possibility in two ways. First, we found that upon nocodazole treatment (10 μg/mL and 30 μg/mL), apical microtubule enrichment was significantly reduced both over time and compared to control-treated embryos (Fig 6H–6J), indicating that apical microtubules can be destabilized. This experiment also demonstrates that our measurement methods (line intensity profiles and apical enrichment) are sensitive enough to detect differences in varying amounts of apical microtubules.
Second, we probed microtubule dynamics at the apical ncMTOC by examining the localization of EBP-2/EB1, a microtubule plus-end-binding protein that associates with growing microtubule plus ends. To do this, we tagged the endogenous ebp-2 locus with GFP using CRISPR/Cas9 to visualize endogenous EBP-2 comets. In control embryos, EBP-2 accumulated at the apical surface and moved along the apical surface and out along lateral microtubule tracks toward the basal part of the cell, consistent with microtubules growing from the apical ncMTOC (Fig 7A–7C, S2 Movie). By contrast, EBP-2 enrichment at the apical surface in GIP-1gut(−) embryos was decreased compared to controls (1.51- versus 1.67-fold enriched, p = 0.02; Fig 7C). Additionally, we saw a significant decrease in EBP-2 enrichment in [GIP-1; AIR-1]gut(−) compared to AIR-1gut(−) embryos (1.47- versus 1.64-fold enriched, p = 0.001, Fig 7C), suggesting that the loss of GIP-1 primarily causes the decrease in apical EBP-2 enrichment. Consistent with this observation, we found a significant decrease in the number of EBP-2 comets in [GIP-1; AIR-1]gut(−) compared to AIR-1gut(−) embryos, and in pooled embryos with GIP-1-depleted genotypes compared to pooled embryos with GIP-1(+) genotypes (Fig 7E, Materials and methods). These results suggest that, unlike at the centrosome, γ-TuRC and AIR-1 are not required to build microtubules at the ncMTOC. Instead, γ-TuRC is required to build a subset of dynamic microtubules alongside other γ-TuRC-independent microtubules at the apical ncMTOC.
We next measured the speed of the EBP-2 comets coming from the apical ncMTOC to determine if the dynamics of microtubule growth were altered (Fig 7D and 7F). In control embryos, apically derived comets had an average speed of 0.558 μm/second, which was significantly slower than the reported speeds for comets originating from centrosomes in the early embryo that had been previously labeled with overexpressed EBP-2::GFP [18] or for endogenous early embryo centrosomal comet speeds we measured (“2-cell,” 0.8884 μm/second, p = 7.68 × 10−13, two-tailed t test, Fig 7D and 7F). We also measured comet speeds from centrosomes in the E8–E16 division (“E8,” 0.563 μm/second), which had similar speeds to apically derived comets and were also significantly slower than comets from centrosomes in the 2-cell embryo (Fig 7D and 7F), suggesting that cell type, cell size, or centrosome size may influence comet speed [54]. Surprisingly, we found that the speed of apically derived comets was significantly increased relative to controls in GIP-1gut(−) (0.721 μm/second), [GIP-1; AIR-1]gut(−) (0.726 μm/second), and [GIP-1; NOCA-1]gut(−); ptrn-1(0) (0.706 μm/second) embryos (p < 0.0005 for all comparisons, Fig 7D and 7F). This increase in speed was not observed in AIR-1gut(−) (0.572 μm/second, p = 0.66) or MZT-1gut(−) embryos (0.585 μm/second, p = 0.33), and comets in GIP-1gut(−) embryos were significantly faster than in MZT-1gut(−) embryos (p = 0.002).
This lack of an increase in apically derived comet speed following AIR-1 or MZT-1 depletion argues that changes in intestinal morphology caused by too few intestinal cells, which should be nearly identical in AIR-1gut(−), MZT-1gut(−), and GIP-1gut(−) embryos, cannot account for the observed increase in comet speeds in GIP-1gut(−) embryos. These results further demonstrate that MZT-1 is not required for γ-TuRC function at the apical ncMTOC. The presence of fewer, faster dynamic microtubules following the depletion of GIP-1 suggests that a limiting microtubule growth factor is normally present at the apical ncMTOC and that the loss of γ-TuRC-based microtubules frees up that limiting factor, permitting faster growth of γ-TuRC-independent microtubules (Fig 7G).
Using a tissue-specific protein degradation system, we tested the role of factors essential for building microtubules at the centrosome in building microtubules at an ncMTOC. These studies reveal two important findings (Fig 7G, S5 Fig): (1) all MTOCs are not equivalent, with different MTOCs requiring distinct proteins to build and localize microtubules and microtubule regulators, and (2) ncMTOCs can be composed of discrete populations of γ-TuRC-dependent and -independent microtubules.
We demonstrate that our adapted ZIF-1/ZF degradation system is a robust method for depleting endogenous proteins in a specific tissue of interest (the primordial intestinal epithelium), thereby allowing us to probe the function of early essential genes in differentiating tissues. We monitored and characterized the effectiveness and efficiency of endogenous protein depletion by adding both GFP and ZF via CRISPR to genes of interest. Degradation of many of these critical centrosomal proteins during intestinal divisions caused mitotic defects, confirming that targeted proteins were depleted before the apical ncMTOC formed. With this adapted method now validated, future studies can omit the GFP and use ZF-tagged CRISPR alleles, which will permit a broader range of quantitative analyses of GFP markers.
A consequence of early depletion of important centrosomal proteins was fewer intestinal cells, causing architectural defects in the intestinal primordium, such as overall shorter apical midlines. However, the overall reduced midline surface, especially in [GIP-1;AIR-1]gut(−) embryos, does not explain the general upward trend in apical enrichment of α-tubulin we observed; we found no evidence of a correlation between midline length and α-tubulin enrichment (see Materials and methods). In addition to changes in intestinal geometry, early depletion of centrosomal proteins likely also caused changes in ploidy. While these changes may have impacted zygotic gene expression, they cannot explain, for example, the observed differences among MZT-1gut(−), GIP-1gut(−), and AIR-1gut(−) embryos, which all have similar nuclear numbers and thus likely similar ploidy defects. In sum, differences in microtubule dynamics do not appear to correlate with ploidy or architecture defect severity (Figs 2 and 7), indicating that these secondary defects alone cannot account for the changes in microtubule dynamics we observe.
Using the ZIF-1/ZF system, we characterized the predicted C. elegans ortholog of MZT1, presenting the first in vivo characterization of a MZT1 ortholog in animal development to our knowledge. As in many systems, MZT-1 colocalizes with other γ-TuRC components, is required for γ-TuRC localization to the mitotic spindle pole, and is essential for viability. Surprisingly, we found that γ-TuRC does not require MZT-1 for its localization to the centrioles and apical MTOC. In addition to localizing γ-TuRC, MZT1 is important for nucleation activity of γ-TuRC in Candida and in human tissue culture cells [38,55]. However, we found that only intestinal GIP-1 depletion, and not MZT-1 depletion, impacted comet number and dynamics, suggesting that the MZT-1 found at the apical MTOC is not important for γ-TuRC activity and may simply be a nonfunctional component of the γ-TuRC complex at these non-PCM sites.
Strikingly, we found that dynamic microtubules were still observed growing from the apical ncMTOC following depletion of GIP-1 and AIR-1, which are essential for centrosomal MTOC activity [20]. This finding indicates that additional mechanisms for generating dynamic microtubules must exist. One exciting possibility is that additional, as yet undiscovered nucleators exist in the cell. Based on previous studies on centrosomal microtubules, these hypothetical molecules might be unique for building microtubules at ncMTOCs.
Rather than additional nucleators, another possible mechanism could be through the action of microtubule-stabilizing and -anchoring proteins, as has been seen for other ncMTOCs. In fact, the exact role of γ-TuRC in vivo is not known. The relatively poor nucleation capacity of γ-TuRC in vitro suggests that factors that activate its nucleation capacity at MTOCs exist in vivo [12,56]. Alternatively, the primary function of γ-TuRC might not be nucleation, as is suggested by imaging studies of centrosomes from γ-tubulin-depleted C. elegans embryos [57]; a large number of microtubules are still found associated with the centrosome following γ-tubulin depletion, but are disorganized relative to the centrioles. These data raise the possibility that γ-TuRC functions in anchoring microtubules onto the PCM and that perhaps dynamic microtubules at the apical ncMTOC are generated from many different types of stabilized microtubule seeds. For example, proteins like PTRN-1/Patronin/CAMSAP and NOCA-1/Ninein could protect and anchor small microtubule seeds that grow in parallel to the γ-TuRC-based microtubules. Evidence for this type of model has been seen in Drosophila oocytes and C. elegans skin cells [45,58]. However, our analysis of GIP-1gut(−); NOCA-1gut(−); ptrn-1(0) triple mutants suggests that in embryonic intestinal cells, the ncMTOC does not require PTRN-1 and NOCA-1/Ninein, even in parallel with GIP-1. Furthermore, microtubules remained organized at the apical MTOC upon depletion of the microtubule polymerase ZYG-9/chTOG and the spindle assembly factor TPXL-1/TPX2, suggesting that additional microtubule regulators remain to be discovered.
A final possibility is that MTOCs could facilitate microtubule growth not by localizing nucleators but instead by increasing the local tubulin heterodimer concentration. Microtubules can be nucleated in vitro in the absence of any additional molecules, depending on the concentration of tubulin. Recent studies of in vitro reconstituted PCM suggest that the centrosome might build microtubules in part through the selective concentration of tubulin [53]. ncMTOCs might similarly concentrate tubulin, leading to localized microtubule growth. Two of our findings are consistent with this possibility. First, we observed α-tubulin enrichment at the apical MTOC following microtubule depolymerization with nocodazole, although we cannot distinguish between free tubulin heterodimers and small protected microtubule seeds. Second, we found a small but significant decrease in apical microtubule enrichment in ZYG-9gut(−) embryos, which can concentrate tubulin in vitro at SPD-5 condensates [53] and could perhaps play a similar role at the apical ncMTOC.
Our finding that microtubules have increased growth speeds following GIP-1 depletion suggests that a limiting growth factor is normally present at the apical ncMTOC. This limiting factor could be sequestered by γ-TuRC itself, or it could be limiting because of competition for it among the large number of growing microtubules. In the first case, we would expect loss of γ-TuRC to release this factor and cause both increased microtubule growth speed and comet number. However, we observed fewer, faster comets and decreased apical EBP-2 enrichment upon GIP-1 depletion. These observations are more consistent with a model in which loss of γ-TuRC leads to fewer growing microtubules, thereby increasing the availability of a limiting factor to growing microtubules and allowing faster growth (Fig 7G). This limiting factor could be tubulin heterodimers themselves; however, the mechanisms for concentrating a pool of tubulin at a membrane are completely unknown.
Finally, we found that γ-TuRC and AIR-1 are not required to form the majority of apical microtubules, raising the question of why these proteins so specifically localize there as a new ncMTOC is being established. One possibility is that the main function of localizing γ-TuRC and AIR-1 to the apical ncMTOC is to effectively remove them from the centrosome at the end of mitosis as centrosomal MTOC function is attenuated. We hypothesize that the ability to remove microtubules from the centrosome is an important step in mitotic exit, as hyperactive MTOC function at the centrosome has been linked to cancer [59–61]. Creating a sink for centrosomal microtubule regulators at an alternative site in the cell would provide a quick and effective way of maintaining the inactivation of MTOC function at the centrosome. Different cell types require a large variety of specific patterns of microtubule organization, and future work will be critical to discover the additional molecular players and mechanisms that contribute to the formation and function of different types of MTOCs.
All data used for quantitative analyses are included as S1 Data. Image files are available upon request.
Nematodes were cultured and manipulated as previously described [62]. Experiments were performed using 1- or 2-day-old adults. The strains used in this study are listed in Table 1.
Two CRISPR editing methods were used to generate knock-in insertions of the ZF::GFP or GFP tag. gip-1(wow3[gfp::gip-1], gip-1(wow5[zf::gfp::gip-1]) and ptrn-1(wow4[ptrn-1::gfp]) were constructed using repair templates with short homology arms (SHAs) and PCR-based screening to test for insertions into the endogenous loci [63]. All other CRISPR-based insertions were achieved using a modified self-excising cassette (SEC), into which the ZF-tag had been inserted N-terminal to GFP (plasmid JF250) [64]. The SEC-derived CRISPR alleles thus also contain additional tags: 3×FLAG for GFP alleles and 3×Myc for TagRFP-T alleles. Cas9 and sgRNAs were delivered using plasmid pDD162, into which the appropriate sgRNA sequence had been added with the Q5 Site-Directed Mutagenesis Kit (NEB). The Cas9/sgRNA plasmid, repair template, and appropriate selection markers were injected into N2 or zif-1(gk117) mutant 1-day-old adult worms. Worms were recovered and processed according to published protocols [63,64]. Successfully edited worms were outcrossed at least two times before being used for subsequent experiments. sgRNA and homology arm sequences are listed in S1 Table.
To assess embryonic viability, 20–30 young adult hermaphrodites of each genotype were singled onto small plates and allowed to lay for 4 hours at 20°C, and then adults were removed and eggs were counted. After 3 days at 20°C, the number of surviving worms present on each plate was counted, and viability for each plate was calculated as the total number of L4s and adults divided by the number of eggs. When comparing N2 and zif-1(−) (JLF155) lethality, two N2 plates had more surviving worms at day 3 than the number of eggs initially counted, and those plates were omitted.
The ZF/ZIF-1 system was executed as previously described, with the following modifications. ZF::GFP tags were inserted into endogenous loci in a zif-1(gk117) mutant background using CRISPR editing techniques (see above). Exogenous ZIF-1 was expressed from either extrachromosomal or integrated arrays (Table 2), which were generated by injecting 50 ng/μL of each plasmid and of a co-injection marker (either pRF4[rol-6(d)] at 100 ng/μL or myo-2p::mcherry at 2.5 ng/μL) into zif-1(gk117) mutant animals, with pBS to inject at a minimum concentration of 150 ng/μL. For comparisons across multiple genotypes, the same array was introduced by mating.
To test the efficiency of AIR-1 depletion—specifically, to test whether zygotic expression of AIR-1::ZF::GFP provided AIR-1 activity in AIR-1gut(−) embryos—we generated air-1(0) embryos with only maternally provided AIR-1::ZF::GFP, which was degraded in the intestinal primordium by ZIF-1 expression. Our genetic strategy was as follows: We used the air-1(ok571) allele, a deletion that removes a large portion of the AIR-1 kinase domain and is a putative null allele. We crossed zif-1(gk117) stIs10220[end-1p::his1::mcherry]; air-1(ok571)/oxTi878[vha-6p::gfp]; wowEx10[elt-2p::zif-1; end-1p::mcherry::tba-1] hermaphrodites with zif-1(gk117) stIs10220; air-1(wow14[air-1::zf::gfp]) males to generate F1 “air-1::zf::gfp/(0)” hermaphrodites, and scored their F2 embryos. We had no markers to distinguish embryonic genotypes, so we scored all progeny and assumed Mendelian segregation of genotypes: 25% air-1::zf::gfp homozygotes, 50% air-1::zf::gfp/(0) heterozygotes, and 25% air-1(0) homozygotes, all with maternally expressed AIR-1::ZF::GFP. Consistent with this predicted distribution, we saw many larvae with perturbed gonad development, a hallmark of the air-1(0) phenotype [65]. We denote these pooled embryos as AIR-1* in Fig 2. If zygotic expression of AIR-1 was providing activity, we would expect to observe a change in the distribution of intestinal nuclear number in AIR-1gut(−) compared with AIR-1gut(−)*. However, intestinal nuclear numbers in AIR-1gut(−)* and AIR-1gut(−) embryos were indistinguishable, suggesting that zygotic expression of AIR-1::ZF::GFP provides little or no AIR-1 rescuing activity and that AIR-1::ZF::GFP depletion is highly efficient.
We used α-GFP immunostaining to test how completely ZF::GFP-tagged proteins are degraded in the intestinal primordium (see S2 Fig). We stained [GIP-1; AIR-1]gut(−) embryos (JLF232), in which heterozygous ifb-2p::zif-1(wowIs3) causes intestine-specific degradation of both proteins in embryos that inherit the transgene. GFP fluorescence survives the staining protocol, so we asked how the perduring GFP signal correlates with Cy3 signal from α-GFP staining. Most JLF232 embryos fell into two categories: strong Cy3 and GFP signal at centrosomes or the apical membrane (n > 40), or no localized Cy3 nor GFP signal in the intestinal primordium (n = 30). As α-GFP antibody staining is highly sensitive, this result suggests that GIP-1 and AIR-1 are both degraded quickly (prior to GFP maturation) and efficiently, and that the loss of visible GFP in live embryos accurately indicates efficient ZF-tagged protein degradation.
We did not find a balancer that allowed us to maintain an EBP-2::GFP; [GIP-1;AIR-1]gut(−) strain. We thus used the following cross strategy to generate these embryos: ebp-2(wow47[ebp-2::gfp]); zif-1(gk117)/eT1[nIs267]; wowEx10[elt-2p::zif-1] hermaphrodites were crossed with JLF298 males to generate ebp-2::gfp; zf::gfp::gip-1 zif-1(gk117)/eT1[nIs267]; air-1::zf::gfp/eT1; wowEx10 F1 hermaphrodites. F1 cross progeny were identified by PCR genotyping for air-1::zf::gfp (air-1 F: GAACGTCTCCCACTTGTTGACATC, ZF R: TTTTTCTACCGGTACCCTCGG). F2 progeny with the wowEx10 array and without the eT1 balancer were selected (ebp-2::gfp; zf::gfp::gip-1 zif-1(gk117); air-1::zf::gfp; wowEx10). zf::gfp::gip-1 was frequently heterozygous, likely due to recombination between eT1 and zf::gfp::gip-1 in the F1 mother. Thus, to ensure we were looking at EBP-2::GFP; [GIP-1;AIR-1]gut(−) embryos, we only scored embryos with four intestinal nuclei (see Fig 2).
JLF36 embryos were treated with nocodazole, as has been previously described [6]. Briefly, trypan blue–coated embryos were affixed to poly-lysine-coated coverslips and submerged in embryonic growth medium (EGM) [66] containing either 0.1% DMSO or 10 or 30 μg/mL nocodazole. E16-stage embryos with their dorsal surfaces facing the coverslip were selected for analysis. The eggshell and vitelline membrane were punctured using a Micropoint nitrogen dye laser (Andor Technology, Belfast, United Kingdom), allowing the EGM + DMSO or nocodazole to reach the embryo. Embryos were imaged immediately following puncturing (T1, 10–45 seconds later) and then again after 10 minutes (T2).
Embryos were fixed and stained as previously described [23]. Briefly, embryos of the appropriate stage were collected and adhered to a poly-lysine-coated slide with a Teflon spacer and covered with a coverslip. Embryos were fixed by freeze-crack followed by 100% MeOH for 5 minutes. Embryos were rehydrated in PBS and incubated with rabbit α-GFP primary antibody (Abcam, Cambridge, United Kingdom, 1/200) or mouse α-PAR-3 [28] either overnight at 4°C or for 1 hour at 37°C. Embryos were washed in PBT and incubated with the appropriate CY3-conjugated secondary antibodies (Jackson ImmunoResearch Laboratories, West Grove, PA, 1/200) and 100 ng/mL DAPI (Sigma-Aldrich, St. Louis, MO) overnight at 4°C or for 1 hour at 37°C and mounted under coverslips in Vectashield (Vector Laboratories, Burlingame, CA). At least 20 mutant embryos were scored for analysis of mutant phenotypes.
Embryos for microscopy were dissected from gravid hermaphrodites incubated for 4–4.5 hours in M9 (at 20°C for EBP-2::GFP comet speed, at 25°C for EBP-2::GFP enrichment, and at room temperature for all other experiments). For live imaging, samples were mounted on a pad made of 3% agarose dissolved in M9. Live imaging was performed on a Nikon Ti-E inverted microscope (Nikon Instruments, Melville, NY) using a 60× Oil Plan Apochromat (NA = 1.4) or 100× Oil Plan Apochromat (NA = 1.45) objective and controlled by NIS Elements software (Nikon). Images were acquired with an Andor Ixon Ultra back thinned EM-CCD camera using 488 nm or 561 nm imaging lasers and a Yokogawa X1 confocal spinning disk head equipped with a 1.5× magnifying lens. For live imaging, images were taken at a sampling rate of 0.5 μm. Images were processed in NIS Elements, the Fiji distribution of ImageJ (“Fiji”) [67,68], or Adobe Photoshop. Fixed images were obtained using a 60× Oil Plan Apochromat objective (NA = 1.4) on either the above system or a Nikon Ni-E compound microscope with an Andor Zyla sCMOS camera.
|
10.1371/journal.ppat.1000624 | EBNA1-Mediated Recruitment of a Histone H2B Deubiquitylating Complex to the Epstein-Barr Virus Latent Origin of DNA Replication | The EBNA1 protein of Epstein-Barr virus (EBV) plays essential roles in enabling the replication and persistence of EBV genomes in latently infected cells and activating EBV latent gene expression, in all cases by binding to specific recognition sites in the latent origin of replication, oriP. Here we show that EBNA1 binding to its recognition sites in vitro is greatly stimulated by binding to the cellular deubiquitylating enzyme, USP7, and that USP7 can form a ternary complex with DNA-bound EBNA1. Consistent with the in vitro effects, the assembly of EBNA1 on oriP elements in human cells was decreased by USP7 silencing, whereas assembly of an EBNA1 mutant defective in USP7 binding was unaffected. USP7 affinity column profiling identified a complex between USP7 and human GMP synthetase (GMPS), which was shown to stimulate the ability of USP7 to cleave monoubiquitin from histone H2B in vitro. Accordingly, silencing of USP7 in human cells resulted in a consistent increase in the level of monoubquitylated H2B. The USP7-GMPS complex formed a quaternary complex with DNA-bound EBNA1 in vitro and, in EBV infected cells, was preferentially detected at the oriP functional element, FR, along with EBNA1. Down-regulation of USP7 reduced the level of GMPS at the FR, increased the level of monoubiquitylated H2B in this region of the origin and decreased the ability of EBNA1, but not an EBNA1 USP7-binding mutant, to activate transcription from the FR. The results indicate that USP7 can stimulate EBNA1-DNA interactions and that EBNA1 can alter histone modification at oriP through recruitment of USP7.
| Epstein-Barr virus (EBV) infections persist for the lifetime of the host largely due to the actions of the EBNA1 viral protein. EBNA1 enables the replication and stable persistence of EBV genomes and activates the expression of other EBV genes by binding to specific DNA sequences in the EBV genome. We have shown that the cellular protein USP7 stimulates EBNA1 binding to its DNA sequences and that EBNA1 recruits USP7 to the EBV genome, which in turn recruits another cellular protein GMP synthetase. The complex of USP7 and GMP synthetase then functions to alter the chromatin structure at a region of the EBV genome that controls EBV persistence. These changes to the EBV genome are likely important for enabling the persistence of EBV genomes in infected cells.
| Epstein-Barr virus (EBV) is a gamma herpesvirus that infects over ninety percent of people worldwide. As part of its latent life cycle, EBV efficiently immortalizes the host cell and predisposes it to a number of malignancies, including Burkitt's lymphoma, nasopharyngeal carcinoma, gastric carcinoma, Hodgkin's disease and a variety of lymphomas in immunosuppressed patients [1]. In latently infected cells, replication and maintenance of the viral genome require the latent origin of replication, oriP and the EBNA1 protein. OriP is comprised of two functional elements, the dyad symmetry (DS) and the family of repeats (FR), which contain four and twenty copies of an 18 bp palindromic EBNA1 binding site respectively [2],[3]. Replication of oriP-containing plasmids requires EBNA1 binding to the DS [4]. EBNA1 binding to the FR is required for the mitotic segregation of the oriP-containing plasmids and transactivation of several latency genes [5],[6].
EBNA1 binds DNA through residues 459–607, which form the DNA binding and dimerization domain (EBNA1-DBD) [7]–[9]. High resolution structures of the EBNA1-DBD, alone and in complex with its DNA binding site, have revealed details of the interaction of EBNA1 with DNA [10]–[12]. EBNA1-DBD comprises two subdomains: residues 504–604, referred to as the core-domain, and residues 461–503, referred to as the flanking domain. The core domain is a β-barrel structure that forms the dimerization interface and makes transient sequence-specific contacts with the DNA through an α-helix [10],[13]. The flanking domain consists of an α-helix (residues 477–489) oriented perpendicular to the axis of the DNA, which contacts the major groove through Lys 477, and an extended chain (amino acids 461–469) that runs along the base of the minor groove of the DNA, making sequence-specific contacts through Lys-461, Gly-463 and Arg-469 [11].
In addition to binding specific DNA sequences, EBNA1 is also known to interact with several host-cell proteins, which in some cases have been shown to mediate EBNA1 functions at oriP [14]–[18]. EBNA1 can also affect cellular processes through sequestration of cellular proteins, as best exemplified by the EBNA1 interaction with the ubiquitin specific protease USP7, also referred to as Herpesvirus Associated Ubiquitin Specific Protease (HAUSP). USP7 was originally identified as a binding partner of the ICP0 protein of herpes simplex virus (HSV) [19] and, since then, several cellular targets of USP7 have been identified including the p53 tumour suppressor protein [20]–[24]. In response to genotoxic stress, USP7 binds and deubiquitylates p53 thereby protecting it from proteasome-mediated degradation. In addition to cleaving polyubiquitin chains, USP7 has been reported to reverse monoubiquitylation in some proteins (eg. p53 and FOXO4), thereby affecting their subcellular localization [25],[26]. Similarly, the Drosophila homologue of USP7 was found to contribute to epigenetic silencing by reversing monoubiquitylation of histone H2B, and this activity required USP7 to be in complex with guanosine 5′ monophosphate synthetase (GMPS) [27].
Our studies on the EBNA1-USP7 interaction have shown that EBNA1 binds the N-terminal domain of USP7 (USP7-NTD), which is distinct from the catalytic domain, and is the the same domain that is bound by p53 [28]. EBNA1 and p53 bind the same pocket in this domain but EBNA1 does so with an affinity that is approximately 10-fold higher than that of p53 [28],[29]. As a result, EBNA1 interferes with the binding and stabilization of p53 by USP7 and with p53-mediated apoptosis in response to DNA damage [29],[30]. In addition, we recently found that EBNA1 disrupts promyelocytic leukemia (PML) nuclear bodies (also called ND10s) in nasopharyngeal carcinoma cells by inducing the degradation of the PML proteins [30]. This activity required USP7 and the EBNA1-USP7 interaction, indicating that this interaction can modulate cellular events in addition to p53 levels.
EBNA1 deletion analysis showed that the USP7 binding sequence in EBNA1 was just N-terminal to the flanking DNA binding domain and subsequent peptide binding assays identified EBNA1 residues 436–450 as sufficient for this interaction [28],[29]. A crystal structure of an EBNA1 peptide bound to the USP7-NTD revealed multiple interactions of EBNA1 residues 442–448 with amino acids in a shallow groove of the TRAF domain formed by the USP7-NTD [29]. In particular interactions mediated by Ser447 in EBNA1 were shown to be critical for USP7 binding. Given the large size of USP7 (135 kDa) and the proximity of its binding site to the EBNA1-DBD residues that are inserted in the DNA minor groove (amino acids 461–469), we wondered whether the USP7 interaction interfered with EBNA1 binding to DNA. Here we report that, contrary to our expectations, USP7 had a large stimulatory effect on the DNA-binding activity of EBNA1 in vitro and can form a ternary complex with DNA-bound EBNA1. Furthermore, USP7 was found to bind GMPS, forming a complex active in histone H2B deubiquitylation, and this complex was recruited to oriP in EBV-infected cells resulting in decreased H2B ubiquitylation.
We initially assessed the effect of USP7 on the DNA binding activity of EBNA1 using electrophoretic mobility shift assays (EMSAs) with a version of EBNA1 that has a shortened Gly-Ala repeat but has wildtype activity for all known EBNA1 functions (referred to as EBNA1; Figure 1A). Purified EBNA1 was incubated with radiolabelled DNA containing a single EBNA1 recognition site (site 1 from the DS element) in presence and absence of excess purified full length USP7. We consistently observed that USP7 stimulated the DNA binding activity of EBNA1 as shown in the representative experiment in Figure 1B (left panel), while no obvious effects on EBNA1-DNA interactions were seen with nonspecific proteins such as BSA (Figure 1B, right panel). Results from multiple experiments showed a 20-fold increase in the DNA binding affinity of EBNA1 in the presence of USP7, resulting in a shift in the dissociation constant (Kd) from 85±7nM for EBNA1 alone to 4.3±0.4 nM for EBNA1 in presence of USP7. This increase in DNA binding affinity was largely dependant on the ability of EBNA1 to bind USP7, as the DNA binding ability of a truncation mutant of EBNA1 (EBNA1452–641) containing the DNA-binding and dimerization region but lacking the USP7 binding site was much less affected by USP7 (on average showing a 4-fold increase in DNA binding in the presence of USP7; Figure 1C).
EBNA1 dimers bound to DNA are known to interact with each other resulting in the crosslinking of multiple DNA fragments through large EBNA1 complexes (referred to as looping or linking interactions) [31]–[33]. These complexes are retained in the wells of the gel in EMSAs as shown in Figure 1B, precluding analysis of the effect of USP7 on the migration of the DNA complexes. The linking interactions of EBNA1 are mediated largely by amino acids 325–376 and to a lesser degree by EBNA1 N-terminal residues [32],[34]. To further evaluate the effect of USP7 on the DNA binding ability of EBNA1 without the confounding effects of DNA linking, we repeated the EMSAs with the EBNA1 truncation mutant 395–641 (Figure 1A), which contains the USP7 binding site and the DNA-binding region but lacks sequences that cause DNA linking. When the DNA binding affinity of EBNA1395–641 was measured in the presence and absence of excess USP7, USP7 was consistently found to stimulate DNA binding by EBNA1395–641 (Figure 2A, left panel), resulting in a 50-fold decrease in the calculated Kd from 233±76 nM for EBNA1395–641 alone to 4±1.8 nM for EBNA1395–641 in presence of USP7. This experiment also showed that the bound DNA migrated more slowly in the presence of EBNA1395–641 and USP7 than with EBNA1395–641 alone, suggesting that USP7 formed a ternary complex with EBNA1395–641 and DNA.
Since EBNA1 is known to bind to the N-terminal TRAF domain of USP7 (USP7-NTD) [28],[29], we examined whether this domain was sufficient to stimulate EBNA1395–641 binding to DNA. When EBNA1395–641 titrations were performed in the presence of excess USP7-NTD, the DNA binding activity was increased 8 to 16-fold in multiple experiments, (Figure 2A, right panel) indicating that the USP7-NTD was partially, but not completely, responsible for the stimulatory effect of USP7 on EBNA1 DNA binding activity. Consistent with the USP7 result, the USP-NTD was found to decrease the migration of the EBNA1-bound DNA suggesting that it can bind the EBNA1-DNA complex.
We also examined the stimulatory effect of USP7 on DNA binding by EBNA1395–641 by incubating a fixed amount of EBNA1395–641 (sufficient to bind a small fraction of the DNA probe on its own) with increasing amounts of USP7 prior to the addition of the DNA binding site Figure 2B, left panel). EMSAs performed in this way showed that USP7 had a dose-dependent effect on the DNA binding activity of EBNA1395–641. The possibility that USP7 itself had some ability to bind the DNA probe was tested by titrating USP7 with the DNA in the absence of any EBNA1, but USP7 alone did not shift the DNA probe even at very high concentrations of USP7 (Figure 2B, right panel lanes 8–12). Similarly, the USP7-NTD on its own did not bind the DNA-probe (Figure 2B, right panel lanes 1–7).
The experiments in Figure 2A indicated that USP7 can bind the EBNA1-DNA complex resulting in a supershift while the titration performed with lesser amounts of USP7 in Figure 2B did not show a supershift. To investigate this discrepancy, we preformed EBNA1-DNA complexes (using EBNA1395–641 as above) then added increasing amount of USP7 (Figure 2C). EMSAs confirmed that USP7 was able to supershift the EBNA1395–641-DNA complex but only at higher concentrations of USP7 (compare lanes 6 and 7 to lanes 2–5). To confirm that the supershifted band contained EBNA1, complexes formed as in lanes 2 and 7 were incubated with an EBNA1-specfic antibody prior to electrophoresis. In both cases the antibody supershifted the bands to the gel wells, whereas no effect of the antibody was seen on the migration of the DNA probe in the absence of EBNA1 (Figure 2C, lanes 8–10). The results indicate that USP7 can form a ternary complex with DNA-bound EBNA1 under some conditions.
During initial EBV infection, EBNA1 assembles on its recognition sites in oriP and remains stably bound to these sites in all types of latently infected cell lines. Therefore it was not possible to determine the effects of USP7 on EBNA1 assembly on oriP using latently infected cells. Instead, we assessed the effect of USP7 on the initial association of EBNA1 with oriP by treating EBV-negative nasopharyngeal carcinoma cells (CNE2Z) with siRNA against USP7 or GFP (negative control) and then transfecting these cells with an oriP plasmid expressing EBNA1 or an EBNA1 mutant (Δ395–450; see Figure 1A) that we previously showed was specifically defective in binding USP7 [14] and a plasmid lacking EBNA1 binding sites (pLacZ) as control for nonspecific DNA binding. Chromatin immunoprecipitation (ChIP) assays were then performed using EBNA1-specific antibodies to assess the degree of EBNA1 association with the the oriP FR and DS elements and lacZ (negative control) as compared to nonspecific rabbit IgG. EBNA1 was readily detected on both the DS and FR elements after siGFP treatment but the association with both elements was greatly decreased by USP7 silencing (Figure 3A, middle panels). As expected, there was little association of EBNA1 with lacZ and this was unaffected by USP7 silencing (right panel). Consistent with the in vitro results, Δ395–450 bound less efficiently to both the DS and FR elements than did wildtype EBNA1, despite being expressed at equivalent levels as EBNA1 (see Figure 3A, left panel). Moreover, unlike wildtype EBNA1, the interaction of Δ395–450 with the FR and DS elements was not affected by USP7 silencing. Therefore we conclude that USP7 can stimulate the assembly of EBNA1 on oriP elements in vivo.
In addition to binding the oriP elements, EBNA1 can interact in a more transient manner with a third region of the EBV genome (referred to as region III), consisting of two lower affinity EBNA1 recognition sites within the BamHI-Q fragment, and this interaction can negatively regulate the Qp promoter used for EBNA1 expression in some types of EBV latency [3],[35],[36]. Due to the transient nature of the EBNA1 interaction with region III, we asked whether USP7 might promote the EBNA1-region III interaction in latently infected cells. D98/Raji cells were used for these experiments since these EBV-infected cells are more transfectable than the Raji cells from which they were derived. D98/Raji cells were transfected with siRNA against USP7 or GFP then ChIP experiments were performed using EBNA1-specific antibody and primer sets for region III. While we did not achieve complete silencing of USP7 in these experiments (Figure 3B, left panel), its down-regulation was consistently found to decrease the association of EBNA1 with region III (Figure 3B, right panel), indicating that USP7 can also modulate EBNA1-DNA interactions in the context of an EBV infection.
The above in vitro analyses raised the possibility that EBNA1 may recruit USP7 to oriP in EBV-infected cells. To test this possibility we conducted ChIP experiments in EBV-positive B-lymphocytes (Raji cells). Antibodies against EBNA1 or USP7 were used to immunoprecipitate these proteins from sheared Raji DNA and compared to non-specific rabbit IgG as a negative control. Immunoprecipitates were analyzed by quantitative real-time PCR using primers specific for the DS and FR regions in oriP and for the promoter region of the BZLF gene, located 40 kb away from oriP. EBNA1 is known to be constitutively bound to the FR and DS elements [37],[38] and, consistent with this, was readily detected on both the FR and DS DNA fragments (with better recovery of the DS element as has been previously observed;[16],[39],[40]) but was not detected on the BZLF1 fragment (Figure 4A). The USP7 antibody consistently isolated more FR DNA fragment than either the DS or BZLF1 fragments (Figure 4A). Recovery of the FR region (but not the DS region) was significantly higher than that of the BZLF1 region with a p-value of 0.0004. The results indicate that USP7 is preferentially recruited to FR and is consistent with the higher enrichment of EBNA1 at the FR.
USP7 is known to regulate p53 levels but this would not seem to explain why it is recruited to oriP. To gain insight into other potential functions of USP7, we used a proteomics approach to identify cellular protein partners of USP7. To this end, increasing amounts of purified USP7 was coupled to resin to generate a series of USP7 affinity columns and a constant amount of human cell extract was passed through each column. Proteins retained on the columns were eluted with 1 M NaCl, followed by 1% SDS, and the recovered proteins were analysed by SDS-PAGE and silver staining (Figure 5A). Only 1 band (at approximately 70 Kda) was observed to be specifically retained on the USP7 column, showing a titratable interaction with USP7 as expected for a specific protein interaction, and this was identified by MALDI-ToF mass spectrometry as GMP synthetase (GMPS).
The interaction between USP7 and GMPS was further examined by glycerol gradient sedimentation analysis of the purified proteins. For these experiments, GMPS, like USP7, was generated using a baculovirus and extensively purified. Analysis of the individual proteins by glycerol gradient sedimentation showed that USP7 migrates close to its calculated molecular mass of 130 Kd indicating that it is monomeric (Figure 5B, top panel). This is consistent with previous analytical centrifugation analyses [28]. GMPS was found to migrate at a similar position as USP7 despite its smaller molecular mass of 77 Kda suggesting that it forms dimers (Figure 5B, middle panel), as occurs for E.coli GMPS [41]. When USP7 and GMPS were combined, their positions in the gradient shifted to a higher molecular weight form, confirming that the two proteins directly interact (Figure 5B, bottom panel). The size of this complex (approximately 200 Kda) suggested that it consisted of one USP7 and one GMPS molecule.
A previous study reported that Drosophila USP7 formed a complex with GMPS in Drosophila embryos and that this complex deubiquitylated histone H2B thereby contributing to polycomb-mediated silencing [27]. This prompted us to investigate whether the human USP7-GMPS complex also functioned to deubiquitylate histone H2B. To this end, we purified total histones from HeLa cells by the acid extraction method and incubated them with purified USP7 (at a MW ratio of USP7∶histones of 1∶1000) for various times prior to Western blot analysis. Histone H2B and its monoubiquitylated form (Ub-H2B) were initially detected using an antibody specific to histone H2B, and USP7 was found to have some ability to deubiquitylate H2B on its own (Figure 6A, left panel). Histone H2A and its monoubiquitylated form were detected in the same assay with antibody specific to H2A, however, in contrast to the H2B results, USP7 was not observed to deubiquitylate H2A (Figure 6A, right panel).
To determine if GMPS affected the ability of USP7 to deubiquitylate H2B, we repeated the experiments including different amounts of GMPS (Figure 6B). The Ub-H2B was more readily detected using an anti-ubiquitin antibody, providing a more robust signal to follow and this band is shown in Figure 6B. We found that the addition of GMPS at amounts stoichiometric to USP7 increased the cleavage of Ub-H2B by USP7 at each time point examined (compare “1∶1” samples to “USP7” samples within each panel). Increasing the amount of GMPS 10-fold had no further stimulatory effect (compare “1∶10” samples to “1∶1” samples in the left panel), while decreasing the amount of GMPS 10-fold abrogated the stimulatory effect (compare “10∶1” samples to “1∶1” samples in the right panel). These results are consistent with GMPS stimulating deubiquitylation of H2B by USP7 by forming a stoichiometric complex with USP7 and are inconsistent with GMPS acting catalytically. We also asked whether the stimulatory activity of GMPS was specific to H2B deubiquitylation or also occurred for other USP7 targets. To this end, we incubated USP7, with or without equal amounts of GMPS, with p53 that had been polyubiquitylated in vitro and we followed the p53 forms by Western blotting with a p53 antibody (Figure 6C). In this case, we saw no obvious difference in the kinetics of cleavage of the ubiquitylated forms by USP7 with or without GMPS, indicating that GMPS does not affect all USP7 targets equally and rather has specificity for Ub-H2B.
To assess whether USP7 regulates histones in human cells, we down-regulated USP7 in HeLa cells with siRNA treatment then prepared total histones as for the in vitro assays. The ratio of monoubiquitylated to nonmodified forms of H2A and H2B were then determined by Western blotting using antibodies against H2A and H2B. An example of the results obtained is shown in the gel images in Figure 6D as compared to results with the same cells treated with siGFP as a negative control. We consistently observed an increase in the ratio of Ub-H2B to total H2B after USP7 silencing, as compared to GFP silencing (negative control), but we did not see a reproducible effect on the H2A monubiquitylated form. Results from three independent experiments are shown in histogram in Figure 6D. Therefore the in vivo studies support the conclusions of the in vitro results, that USP7 can regulate H2B monoubiquitylation.
We next investigated the relevance of the USP7-GMPS interaction for EBNA1, in particular whether GMPS could form part of the USP7-EBNA1-DNA complex. We examined this in two ways: First, we tested possible interactions between DNA-bound EBNA1395–641 with GMPS with and without USP7 by EMSAs (Figure 7). The binding of EBNA1395–641 to the DNA probe was assessed on its own or after incubation of the same amount of EBNA1 with USP7 or GMPS and the migration of the DNA complexes was assessed. As observed above, USP7 shifted the EBNA1-DNA complex to a slower migrating form indicative of a ternary complex (Figure 7, compare lanes 2 and 3). On the other hand, the same amount of GMPS did not alter the mobility of the EBNA1-DNA complexes (Figure 7, compare lanes 2 and 4). This was expected since there is no evidence of a direct interaction between EBNA1 and GMPS. However, when USP7, GMPS and EBNA1 were combined (the same amounts as when tested individually), and then added to the DNA, these complexes shifted to a position higher than that of the USP7-EBNA-DNA ternary complex as shown in lanes 5 and 6 of Figure 7 (compare to lane 3). However neither GMPS, USP7 nor GMPS+USP7 interacted with the DNA in the absence of EBNA1 (Figure 7, lanes 8–10). The results suggest that USP7 mediates an interaction between GMPS and the EBNA1-DNA complex resulting in the formation of a quaternary complex.
We also examined the possible association between USP7-GMPS complexes and EBNA1 in vivo, by determining if GMPS localized with EBNA1 and USP7 on EBV chromatin. ChIP experiments performed on Raji cells, showed that, like USP7, GMPS was preferentially detected at the FR element of oriP over the DS element or the BZLF1 region (Figure 4A, right panel). This is consistent with the recruitment of the USP7-GMPS complex to the FR through EBNA1.
We next investigated whether recruitment of GMPS to the FR was dependent on USP7, as suggested by the EMSA experiments. These experiments required down-regulation of USP7 by siRNA treatment and could not be performed in Raji cells due to their low transfection efficiency. Instead, the more readily transfectable D98/Raji fusion cells were used, which retain the EBV genomes from Raji cells [42]. USP7 was confirmed to be down-regulated in these cells following treatment with siRNA against USP7 but not siRNA against GFP (negative control), while GMPS levels were not affected (Figure 4B, left panel). ChIP analysis of GMPS from these cells showed that, as in Raji cells, GMPS was preferentially localized to the FR region, and that down-regulation of USP7 resulted in decreased levels of GMPS at the FR (P value 0.01 relative to FR-siGFP samples; Figure 4B, right panel).
If the USP7-GMPS complex functions to deubiquitylate histone H2B, then the loss of this complex from the FR would be expected to increase the level of Ub-H2B in this region. We investigated this possibility by performing ChIP experiments with and without USP7 silencing, using an antibody that recognizes only the ubiquitylated form of H2B [43]. To control for possible differences in the number of histones at each region we performed the same experiment with antibody against total histone H2B and expressed the Ub-H2B as a ratio of this value. In Figure 4C (left panel) the change in the fraction of Ub-H2B after USP7 silencing is shown from multiple experiments (in relation to siGFP treatment). While we saw considerable variability on the level of Ub-H2B at the BZLF1 region, we consistently observed that USP7 silencing resulted in increased levels of Ub-H2B at the FR and had little effect on Ub-H2B levels at the DS. The results support the model that USP7 is needed for recruitment of GMPS to the FR and subsequent deubiquitylation of histone H2B.
Since EBNA1 binding to the FR is known to activate transcription from the LMP1 and Cp promoters [44],[45], we examined the possibility that the recruitment of the USP7-GMPS complex to the FR might also affect H2B ubiquitylation at these promoters. To this end, ChIP was performed on D98/Raji cells before and after silencing USP7, using antibodies against Ub-H2B and total H2B. The recovery of the LMP1 and Cp promoter regions was quantified for each treatment and the change in the fraction of Ub-H2B after USP7 silencing was determined. Silencing of USP7 consistently resulted in increased Ub-H2B at both the LMP1 and Cp promoters, with the strongest effect on the Cp promoter, whereas H2B ubiquitylation at the oriLyt region of EBV (negative control) was not affected by USP7 silencing (Figure 4C, right panel). The results suggest that the USP7-GMPS complex not only affects H2B ubiquitylation at the FR but also at promoters controlled by the FR.
The above observations suggest that EBNA1-mediated recruitment of the GMPS-USP7 complex to the FR may contribute to transcriptional activation by this element through alteration of Ub-H2B at the FR and/or promoters under FR control. To test this possibility, we treated EBV-negative CNE2Z cells with siRNA against USP7 or GFP then co-transfected them with a reporter plasmid in which expression of chloramphenical acetyl transferase (CAT) is under FR control and with a plasmid expressing either EBNA1, the EBNA1 Δ395–450 mutant that is unable to bind USP7 or no EBNA1 (oriP plasmid). CAT assays were then performed on each sample to assess degree of transcriptional activation (Figure 8). As expected strong transcriptional activation was seen after siGFP treatment in the presence of EBNA1 but not in its absence and, as previously reported [14], Δ395–450 had slightly reduced transcriptional activity. USP7 silencing caused a significant decrease in transcriptional activation by EBNA1 (P value 0.004) but did not significantly affect transactivation by Δ395–450. These results support the model that recruitment of the USP7-GMPS complex by EBNA1 contributes to EBNA1-mediated transcriptional activation.
EBNA1 forms a stable complex with host cell USP7 and this interaction can promote cell survival, at least in part through interfering with p53 stabilization by USP7 and through disrupting PML nuclear bodies [14], [28]–[30]. Here we provide the first evidence that the EBNA1-USP7 interaction also contributes to EBNA1 functions at EBV oriP. This study stemmed from the unexpected observation that USP7 greatly stimulated the DNA binding activity of EBNA1 in vitro and could form a ternary complex with DNA-bound EBNA1. EBNA1 appears to be constitutively bound to oriP elements in latent EBV infections in proliferating cells [37],[38] and, in these cases, the functional relevance of these observations for oriP-related functions most likely lies in the ability of USP7 to form a ternary complex with DNA-bound EBNA1, as verified at the FR element in EBV-infected cells. In keeping with this hypothesis, we found that USP7 within this complex can mediate an interaction with GMPS which promotes deubiquitylation of histone H2B and that USP7 contributes to EBNA1-mediated transcriptional activation. However we have also shown that USP7 can stimulate the assembly of EBNA1 on oriP elements in transfected plasmids suggesting that USP7 might play a role in the initial association of EBNA1 with these elements upon initial EBV infection, and/or during the switch from the EBV latency form in nonproliferating cells, in which EBNA1 is not expressed (referred to as the latency program [46]), to latency forms in proliferating cells in which EBNA1 is expressed and bound to oriP. In addition, we have shown that USP7 can stimulate EBNA1 binding to region III in the EBV genome which, under some circumstances, negatively regulates EBNA1 expression [35],[36], raising the possibility of a role for USP7 in EBNA1 autoregulation from the Qp promoter.
We have previously shown that EBNA1 residues 441–450 bind to the USP7-NTD [28],[29]. The ternary complex formed between USP7 and DNA-bound EBNA1 also appears to require the interaction of the USP7-NTD with the EBNA1 441–450 region for the following two reasons. First, the USP7-NTD was sufficient to supershift the EBNA1-DNA complex. Second, USP7 did not supershift the complex formed by DNA and EBNA1452–641, which lacks the USP7 binding site but retains full DNA binding activity. However, it is curious that we observed partial but not complete stimulation of EBNA1 DNA binding by the USP7-NTD. We had previously assessed the ability of all USP7 stable domains to bind EBNA1 by examining the retention of partially proteolysed USP7 on an EBNA1 affinity column and only the USP7-NTD was found to bind EBNA1 [28]. However, this does not eliminate the possibility that other regions of USP7 might have weak affinities for EBNA1. Our in vitro data are consistent with a model in which the USP7-NTD binds EBNA1 residues 441–450 to bring USP7 to EBNA1, enabling subsequent weaker or less specific interactions of other regions of USP7 with the EBNA1 DNA binding or C-terminal regions (452–641). This might explain why the DNA binding activity of EBNA1452–641 was weakly stimulated by USP7. Another possible interpretation of the in vitro data is that the interaction of the USP7-NTD with EBNA1 is stabilized by the rest of USP7 due to effects on the structure of the USP7-NTD. However we do not think this is likely because the USP7-NTD is a TRAF domain that is stably folded in the absence of the rest of USP7 [29],[47].
While stoichiometric amounts of USP7 were sufficient to stimulate the DNA binding activity of EBNA1, only at higher USP7 concentrations was USP7 observed to be stably associated with the EBNA1-DNA complex in vitro. This indicates that the affinity of USP7 for free EBNA1 is higher than for DNA-bound EBNA1 and that a higher effective concentration of EBNA1 or USP7 may be necessary to drive the interaction of these proteins on DNA. This conclusion is also supported by the observation that USP7 is preferentially associated with EBNA1 on the FR element over EBNA1 on the DS element of oriP. The FR element is bound by 20 EBNA1 dimers as compared to 4 EBNA1 dimers at the DS element and, in both cases, the dimers within the element interact with each other to form a larger EBNA1 complex [31],[32]. As a result the effective concentration of EBNA1 at the FR is higher than at the DS and this may drive recruitment of USP7.
An increasing number of human cellular protein binding targets of USP7 have been identified including p53, Mdm2, FOXO, March 7 and PTEN, all of which can be deubiquitylated by USP7 [20]–[24],[26]. Our proteomic profiling of USP7 protein interactions identified GMPS as another USP7 binding partner. We expect that other USP7 binding partners were not identified by this method due to their low abundance or transient nature of the interaction in response to particular stimuli (such as occurs with the USP7-p53 and USP7-FOXO interactions). The interaction of USP7 with GMPS is unique in that it appears to affect the activity of USP7 for specific substrates, as opposed to being a substrate itself. This is supported by the fact that GMPS levels are not altered when USP7 is silenced (as shown in Figure 4B).
The finding that human USP7 forms a stable complex with GMPS fits well with the observations of van der Knaap et al [27], where Drosophila USP7 was found to co-purify with GMPS. Our glycerol gradient sedimentation analyses indicated that human USP7 and GMPS form a 1∶1 complex and in vitro assays show that GMPS stimulates the ability of USP7 to deubiquitylate H2B (but not H2A), as observed for the Drosophila GMPS-USP7 complex. Van der Knapp et al [27] also showed that the stimulation of Drosophila USP7 activity by GMPS did not require the catalytic activity of GMPS. Our in vitro results are consistent with this conclusion because stimulation of USP7 deubiquitylation activity for H2B required stoichiometric amounts of GMPS (indicative of formation of a USP7-GMPS complex) and did not occur with substoichiometric amounts of GMPS (as would be expected for an enzymatic activity). Although our results are largely in agreement with those of van der Knaap et al [27], there are subtle differences in the findings of the two studies. First, Drosophila USP7 was not found to deubiquitylate H2B in vitro in the absence of GMPS while we found that human USP7 was able to cleave Ub-H2B in vitro but that this activity was stimulated by GMPS. Second, In Drosophila, GMPS was found to stimulate deubiquitylation of p53 by USP7 and we have not observed this effect with human USP7. It is presently unclear whether these discrepancies are the result of the different in vitro reaction conditions and protein concentrations or reflect genuine differences in the Drosophila and human USP7.
ChIP assays consistently showed higher recruitment of USP7 and GMPS to the oriP FR over the DS and the BZLF1 promoter region, however some degree of interaction of USP7 and GMPS was also detected at the DS and BZLF1 regions as compared to the IgG negative control. This may indicate that these proteins are wide spread on chromatin where they could regulate multiple processes that are affected by H2B ubiquitylation [48]. H2B monoubiquitylation has been reported to be associated with increased transcription through effects on both initiation and elongation [43], [49]–[51], however in some instances H2B monoubiquitylation appears to inhibit transcription [52]–[55]. Therefore the contribution of H2B monoubiquitylation to gene expression is complicated and possible contributions to other DNA processes such as DNA replication are largely unexplored. We have observed that USP7 silencing increases H2B ubiquitylation at the FR as well as at LMP1 and Cp promoters and decreases transcriptional activation from the FR element, suggesting that H2B ubiquitylation is inhibitory to transcription controlled by the FR. This is consistent with our previous observation that the EBNA1 mutant that fails to bind USP7 has decreased transcriptional activation function [14].
The increased detection of USP7 and GMPS at the FR element and their effect on Ub-H2B levels in this region, suggests that EBNA1 can employ the USP7-GMPS complex for its own purposes, at least in part by decreasing the level of Ub-H2B. In addition to functioning in transcriptional activation, the EBNA1-bound FR element mediates the segregation of the EBV episomes in mitosis [5],[6],[16],[56], may enhance DNA replication from the DS [2],[57] and causes an impediment to replication fork progression [4],[58],[59]. It is conceivable that any of these processes could be affected by the state of H2B ubiquitylation, since EBV genomes in latent infection are known to exist as nucleosomal arrays [60]. We have previously shown that EBNA1Δ395–450 that does not bind USP7 has increased DNA replication activity [14], suggesting that H2B monobiquitylation could promote DNA replication but other interpretations are also possible.
Histone modifications at oriP are just beginning to be examined and so far these studies have been focused on histone H3 acetylation and methylation of the oriP DS region. Acetylated histone H3 is generally enriched at the DS but a decrease was observed at late G1 that appears to account for the delayed replication of EBV genomes [61],[62]. Histone H3 dimethyl K4 was also enriched at the DS region while H3 methyl K9 was decreased at this region [61],[63]. Our findings indicate that monoubiquitylation of H2B is another histone modification that is modulated at oriP and that this modification is affected by EBNA1. We had previously shown that EBNA1 binding to USP7 serves to alter cellular processes in order to facilitate cell survival [29],[30]. We now present evidence that the USP7 interaction is not limited to soluble EBNA1 but also occurs with EBNA1 bound to EBV episomes where it could regulate the plasmid maintenance and transcriptional functions of EBNA1 in EBV latent infection.
EBNA1395–641 was expressed fused to a hexahistidine tag at the N-terminus in Escherichia coli from plasmid pET15b. This construct was generated by PCR amplification of EBNA1 sequences encoding amino acids 395–641 from pc3oriPEBNA1 and ligation between the Nde1 and BamH1 sites of pET15b. BL21 pLysS cells containing pET15b- EBNA1395–641 were grown to OD600nm of 0.5 then induced for 3 hrs at 37°C by the addition of IPTG (0.1 mM final concentration). Cells were lysed in 50 mM NaH2PO4 pH 8.0, 300 mM NaCl, 10 mM imidazole, 20 mM β-mercaptoethanol, 0.5 mM PMSF, 1 mM benzamidine and EBNA1395–641 was purified on Ni-NTA Agarose resin (Qiagen) then dialyzed against 50 mM Tris pH 7.5, 300 mM NaCl, 20 mM β-mercaptoethanol, 1 mM PMSF. EBNA1452–641 was purified from E.coli as previously described [64]. EBNA1 (lacking most of the Gly-Ala repeat) was purified from insect cells as described previously [14].
Full length USP7 and USP7-NTD containing amino acids 56–205 were purified as according to Holowaty et al [14]. GMPS was expressed in insect cells from a baculovirus. The GMPS baculovirus was constructed by PCR amplification of full-length GMPS cDNA in pOTB7 (ATCC number 7515509) using the primers: GCAGGATCCCATATGGCTCTGTGCAACGGAGAC (N-terminus) and GCACTCGAGTTACTCCCACTCAGTAGTTCC (C-terminus). The amplification product was digested with BamHI and XhoI and cloned between the same sites of pFastBac HT B (invitrogen). Bacmids were obtained by transformation of competent DH10Bac E. coli (invitrogen) with GMPS pFastBac HT B, then Spodoptera frugiperda (SF9) insect cells were transfected with the bacmids to generate the baculovirus according to manufacturer's specifications. Culture media containing the baculovirus was harvested 5 days post-transfection and amplified twice. To generate GMPS for purification, ten 15 cm plates of High Five cells at 80% confluency were infected with the GMPS baculovirus. Cells were harvested 50 hrs post-infection, washed with PBS and lysed in 10 mls of 20 mM Tris-HCl pH 8, 0.5 mM DTT, 0.5 mM EDTA, 10% glycerol and complete protease inhibitor cocktail (Roche). The lysate was sonicated, incubated 30 min on ice, then clarified by centrifugation at 64,000×g for 15 min at 4°C. The clarified lysate was incubated with 250 µl of a nickel resin (Sigma) for 1 h (with rotation) then transfered to a column. The resin was washed 3 times with 4 column volumes of column buffer (50 mM NaH2PO4, 300 mM NaCl and 10 mM imidazole) and the His-tagged GMPS was eluted from the column with column buffer containing 250 mM imidazole. EDTA and DTT were added to the elutions to a final concentration of 10 mM and the eluted protein was dialyzed overnight against 50 mM HEPES pH 7.9, 50 mM NaCl, 10% glycerol, 0.1 mM EDTA and 0.1 mM DTT then stored in aliquots at −80°C.
DNA probes for EBNA1 EMSAs were generated by end-labeling a 20-mer oligonucleotide corresponding to site 1 of the DS element (5′-CGGGAAGCATATGCTACCCG-3′) with γ-32P-ATP and annealing it to its complementary sequence. In assays containing EBNA1 and either USP7 or GMPS, EBNA1 was preincubated with USP7 or GMPS at room temperature (RT) for 10 minutes prior to adding the labeled DNA, except in Figure 2C, where EBNA1 was incubated with labeled DNA for 10 minutes at RT first, followed by addition of increasing amounts of USP7 and further incubation at RT for 10 minutes. In Figures 1 and 2A, 10 pmols of USP7 was used along with the indicated amounts of EBNA1. For samples containing EBNA1 and both USP7 and GMPS, USP7 and GMPS were preincubated together at 4°C for 5 minutes before the addition of EBNA1 and further incubation at RT for 10 minutes. The EMSAs in Figure 7 used 2 pmol EBNA1 dimer and 64 pmols of USP7 and GMPS. Protein mixtures were incubated with 10 fmoles of labeled DNA at RT for 10 minutes in the presence of 1 µg salmon sperm DNA in 20 µl binding buffer (20 mM Tris pH 7.5, 200 mM NaCl). 4 µl of 6× DNA Loading Dye (10 mM Tris-HCl pH 7.6, 0.03% bromophenol blue, 0.03% xylene cyanol FF, 60% glycerol, 60 mM EDTA; MBI Fermentas, R0611) was then added to the reactions prior to electrophoresis on a 10% polyacrylamide gel. Bands were visualized by autoradiography.
Purified USP7 was covalently coupled to Affi-Gel 10 (Bio-Rad) at concentrations of 0, 0.5, 1 or 2 mgs per ml of resin in 50 mM HEPES pH 7.5, 50 mM NaCl, 1 mM DTT, 5% glycerol. The resin was then blocked in ethanolamine, equilibrated in column buffer (50 mM HEPES pH 7.5, 100 mM NaCl, 1 mM DTT, 0.1 mM EDTA, 10% glycerol) and used to generate 40 µl microcolumns as previously described [14],[65]. Whole HeLa cell lysates were generated as in Holowaty et al [14] and equal amounts were applied to each microcolumn. The columns were washed in column buffer then sequentially eluted in column buffer containing 1 M NaCl then the same buffer containing 1% SDS. Column eluates were analysed by SDS-PAGE and silver staining. The band running at 70 kDa was excised and prepared for MALDI-ToF mass spectrometry analysis as previously described [14]. Recovered peptides were analysed on a Voyager DE-STR instrument (Applied Biosystems) and the protein was identified by mass fingerprinting using ProFound software.
50 µg of purified USP7 was incubated with 25 µg of purified GMPS in a total volume of 25 µl of 50 mM HEPES pH 7.9, 50 mM NaCl, 10% glycerol, 0.1 mM EDTA, 0.1 mM DTT for 1 hour at room temperature. Control samples were also generated in which USP7 or GMPS were incubated individually. The mixtures were then diluted to 500 µl in 50 mM HEPES pH 7.9, 5% glycerol, 200 mM NaCl and 0.5 mM EDTA and loaded onto 11.5 ml 10%–20% glycerol gradients formed in the same buffer. Gradients were subjected to centrifugation in a SW41 rotor at 34,000 rpm for 18 hours at 4°C. Fractions of 500 µl were collected from the top of each gradient and 30 µl of each fraction was analyzed on an 8% SDS-polyacrylamide gel. Proteins were visualized by colloidal blue staining. Aldolase and catalase were analyzed on identical gradients as size markers.
Histones for in vitro assays were prepared by acid extraction as described by Kao and Osley [66]. Briefly, HeLa cells at 70% confluence were lysed in 10 mM HEPES pH 7.9, 1.5 mM MgCl, 10 mM KCl, 0.5 mM DTT, 1.5 mM PMSF and 1 mM NEM, then hydrochloric acid was added to a final concentration of 0.2 M. The lysate was incubated on ice for 30 min, then subjected to centrifugation at 10,000×g for 10 min at 4°C. The supernatant fraction, containing the histones, was dialyzed against 0.1 M acetic acid, then against distilled water and store at −70°C. Prior to use, the histones were diluted to 1 mg/ml and adjusted to a final concentration of 50 mM HEPES pH 7.9, 100 mM NaCl and 1 mM DTT. 200 µg of histones were incubated at 37°C with 0.2 µg USP7, with or without 0.1 µg GMPS (1∶1 USP7∶GMPS), 0.01 µg GMPS (10∶1 USP7∶GMPS) or 10 µg GMPS (1∶10 USP7∶GMPS) as indicated in a 200 µl reaction. Samples were collected at the indicated times and mixed with SDS-PAGE loading buffer to stop the reactions. Samples were analysed by electrophoresis on 15% SDS-polyacrylamide gels and the levels of ubiquitinated H2B and H2A were visualized by Western blotting using antibodies against H2B (Upstate Biochemicals), H2A (Upstate Biochemicals) and ubiquitin (Sigma).
Ubiquitylated p53 was generated by in vitro reactions with Mdm2. To this end human p53 and Mdm2 were cloned into pET15b (Novagen), expressed in E.coli and purified by virtue of the hexahistidine tag using standard metal affinity purification procedures. P53 was ubiquitylated in vitro as previously described [67]. Briefly, 5 µg p53 and 5 µg Mdm2 were incubated for 90 min at 30°C with 500 ng E1 (Calbiochem), 1 µg UbE2D2 (Boston Biochem), 50 µg ubiquitin (Boston Biochem) in a 200 µl reaction mixture containing 50 mM Tris pH 7.6, 5 mM MgCl2, 2 mM ATP, 2 mM DTT. Ubiquitylation was confirmed by Western blot analysis of a 10 µl sample using p53 monoclonal antibody PAb1801 [68] and the remaining mixture was stored at −80°C. For deubiquitylation assays, 5 µg (10 µl), poly-ubiquitylated p53 was incubated with 0.5 µg USP7 with or without 5 µg of GMPS in an 10 µl reaction. The samples were collected at the indicated time points, mixed with SDS-PAGE loading buffer and subjected to 10% SDS-PAGE. p53 was detected by Western blotting using p53 antibody PAb1801.
HeLa cells were transfected 3 times during a seven day period with USP7 siRNA (100 pmols, 200 pmols and 200 pmols, respectively) or with negative control siRNA against GFP [30] using Lipofectamine 2000 (Invitrogen). USP7 siRNA sequence was CCCAAATTATTCCGCGGCAAA as described in Tang et al 2006 [69]. Cells were then harvested and split into two equal samples. One sample was used to verify USP7 silencing by Western blotting using rabbit serum against USP7 [30] and anti-actin antibody (Calbiochem) as a loading control. The other sample was used to isolate the histones by acid extraction as described above and to quantify the levels of ubiquitylated histones H2B and H2A by Western blotting for these histones as described above. In each case, the amount of ubiquitylated histone was determined by normalizing the intensity of this band to that of the unmodified histone band (set to 1).
ChIP assays were performed for GMPS and USP7 in the EBV-positive, Raji Burkitt's lymphoma cells as previously described [16] using anti-USP7 rabbit antibody (Bethyl Laboratories.Inc) or rabbit antiserum raised against full length recombinant GMPS purified from insect cells. Rabbit IgG (Santa Cruz) and anti-EBNA1 R4 rabbit antibody [14] were also used as negative and positive controls, respectively. Quantitative real-time PCR was performed with a Platinum SYBR Green qPCR superMix-UDG (Invitrogen) in a Rotorgene qPCR System (Corbett Research), using 1/50th of the ChIP samples or 1/2500th of DNA samples prior to immunoprecipitation (input) and the previously described primer sets for the DS and FR elements and the BZLF1 promoter region [16]. Values obtained for ChIP samples were normalized to input samples with the same primer sets. For ChIP assays involving USP7 depletion, D98/Raji cells [42] were subjected to three rounds of transfection (every 24 hours) with siRNA against USP7 or with siRNA against GFP as described above. Samples were prepared as for the ChIP experiments in Raji cells except that antibodies against EBNA1, histone H2B (Upstate Biochemicals) and mono-ubiquitylated histone H2B (MediMabs Inc, Montreal) were used. Primer sets used to assess recovery of the LMP1 promoter region were CAATCAGAAGGGGGAGTGCG and ACAGCCTTGCCTCACCTGAAC, of Cp promoter region were AACCTTGTTGGCGGGAGAAG and GGCGAATTAACTGAGCTTGCG, and of oriLyt region were CGTCTTACTGCCCAGCCTACT and AGTGGGAGGGCAGGAAAT. Experiments examining EBNA1 binding to region III used the primer sets GACCACTGAGGGAGTGTTCCACAG and ACACCGTGCGAAAAGAAGCAC described in Yoshioka et al [36].
CNE2Z cells [70] were plated in 6 cm dishes and transfected with 50 pmols of siRNA against GFP or siRNA against USP7. siRNA transfections were repeated twice at 24 hour intervals for a total of 3 rounds of siRNA transfection over 72 hours. Cells were then moved to 10 cm dishes and transfected with 5 µg of pc3OriP, pc3OriPEBNA1 or pc3OriPΔ395–450 and 250 ng pLacZ plasmid containing LacZ cDNA. 24 hours post-transfection, cells were fixed with 1% formaldehyde, lysed in RIPA buffer (20 mM Tris pH 8.0, 150 mM NaCl, 1% NP40, 0.1% Sodium Deoxycholate, 1 mM PMSF) containing protease inhibitor cocktail (Sigma, P8340) and sonicated briefly to shear the DNA. Clarified lysates were precleared with Protein A/G beads (Santa Cruz, SC-2003) prior to immunoprecipitation with EBNA1 R4 antibody and normal rabbit IgG (Santa Cruz, SC-2345). Protein cross links were reversed in the immunoprecipitated DNA by incubating at 65°C for 16 hrs. DNA was purified using QIAquick Gel Extraction Kit (Qiagen, 28704) and analyzed by quantitative RT-PCR using LightyCycler 480 DNA SYBR Green I Master (Roche, 04707516001) and a Rotorgene Q-PCR system (Corbett Research). Primers used for DS are as described above. Primers used for FR and lacZ quantification were CCCGGATACAGATTAGGATAGC and TGTTGCCATGGGTAGCATA for FR and ATATTGAAACCCACGGCATGGTGC and TTTGATGGACCATTTCGGCACAGC for lacZ.
EBNA1 transactivation assays were performed as described previously [71] with the following modifications. CNE2Z cells were transfected with siRNA against GFP or USP7 as described above, then were moved to 10 cm dishes 24 hour prior to transfection with 2 µg of pFRTKCAT reporter construct (kindly provided by Bill Sugden) and 180 ng of pc3OriP or pc3Orip containing expression cassettes for EBNA1 [15] or EBNA1Δ395–450. 48 hrs later, cells were harvested and lysed using three rounds of freezing and thawing. 15 µg of total protein from each sample was assayed for chloramphenicol acetyltransferase activity using several reaction times and results from a point in the linear range was reported.
|
10.1371/journal.pbio.1000382 | Dynamic Assignment and Maintenance of Positional Identity in the Ventral Neural Tube by the Morphogen Sonic Hedgehog | Morphogens are secreted signalling molecules that act in a graded manner to control the pattern of cellular differentiation in developing tissues. An example is Sonic hedgehog (Shh), which acts in several developing vertebrate tissues, including the central nervous system, to provide positional information during embryonic patterning. Here we address how Shh signalling assigns the positional identities of distinct neuronal subtype progenitors throughout the ventral neural tube. Assays of intracellular signal transduction and gene expression indicate that the duration as well as level of signalling is critical for morphogen interpretation. Progenitors of the ventral neuronal subtypes are established sequentially, with progressively more ventral identities requiring correspondingly higher levels and longer periods of Shh signalling. Moreover, cells remain sensitive to changes in Shh signalling for an extended time, reverting to antecedent identities if signalling levels fall below a threshold. Thus, the duration of signalling is important not only for the assignment but also for the refinement and maintenance of positional identity. Together the data suggest a dynamic model for ventral neural tube patterning in which positional information corresponds to the time integral of Shh signalling. This suggests an alternative to conventional models of morphogen action that rely solely on the level of signalling.
| In many developing tissues, the pattern in which cell types are generated depends on secreted factors called morphogens. These signalling molecules are produced in specific locations and at specific concentrations, thereby forming concentration gradients. Different target genes are induced at specific distances from the source of the morphogen, and therefore the spatial pattern of gene expression correlates with this concentration gradient. In this study, we examined how cells respond to a morphogen gradient to produce the appropriate pattern of cellular differentiation. We focused on the morphogen Sonic Hedgehog (Shh), which specifies the pattern of different types of neurons in the ventral regions of the neural tube (the embryo's precursor to the central nervous system). We show that in addition to the concentration of Shh, the duration of Shh signalling also contributes to the patterning of the ventral neural tube. A consequence of this is that the genes defining different cellular identities are expressed in a characteristic temporal progression. In addition, sustained Shh signalling is required for more ventral cell types; otherwise they revert to their previous cellular identity. Together these results indicate that dynamic and sustained signalling by Shh is required for the patterning of the ventral neural tube, challenging conventional models of morphogen action that rely solely on the concentration of signal perceived by cells at specific positions in the morphogen gradient.
| A defining feature of embryogenesis is the specification of a large variety of cell types in stereotypical spatial and temporal patterns. A common mechanism by which this is achieved involves the deployment of morphogens [1],[2],[3]. These secreted molecules are proposed to establish signalling gradients within developing tissue that provide the positional information that guides the pattern of gene expression and cellular differentiation. Of central importance, therefore, is to understand the nature of the positional information produced by the morphogen.
Several families of secreted proteins, including members of the Hedgehog (Hh), Transforming Growth Factor beta (TGF-β), and Wingless (Wnt) families, operate as morphogens during embryonic development [2],[4]. One example where progress has been made in understanding the mechanism of morphogen action is Sonic Hedgehog (Shh) signalling in the vertebrate central nervous system [5],[6],[7]. Shh is produced from the ventral midline of the neural tube and underlying notochord [8],[9],[10],[11],[12],[13] and spreads to form a ventral-to-dorsal gradient [14]. Within responding cells, Shh signalling regulates the activity of Gli transcription factors (Gli1, 2, and 3) to produce a net increase in their transcriptional activator function [15],[16],[17],[18],[19]. This, in turn, controls the expression of a set of transcription factors in ventral progenitor cells that subdivide the neuroepithelium into five molecularly distinct domains arrayed along the dorsal-ventral (DV) axis [5],[20]. Each progenitor domain generates one of five different neuronal subtypes; from dorsal to ventral, the domains are termed p0, p1, p2, pMN, and p3 and generate V0, V1, V2 neurons, motor neurons (MNs), and V3 neurons, respectively [5],[20],[21].
The two most ventral neural progenitor domains, p3 and pMN, are defined by the expression of the transcription factors Nkx2.2 and Olig2, respectively [22],[23]. In response to Shh signalling, Olig2 is expressed first, in a small group of ventral neural progenitors [14],[24]. Its expression then gradually expands dorsally. As this happens, the ventral cells that originally expressed Olig2 induce Nkx2.2 expression and downregulate Olig2. Nkx2.2 expression then also expands dorsally, downregulating Olig2 in its wake [22],[23],[24], however the dorsal expansion of Nkx2.2 is more limited than Olig2. The end result of the sequential induction and expansion of these two factors is two adjacent but spatially distinct progenitor domains: Olig2 expressing pMN progenitors located dorsal to Nkx2.2 expressing p3 cells. Importantly, the order of appearance and the final position of these progenitor domains correspond to their requirement for increasing concentrations of Shh [21],[24],[25]—the more ventrally located progenitor domain emerges later and requires higher Shh concentrations. One explanation for this, which would accord with the conventional model for morphogen interpretation [1],[3], is that Shh concentration increases over time [14]. As a consequence of this, the threshold concentration for Nkx2.2 activation would be reached later than that of Olig2 [14]. However, our previous work suggested a more complex mechanism of morphogen gradient interpretation in which the signalling pathway does not linearly transduce Shh concentrations required for Olig2 and Nkx2.2 induction [24].
In this non-canonical model of morphogen interpretation, not only the level of intracellular signalling but also its duration plays a crucial role. In vitro experiments indicate that signal transduction is not linearly proportional to Shh concentration at the concentrations of Shh that induce Nkx2.2 and Olig2 [24]. In fact, initially in cells exposed to these concentrations of Shh, signal transduction is saturated resulting in similarly high levels of intracellular Gli activity. However, cells gradually adapt their response to Shh so that they become increasingly less sensitive to the ligand. This means that the time for which intracellular signalling is maintained above a particular threshold is proportional to the extracellular concentration of Shh. Consequently, the concentration of Shh that induces Nkx2.2 maintains high levels of intracellular signalling for a longer period of time than the lower, Olig2-inducing concentration. The gradual adaptation of cells to Shh is controlled, at least in part, by the Shh mediated upregulation of Ptc1, the receptor for Shh that is also a negative regulator of the pathway [26],[27],[28],[29],[30],[31]. High Shh concentrations bind to and suppress more Ptc1 than lower Shh concentrations, thereby sustaining intracellular signal transduction for longer [32],[33]. Importantly, a longer period of signalling is required for the induction of Nkx2.2 than for Olig2. Together, therefore, the data indicate that the differential response to the concentrations of Shh that induce Nkx2.2 or Olig2 is a function of the duration of intracellular signalling rather than the levels of signalling. Hence, at these concentrations, increasing the extracellular concentration of Shh results in an increase in the duration of intracellular signalling rather than a change in the level of signalling. We termed this mechanism “temporal adaptation” [24].
Whether this mechanism is relevant for the provision of positional information throughout the concentration range of Shh in the ventral neural tube remains an open question. Moreover, the kinetics of Shh signalling that assign positional identity are unclear, as is the length of time that Shh signalling is required to maintain positional identity once established. To address these issues we first focused on the specification of V0, V1, and V2 neurons, which arise from progenitors located dorsal to the pMN and require lower concentrations of Shh than MNs and V3 neurons [34]. We show that these concentrations of Shh do not saturate the signal transduction pathway and Gli activity levels are proportional to Shh concentration. Nevertheless, we provide evidence that the duration of Shh signalling plays an important role in the specification of V0–V2 neurons, such that increasing durations of signalling promote the generation of more ventral identities. Moreover, we provide in vivo evidence that the assignment of progenitor identity in this region of the neural tube is progressive. Together these data indicate that the duration of Shh signalling is important for all ventral progenitors to acquire their correct positional identity. We further show that sustained Shh signalling is required to maintain appropriate progenitor pattern in the ventral neural tube even after positional identity has been induced. Progenitors revert to antecedent identities if signalling is interrupted. Thus the allocation of cell identity in the ventral neural tube appears more plastic than for other well-described morphogen patterned tissues and we discuss the implications for models of tissue patterning by morphogen gradients. Together, the data suggest a model for ventral neural tube patterning in which positional identity of progenitors is dynamic and determined by the level and duration—the time integral—of Shh signalling.
Intracellular signal transduction by Shh results in increased transcriptional activity of the Gli family of zinc finger transcription factors [19]. Our previous studies demonstrated that exposure of neural cells to concentrations of Shh in excess of 1 nM, which induce p3 and pMN progenitors, saturate intracellular signalling and generate similar high levels of Gli activity during the first 12 h [24]. Lower concentrations of Shh are associated with the production of several subtypes of interneurons, dorsal to MNs, including V0, V1, and V2 neurons [34]. We therefore assayed the profile of Gli activity induced by lower concentrations of Shh. For this, intermediate neural plate ([i]) explants from Hamburger and Hamilton (HH) stage 10 chick embryos in ovo electroporated with a Gli reporter (GBS-luc; Figure 1A; see Materials and Methods; [24]) and normalization plasmids were cultured in the presence of 0.1–2 nM Shh. The level of Gli activity in these explants was measured 6 h after the start of culture and compared to the levels of Gli activity in explants cultured in the absence of Shh (Figure 1B). These data indicated that saturating levels of signal transduction, corresponding to Gli activity ∼25-fold higher than basal levels, were reached at ∼1 nM Shh (Figure 1B). This is consistent with our previous findings [24]. At concentrations less than 1 nM Shh, however, the level of Gli activity was a function of Shh concentration and half maximal levels of Gli activity (10–15-fold induction) were obtained between 0.25 nM and 0.5 nM Shh (Figure 1B).
We next compared the dynamics of Gli activity in explants exposed to 0.25 nM and 2 nM Shh for 6–24 h (Figure 1C). For both concentrations, maximum levels of Gli activity were recorded at ∼6 h and subsequently decreased over time. At 2 nM Shh, Gli activity level at 6–12 h was similar to that generated by 1 nM Shh. In contrast to 1 nM Shh, however, cells exposed to 2 nM Shh contained higher levels of Gli activity at 18–24 h, indicating that the rate of signalling decline is inversely correlated with morphogen concentration. This is consistent with the “temporal adaptation” mechanism [24]. At 0.25 nM, Shh signalling pathway was not saturated. Thus levels of Gli activity reached in this condition were lower; nevertheless the level of signalling declined over time and was indistinguishable from pre-stimulus levels by 18–24 h. Consequently, lower concentrations of Shh sustain signal transduction above basal levels for shorter periods of time (Figure 1C, inset). Together these data suggest that exposure of neural cells to Shh generates distinctive temporal profiles of intracellular signalling. Gli activity reaches a peak within ∼6 h of exposure to Shh, and these peak levels correspond linearly with Shh concentration for low concentrations and are saturated at high concentrations. Subsequently, for all concentrations, intracellular signalling declines from the peak as cells adapt and the length of time it takes for Gli activity to return to pre-stimulus levels increases with Shh concentration.
We next assayed how the observed dynamics of Shh signalling regulate the specification of different neuronal subtypes. We focused our attention on Shh concentrations <1 nM and the generation of V0, V1, and V2 neurons within the intermediate region of the neural tube (Figure 2A) [34]. Explants were exposed to 0.05–0.5 nM Shh for 48 h, then assayed for Evx1, En1, Chx10, and MNR2 expression, markers of V0, V1, and V2 neurons and MNs, respectively (Figure 2B, S1A, S1B, S1C, S1D, S1E, S1F; [35],[36],[37],[38]). In the absence of Shh, most cells express Pax7 (Figure S2; [39]) and few, if any, cells expressed V0–V2 or MN markers (Figure 2B). Addition of 0.05–0.3 nM Shh resulted in a reduction in Pax7 expression, induction of the intermediate progenitor marker Dbx1 (Figure S2), and the production of cells expressing Evx1, En1, or Chx10 (Figure 2B). There was a trend for the induction of increasing numbers of more ventral cell types as Shh concentration was increased. For instance, the production of V0 neurons peaked at ∼0.1 nM while the maximum numbers of V1 neurons were observed at ∼0.25 nM Shh and MNs were only induced in significant numbers by concentrations of Shh >0.4 nM Shh (Figure 2B). Nevertheless, a mixture of neuronal subtypes was generated in response to all Shh concentrations tested in the range of 0.05–0.4 nM. Thus, these data indicate that, although low concentrations promote the generation of V0–V2 neurons, incremental differences in concentration are not sufficient to generate distinct neuronal subtypes. Since the Shh concentrations used in these experiments produced distinct levels of Gli activity (Figure 1B), the data suggest that the level of signalling is not sufficient to separate the specification of V0, V1, and V2 neurons.
As different Shh concentrations could not delineate clear transitions in the production of different neuronal subtypes, we asked whether the length of time cells are exposed to Shh could distinguish the generation of ventral neurons. For this, explants were exposed to 0.5 nM Shh for fixed times ranging from 6 h to 48 h. After the indicated period of Shh exposure, Shh was removed and replaced with media lacking Shh; neuronal subtype identity was then assayed in all explants at 48 h (Figure 2C, 2D). Exposure to Shh for only the first 6 h of the 48 h culture period was sufficient to induce V0 and V1 neurons, but few, if any, V2 neurons were generated (Figure 2D). Increasing the duration of Shh signalling resulted in the progressive generation of more ventral cell types: 12 h exposure resulted in V2 generation and a marked reduction in V0 generation, while 18 h exposure was required for the appearance of MNs (Figure 2D). Thus different durations of signalling influenced the neuronal subtype generated in response to a fixed Shh concentration such that more ventral neuronal subtypes were generated in explants exposed for longer periods of time. Notably, different durations of Shh signalling generated well-separated peaks of V0 and V2 neurons, in contrast to the overlapping peaks of neuron production following exposure to different Shh concentrations (compare Figure 2B with 2D). Nevertheless there was still significant overlap in the generation of V0 and V1 neurons (Figure 2D). This might reflect limitations in the resolution of the explant assay or the action of additional signals in the generation of these cell types [40].
To confirm the involvement of signal duration in the assignment of positional identity, we assayed the expression of transcription factor markers of p0–p2 progenitors, which generate V0–V2 neurons. Explants were assayed at 48 h after exposure to 0.5 nM Shh for different periods of time (Figures 2C, 2E, S1G, S1H, S1I, S1J). V0 neurons are generated from progenitors that express Dbx1 ventral to the Pax7 boundary; progenitors of V1 neurons express Pax6 but lack Nkx6.1 and Dbx1 expression; V2 progenitors express Nkx6.1 but not Olig2 (Figure 2A; [41]). Consistent with the profile of neuronal subtype generation, increasing the duration of Shh signalling resulted in a gradual ventralization of progenitor cells. In the absence of Shh, progenitor cells expressed Pax7. Exposure to Shh for 6 h induced Dbx1 expression and decreased the number of Pax7 expressing cells by ∼40% (Figure 2E). Longer periods of exposure resulted in a gradual decrease in Dbx1 and Pax7 expression and an increase in the expression of Nkx6.1; finally Olig2 expression, the pMN marker, was detected (Figure 2E). Thus, similar to neuronal subtype identity, the response of the transcriptional markers of progenitor domains to different durations of Shh exposure corresponds to their DV position in the neural tube.
We sought to rule out the possibility that the positional identities induced by different times of Shh exposure were the result of temporal changes in the competence of progenitors. If this were the case, the time at which cells received Shh signalling, rather than duration for which they were exposed to Shh, would determine positional identity. We therefore assessed the effect of adding Shh to explants at different time points. Explants were cultured in the absence of Shh for 12–24 h and then exposed to 0.5 nM or 4 nM Shh for an additional 6–24 h (Figure S3). Assaying the expression of Dbx1, Nkx6.1, Olig2, and Nkx2.2 revealed that the response of explants was offset by the same amount of time that the addition of Shh was delayed. For example, Dbx1 and Nkx6.1 were induced by exposure to 0.5 nM Shh for 6 h and 18 h, respectively, regardless of whether Shh exposure was initiated at 0 h or 24 h after the start of the culture period (Figure S3A). Similarly, 12 h of exposure to 4 nM Shh induced Olig2, while longer times were required for Nkx2.2 induction, whether Shh was added immediately after explanting or following 12–24 h ex vivo (Figure S3B). These data argue against an intrinsic timing mechanism that over time changes the competence of progenitors to generate different neuronal subtypes. Instead the data provide strong support for the idea that the duration of Shh signalling plays a central role in determining positional identity. Thus, throughout the ventral neural tube, progressively more ventral fates are generated as the concentration and time of exposure to Shh are increased.
To test directly whether progenitors progressively adopt more ventral identities as the duration of Shh exposure is increased, we assayed the dynamics of Dbx1, Nkx6.1, and Olig2 expression in [i] explants exposed to 0.5 nM Shh for 6–36 h (Figure 3A). At 18 h, explants expressed the three markers (Figure 3B). Exposure to 0.5 nM Shh for longer periods of time resulted in the gradual increase in the number of Olig2 and Nkx6.1 expressing cells and a reduction in Dbx1 expression and only a small number of V0–V2 neurons were present in explants exposed for 36 h to 0.5 nM Shh (Figures 3B, S4A, S4B, S4C). These data suggest that continued Shh signalling after 18 h promotes the acquisition of more ventral positional identities at the expense of intermediate fates. Notably, by 18 h the level of Shh signalling in cells exposed to 0.5 nM Shh was significantly reduced from the peaks attained at 6–12 h (Figure 1C; [24]). Nevertheless, this level of signalling appears sufficient for the progression of progenitor identity from intermediate to pMN.
A consequence of the sequential establishment of positional identity is that the ventral limit of expression of a progenitor marker such as Dbx1, which identifies p0 progenitors, should be displaced dorsally as development proceeds. In this view, cells that have expressed Dbx1 will contribute to progenitor domains ventral to p0. To test this, we took advantage of a Dbx1-Cre mouse line [42] and the inducible reporter allele ROSA26floxSTOPfloxYFP (Figure 3C; Materials and Methods; [43]). In these embryos the progeny of any cell that has expressed Dbx1 is indelibly marked by the expression of YFP. In embryos assayed at E11.5, the ventral limit of YFP expression was more ventral than the extant p0 domain, identified by Dbx1 (unpublished data) and Cre recombinase expression (Figure 3D), and some Dbx1 progeny were detectable within Olig2 positive progenitors (Figure 3E). Thus, Dbx1 is initially expressed in more ventral progenitors and its expression is refined, eventually becoming restricted to the p0 domain. These data provide evidence that ventral neural tube patterning proceeds via the progressive assignment of ventral identities.
The progressive assignment of positional identity by Shh signalling prompted us to question whether continued signalling is required to maintain specific ventral identities once they are induced. To test this, we blocked Shh signalling in explants at different time points using the inhibitor cyclopamine [32],[44], an antagonist of Smoothened, the essential intracellular transducer of the pathway [45],[46]. After explants had been exposed to 0.5 nM Shh for 18 h, they were cultured for an additional 6 h, 12 h, or 18 h in fresh media containing 500 nM cyclopamine (Figures 4A–4C, S3D). Explants that had been exposed to 0.5 nM Shh for 18 h contained significant numbers of Dbx1, Nkx6.1, and Olig2 expressing cells (Figure 3B). However, when Shh signalling was inhibited from 18 h, a marked reduction in the expression of ventral progenitor markers Nkx6.1 and Olig2 was apparent at 36 h (Figure 4B, 4C), compared to explants exposed to 0.5 nM Shh for 18 h (Figure 3B) or for 36 h (Figure 4B, 4C). Consistent with this, MN production was reduced when signalling was blocked at 18 h (Figure S4B, S4C). Notably, Olig2 expression was more severely affected than Nkx6.1 (Figure 4C). This suggests that Olig2 is more sensitive than Nkx6.1 to removal of signalling and that the remaining Nkx6.1 expressing cells in these explants represent p2 progenitors. Concomitant with the decrease in ventral progenitor markers, the numbers of Dbx1 expressing progenitors and V0–V2 neurons increased (Figure 4C, S4B, S4C). This change in progenitor identity inversely correlated with the concentration of Shh required for induction and indicates that cells gradually revert to a more dorsal identity when Shh signalling is blocked. These data suggest, therefore, that Shh signalling is required not only to induce ventral identity but also to maintain the identity after it has been assigned.
We asked whether the requirement for Shh signalling to maintain progenitor identity is a general property of the response of neural progenitors and is therefore also observed in cells exposed to higher Shh concentrations. For these experiments, explants were exposed to 4 nM Shh for 18 h, and then placed into fresh media containing either 4 nM Shh, or lacking Shh, or containing 4 nM Shh and 500 nM cyclopamine, to block signalling. Incubation of explants was continued for either an additional 6 h, at which time point the level of Gli activity assayed, or an additional 18 h in order to assay Olig2, Pax6, and Nkx2.2 expression (Figure 4D). Six hours after the addition of 500 nM cyclopamine, Gli activity levels had returned to basal levels (Figure 4E, iv). By contrast, 6 h after removal of Shh from the medium, the levels of Gli activity were significantly lower but remained 2–3-fold above basal levels (Figure 4E, iii). After 18 h exposure to 4 nM Shh, explants expressed a mixture of Olig2 and Nkx2.2 (Figure 4F, i; see also [24]); a few cells expressed both Nkx2.2 and Olig2 but the majority of cells expressed one or the other marker. In explants in which exposure to 4 nM Shh was maintained for 36 h, Nkx2.2 expression was consolidated and Olig2 was downregulated (Figure 4F, ii). By contrast, in explants in which Shh was removed and incubation continued for an additional 18 h in the absence of Shh, most cells reverted to expressing Olig2 alone and Nkx2.2 had largely been downregulated (Figure 4F, iii). This indicates that Gli activity levels 2–3-fold above basal are sufficient to maintain Olig2, but not Nkx2.2 expression. The complete inhibition of Shh signalling with 500 nM cyclopamine resulted in the downregulation of both Nkx2.2 and Olig2 (Figure 4F, iv). These explants expressed high levels of Pax6, a marker of the intermediate neural tube (Figure 4F, iv; [25]). These data demonstrate that continuous Shh signalling is required for maintaining the identity of all ventral neural progenitors. In addition, it suggests that the maintenance of different target genes requires different thresholds of Gli activity.
These data provide an explanation for the discrepancy in the identity of progenitor and post-mitotic cells observed in the data from initial experiments using [i] explants (Figure 2D, 2E). In these experiments, the subtype identity of neurons generated by a specific duration of signalling did not always correspond to the progenitor markers expressed in the same conditions. For example, 18–24 h exposure to Shh was sufficient to induce large numbers of MNs in explants assayed at 48 h, however there was no detectable Olig2 expression in explants assayed in the same conditions. This is consistent with the idea that pMN progenitors were present in these explants at an earlier time point, but by the time the assay was performed, these progenitors had disappeared, leaving behind the apparently anomalous post-mitotic neurons they had generated. Together, these data identify a function for Shh signalling in maintaining progenitor identity and suggests a dual role for Shh signalling in the neural tube: first to assign progenitor identity and then to maintain the established positional identity of ventral progenitor cells.
The ex vivo data prompted us to ask whether an extended period of Shh signalling is required in vivo to maintain appropriate patterns of gene expression. We first used cyclopamine to block Shh signalling by administering the drug in ovo to HH st.18 embryos. At this developmental stage, corresponding to ∼36 h after the time at which intermediate regions are explanted from embryos, progenitor domains are well established (Figure 5A) and neurogenesis is under way. Expression of progenitor markers of the ventral neural tube was then analyzed 24 h later, after administration of cyclopamine or vehicle alone; this corresponded to ∼HH st.22. At thoracic levels, altered ventral patterning was observed in 80% of the embryos exposed to cyclopamine (Figure 5B). In these embryos, the number of Nkx2.2+ cells was markedly reduced and the domain of Olig2 shifted ventrally (Figure 5B, ii). Furthermore, the pMN domain was affected in 60% of the samples, with fewer cells expressing Olig2 (n = 12; Figure 5B, iii). Thus, similar to the ex vivo results, continued Shh signalling appears to be required to maintain the appropriate pattern of progenitor identities in the ventral neural tube, and if Shh signalling is interrupted cells transform into a more dorsal identity.
To corroborate these data, we blocked Shh signalling, in ovo, by electroporation of Ptc1Δloop2, a dominant active version of the Shh receptor that cell autonomously blocks intracellular Shh signal transduction (Figure 5C; [47]). Embryos transfected with Ptc1Δloop2 at HH st.18 were assayed for the expression of progenitor markers of the ventral neural tube 24 h later, at ∼HH st.22. Similar to the results of cyclopamine exposure, the inhibition of Shh signalling with Ptc1Δloop2 resulted in the inhibition of Nkx2.2 and Olig2 expression (Figure 5C). The effects of Ptc1Δloop2 appeared cell autonomous, as only transfected cells displayed obvious changes in gene expression, consistent with the cell autonomous inhibition of Shh signalling imposed by Ptc1Δloop2 [47].
To test whether an extended period of Shh signalling is also necessary in the mouse neural tube, we deleted Shh from the floor plate at ∼e9.5. Mice containing a conditional null allele for Shh (Shhflox/flox; [48]) and a Brn4cre transgene (Bcre32 [Tg(Pou3f4-cre)32Cre]; [49]), which directs cre expression to neural progenitors from ∼e9.0, were used. We confirmed that Shh is lost from the cervical neural tube of Shhflox/flox;Brn4cre embryos by e10.5 but left notochord expression of Shh unaffected (Figure S5A, S5B). Moreover, in contrast to embryos lacking Gli2 [50], in which the notochord remains abutting the neural tube, in Shhflox/flox;Brn4cre embryos the notochord separated normally from the overlaying neural tissue. Assaying expression of Nkx2.2 and Olig2 at e9.5 revealed that the induction and extent of expression of both proteins were similar in Brn4cre;Shhflox/flox embryos and littermate controls (Figure 5D, 5E). By contrast, analysis of e10.5 Brn4cre;Shhflox/flox embryos revealed a marked decrease in the number of cells expressing Nkx2.2 and Olig2 and a ventral retraction in the domains of expression (Figure 5F, 5G). Notably, the expression of Olig2 expression was more affected than Nkx2.2. In addition the number of cells expressing Nkx6.1 was decreased; concomitantly there was a ventral shift in the expression of Dbx1 and Pax6 (Figure S5C, S5D; unpublished data). By e12.5 embryos, the expression of both Nkx2.2 and Olig2 was severely affected in Brn4cre;Shhflox/flox and only a few cells expressing either marker could be detected (Figure 5H, 5I). Thus, similar to chick, prolonged exposure to Shh is required to maintain normal ventral pattern in the neural tube. Moreover, these data identify a crucial role for floor plate derived Shh in the patterning of the neural tube. Together, these results indicate that, even after progenitor identity has been assigned, continued Shh signalling is necessary to maintain progenitor identity in vivo in the ventral neural tube.
The continuous requirement for Shh signalling in vivo, coupled with the plasticity of progenitor identity in explants, raised the possibility that during normal embryonic development some progenitor cells may transiently express a transcriptional code characteristic of a more ventral progenitor population before reverting to a more dorsal identity. To test this, we used Olig2Cre mice to mark the progeny of cells that have expressed Olig2 [24]. Examining E10.5 mice embryos harbouring the Olig2cre allele and the ROSA26-flox-STOP-flox-lacZ or the ROSA26-flox-STOP-flox-GFP lineage tracers [51] revealed cells expressing the lineage marker within the pMN domain and their MN progeny (Figure 5J, 5K). In addition, however, a small number of cells positive for the reporter were observed dorsal to the pMN domain. These were located adjacent to the dorsal limit of the pMN domain, defined by Olig2 expression, but did not express Olig2 protein (Figure 5J, j). This suggests that some cells residing in the p2 progenitor domain transiently expressed Olig2 and activated Cre expression. This was confirmed by examining e11.5 embryos. In these embryos a few Chx10 and GATA3 expressing V2 neurons, which are generated from Olig2 negative, p2 progenitors, contained the reporter, indicating that they had been generated from progenitors that had previously expressed Olig2 (Figure 5K, k and unpublished data). Together these data confirm the plasticity of progenitor identity in vivo and emphasize the dynamic way in which positional information is supplied in the neural tube.
In this study we provide evidence that the interpretation of Shh morphogen gradient in the neural tube is dynamic and entails a prolonged period of intracellular signalling to elaborate and maintain gene expression in progenitor cells. Shh induced intracellular Gli activity is required for the patterning of progenitors [22],[52]. We show that the Gli activity in neural cells reaches a peak ∼6 h after exposure to Shh. Below 1 nM Shh, the intensity of this peak of Gli activity correlates with the concentration of Shh; above 1 nM Shh intracellular signalling appears saturated (Figure 1B). For all Shh concentrations, however, cells progressively adapt their response to Shh, resulting in a gradual decline in Gli activity levels from 6 h onwards. Thus, the time it takes for Gli activity to return to basal levels (i.e. “the duration of signalling”) is proportional to Shh concentration for all concentrations. Hence, different concentrations of Shh generate distinct temporal profiles of Gli activity (Figure 1C). The features of this profile—the duration and the level of Gli activity—are used to allocate positional identity to neural progenitors (Figure 6). Progressively higher levels and longer durations of intracellular Gli activity specify more ventral identities (Figures 2, 3) and, even for the interneurons that are generated by below saturating concentrations of Shh, the duration of signalling plays a key role. These data indicate, therefore, that the temporal adaptation mechanism of morphogen interpretation allocates positional identity throughout the ventral neural tube. This offers a unifying view of ventral neural tube patterning that differs significantly from conventional models of morphogen action [3],[24]. Consistent with the dynamic aspect of this model, Shh target genes initially occupy territories that are ventral to their final positions and are progressively refined dorsally over time (Figure 3D, 3E). Moreover, even after a positional identity has been assigned to a cell, it remains susceptible to change. For the more ventral progenitor domains, Gli activity above basal levels is required to maintain an identity (Figures 4, 5). Thus, blockade of signalling results in ventral progenitors losing their identity and acquiring an antecedent, more dorsal identity. This emphasizes a second important difference between the mechanisms of Shh interpretation and standard morphogen models, which assert that positional identity of cells can ascend but not descend a gradient [3],[53]. Together, the data suggest a model in which the positional information provided by Shh can be represented by the time integral—the cumulative level and duration—of signalling and the identity of progenitors is assigned in a progressive and dynamic manner.
The use of signal duration in the assignment of positional identity contrasts with other models of morphogen action that contend that the level of intracellular signalling is the main, or sole, factor that determines the response [3]. For example, different levels of nuclear Smad activity, the transcriptional mediator of TGFβ signalling, have been proposed to be sufficient to mediate the graded response necessary for mesoderm patterning [54],[55],[56]. Similarly, the graded response of cells to Dpp in Drosophila appears to correlate with the amount of activated Mad transcription factor [57],[58],[59] and the anterior-posterior patterning of Drosophila embryos depends on the amount of Bcd protein in each nucleus [60],[61]. In the case of Shh signalling, the use of duration as well as level of signalling to provide positional information introduces an additional dimension to the process that could be relevant to other responding tissues. Notably, anterior-posterior patterning of the developing limb appears to depend on both a concentration and temporal gradient of Shh signalling [62],[63],[64] and, analogous to the neural tube, the specification of digit identity is progressive with the sequential induction of more posterior identities [65]. Likewise, Hh signalling in the wing disc of Drosophila embryos results in the induction of higher response genes later than lower response genes [66]. Thus, the mechanism of morphogen interpretation operating in the neural tube might be common to other tissues patterned by graded Hh signalling. Furthermore, since negative feedback is a general feature of most signal transduction pathways [67], the graded response of cells to other morphogens might result in similar temporal dynamics and exploit analogous mechanisms.
What determines the temporal dynamics of signalling in responding cells? The Shh dependent upregulation of Ptc1, and possibly other inhibitors of Shh signalling, results in the cell autonomous desensitization of cells to Shh over time [23],[24]. This negative feedback is likely to play a significant part in defining the signalling kinetics in responding cells. It would account for the temporal profile and in particular the rate of decline, and hence the duration of signalling, in cells exposed to a fixed concentration of Shh. However, in addition to negative feedback, the period of time a cell is exposed to Shh will also determine the duration of signalling. Progenitors in vivo are subject to a changing gradient of Shh and as a result the time that Shh is available might limit the duration of signalling [14]. This could be particularly relevant in intermediate regions of the neural tube, which generate V0–V2 neurons. It takes a longer time for the Shh gradient to extend to these positions [14] and then the growth of the neural tube and the sequestration of Shh ventrally might limit the period of time for which significant concentrations of Shh are maintained at this distance [23],[26],[47]. In agreement with this idea, significant numbers of V0 and V1 neurons are only generated in vitro when explants are exposed to Shh for less than 18 h (Figure 2D). Longer exposure results in cells switching identity towards more ventral fates. A similar model has been proposed for the temporal changes in Hh distribution in the anterior compartment of the Drosophila wing disc [68]. In this view, the duration of Shh exposure is decisive for specifying the intermediate identities, p0–p2, while the more ventral pMN and p3 identities are chiefly dependent on maintaining appropriate high concentrations of Shh, which are transformed into corresponding periods of intracellular signal transduction by negative feedback. However, the critical test of this model awaits the capability to assay directly Shh protein distribution and intracellular signalling in vivo in the developing neural tube.
The central role of Shh in providing positional information to neural progenitors does not exclude the possibility that other extrinsic signals contribute to ventral neural tube patterning. Members of the BMP and Wnt families as well as Notch and retinoic acid signalling have been implicated in aspects of neural tube development and cell fate allocation [6],[40]. It is possible that these interact, either directly or indirectly, with Shh signalling to refine and modify the positional identity provided by Shh signalling. This could increase the precision or reliability of patterning in vivo. In this context, it is notable that the in vitro generation of V0 and V1 neurons overlap in explants exposed to Shh (Figure 2D). Both retinoic acid and Notch signalling have been implicated in the generation of these neuronal subtypes and it is possible that one or both of these signals are involved in the spatially segregated generation of these interneurons in vivo [40],[69].
The dynamic nature of Shh mediated pattern formation is emphasized by the requirement for continued Shh signalling to maintain the more ventral progenitor identities even after the gene expression profiles that define these progenitor domains have been induced. Thus, progenitors of MNs and V3 neurons revert to more dorsal identities if signalling is interrupted. This provides an explanation for the requirement for Shh signalling up to the last cell division in order to specify MN subtype identity [39]. Importantly, this also identifies a role for floor plate produced Shh for DV patterning of neural progenitors despite the initial induction of ventral gene expression by Shh secreted from the notochord [14]. Moreover, reversibility of the expression of Hh target genes has also been observed in other tissues, such as the Drosophila wing disc [70], suggesting it is a common feature of the pathway. It remains to be determined whether the expression of Nkx2.2 and Olig2 eventually becomes independent of continued Shh signalling in the neural tube and if so at what time point. Nevertheless, the extended requirement for Shh signalling to maintain correct tissue pattern highlights a difference with other morphogens. In the case of mesoderm induction by Activin [53] and BMP signalling in the telencephalon [71] cells maintain the gene expression profile associated with the highest concentration of morphogen to which they have been exposed, even if the ligand exposure is limited to as little as 20 min [53]. This has been dubbed “the ratchet effect” because cells appear to respond to increases but not decreases in morphogen concentration [3],[53]. As a result, prolonged exposure to TGF-β morphogens is unnecessary for patterning [53]. One consequence of a self-sustaining memory for the highest morphogen response is that it obviates the need to preserve a stable gradient for a long period time. This might be important in tissues that develop quickly and/or undergo movements that preclude the establishment of a stable gradient. However, cells would be vulnerable to transient fluctuations in signalling and could adopt an identity inappropriate for their position in response to a brief exposure to an anomalously high level of morphogen. Thus, the requirement for an extended period of Shh signalling to maintain positional identity in the neural tube emphasizes the ongoing refinement of progenitor boundaries during development and suggests a way to buffer fluctuations in ligand concentration and enhance the precision of the response.
The delayed induction of some Shh target genes in progenitor cells suggests that the transcriptional state of cells is important for progenitor identity specification. This points to a model where the transcriptional network downstream of Shh signalling together with the temporal profile of Gli activity in a cell accounts for differential gene expression. Hence, Gli dependent induction of some genes might take place only after changes in the transcriptional state of a responding cell has been changed by an early period of Gli activity. This would explain the delayed induction of some progenitor markers and a requirement for ongoing signalling for their expression. A well-described example of this strategy is found in the Dpp response of cells of the dorsal ectoderm of Drosophila embryos [72]. In this case, Dpp induces expression of the transcription factor Zen. Then, Zen together with continued Dpp signalling activates a second gene, Race. Thus, the expression of Race takes longer than Zen and requires Dpp signalling to be sustained after Zen is induced. This strategy for morphogen dependent gene regulation has been referred to as “sequential cell context” [73] and represents an example of a feed-forward loop [74]. In the neural tube, it is well established that the transcriptional cross-repression between target genes downstream of Shh plays a key role in defining the spatial extent of progenitor domains [21]. It seems likely that these same regulatory interactions will also be involved in the dynamics of gene expression. For example, the repression of Nkx2.2 by Pax6 determines the dorsal limit of Nkx2.2 expression [25]; moreover, Nkx2.2 expression appears to shift increasingly dorsal in mouse embryos lacking Pax6 (N.B. et al., unpublished observation). Thus both the temporal and spatial pattern of Nkx2.2 appears to be influenced by Pax6. The reciprocity of many of the cross-regulatory interactions within the transcriptional network might also account for the prolonged requirement for Shh signalling to sustain progenitor identity. In this view, continued Gli activity would be required to ensure that the extant transcriptional state of a responding cell does not revert to a previous state. This strategy of morphogen interpretation suggests a dynamic mechanism in which the positional identity of a cell is determined by the combined action of the ligand gradient and the state of the transcriptional network in the responding cell itself.
More generally, the mechanism of Shh morphogen interpretation is reminiscent of “integral feedback control,” a common control strategy in systems engineering in which the time integral of the difference between the actual output and a target output is fed back into the system to correct the output [75]. In the case of Shh signalling, the Gli dependent induction and gradual accumulation of Ptc1 could be viewed as the “time integral” that corrects the sensitivity of cells to Shh; hence increasing amounts of Ptc1 result in the increasing desensitization of cells to Shh signalling (Figure 6A, 6B). The use of integral feedback control has been noted in several biological systems where it can act as a “gain” control to allow sensing over large concentration ranges or as a means to re-establish homeostasis after a system is disturbed [75],[76],[77]. An additional characteristic of this mechanism is that it provides a way to transform different levels of input into corresponding durations of signal output and it is this feature that appears to be exploited by cells to interpret graded Shh signalling. This would reconcile the competing models of pattern formation that have emphasized either concentration or time-dependent mechanisms [73]. Furthermore, in contrast to concentration-based mechanisms of morphogen interpretation where concentrations above the saturation limit elicit the same response, a temporal adaptation mechanism allows cells to discriminate between saturating concentrations of ligand (Figure 6B). This would therefore extend the range of concentrations that can be discerned by a cell and shift the functional range to higher concentrations. The effect would be to increase the potential set of responses that can be elicited by a single signal and render the process less susceptible to the higher levels of noise associated with low concentrations of an extrinsic ligand [78]. A trade off with this mechanism, however, is that it requires patterning to occur over a comparatively long time period in which cell position remains stable relative to the gradient. Nevertheless, despite this limitation, it is possible that the interpretation of other morphogens may use similar mechanisms. For example, a correspondence between the activation of target genes and duration of signalling has been noted for Nodal in zebrafish embryos [79] and Dpp in the Drosophila wing disc [80]. In addition, brief exposure to Wg signalling is sufficient to induce low but not high responses in the wing disc [81]. Thus, the elucidation of the mechanism by which Shh provides positional information to neural cells suggests strategies of morphogen interpretation that may be generally applicable in developing tissues.
In ovo electroporation, fixation, sectioning of embryos, and immunocytochemistry were performed as described ([21]). Antibodies used were rabbit anti-Arx (a gift from J. Chelly), goat anti-β-galactosidase (Biogenesis), rabbit anti-Chx10 (a gift from T. Jessell), rabbit anti-Cre (Covance), Rabbit anti-Dbx1 [34], mouse anti-En1 (DSHB), mouse anti-Evx1 (DSHB), sheep anti-GFP (Biogenesis), mouse anti-MNR2 (DSHB), rabbit anti-Nkx2.2 (a gift from T. Jessell), mouse anti-Nkx2.2 (DSHB), mouse anti-Nkx6.1 (DSHB), guinea-pig anti-Olig2 [82], mouse anti-Pax6 (DSHB), mouse anti-Pax7 (DSHB), and mouse anti-Shh (DSHB).
The open neural plate of HH stage 10 chick embryos was isolated in L-15 (Gibco) media and the intermediate region dissected following Dispase treatment (Gibco). These [i] explants were embedded in Collagen Type I containing DMEM (Gibco). Culture medium contained F-12/Ham (Gibco) supplemented with 2 mM of Glutamine, 50 U/ml of Penicillin, 50 µg/ml of Streptomycin and Mito Serum (BD). Recombinant Shh protein was produced as described [39]. Explants were fixed in 0.1 M phosphate buffer pH7.2 containing 4% PFA prior to immunostaining. Two regions containing approximately 200 cells were chosen from each of 4 or 5 explants for quantitations. For luciferase assays in explants, GBS-luc [83] was electroporated with normalization plasmid (pRL-TK; Promega) 2 h prior to the dissection of the explants [24]. Gli activity was measured using the Dual Luciferase Assay (Promega) and compared to levels of Gli activity in transfected untreated explants.
Eggs were incubated to HH st.18 and windowed. Embryos were treated with 25 µl of 1 mg/mL solution of cyclopamine (Toronto Research Chemicals) in 45% 2-hydropropyl-β-cyclodextrin (HBC; Sigma) [44]. The embryos were re-incubated for a further 24 h and then processed for immunocytochemistry.
The Dbx1-Cre, Olig2-Cre, ShhFlox, and Bcre32 (Tg(Pou3f4-cre)32Cre) mouse lines have been described previously [24],[42],[48],[49] and the conditional lineage tracing alleles engineered in the ROSA26 locus have been described [43],[51]. All studies in mice were carried out with appropriate permissions and in accordance with the Institutional Animal Care and Use Subcommittee of the National Institute for Medical Research.
|
10.1371/journal.pntd.0004806 | Clonorchis sinensis Co-infection Could Affect the Disease State and Treatment Response of HBV Patients | Clonorchis sinensis (C. sinensis) is considered to be an important parasitic zoonosis because it infects approximately 35 million people, while approximately 15 million were distributed in China. Hepatitis B virus (HBV) infection is a major public health issue. Two types of pathogens have the potential to cause human liver disease and eventually hepatocellular carcinoma. Concurrent infection with HBV and C. sinensis is often observed in some areas where C. sinensis is endemic. However, whether C. sinensis could impact HBV infection or vice versa remains unknown.
Co-infection with C. sinensis and HBV develops predominantly in males. Co-infected C. sinensis and HBV patients presented weaker liver function and higher HBV DNA titers. Combination treatment with antiviral and anti-C. sinensis drugs in co-infected patients could contribute to a reduction in viral load and help with liver function recovery. Excretory-secretory products (ESPs) may, in some ways, increase HBV viral replication in vitro. A mixture of ESP and HBV positive sera could induce peripheral blood mononuclear cells (PBMCs) to produce higher level of Th2 cytokines including IL-4, IL-6 and IL-10 compared to HBV alone, it seems that due to presence of ESP, the cytokine production shift towards Th2. C. sinensis/HBV co-infected patients showed higher serum IL-6 and IL-10 levels and lower serum IFN-γ levels.
Patients with concomitant C. sinensis and HBV infection presented weaker liver function and higher HBV DNA copies. In co-infected patients, the efficacy of anti-viral treatment was better in patients who were prescribed with entecavir and praziquantel than entecavir alone. One possible reason for the weaker response to antiviral therapies in co-infected patients was the shift in cytokine production from Th1 to Th2 that may inhibit viral clearance. C. sinensis/HBV co-infection could exacerbate the imbalance of Th1/Th2 cytokine.
| Clonorchiasis and hepatitis B infection are infectious diseases that affect millions of people worldwide, especially in China. These two diseases are caused by two different pathogens, C. sinensis and hepatitis B virus, respectively. Concurrent infection between HBV and C. sinensis is often observed in some areas where C. sinensis is endemic. Both diseases share the same target organ, but there is little known on whether concomitant clonorchiasis could have an impact on HBV infection and the efficacy of antiviral treatment. In this study, we showed for the first time that co-infection with C. sinensis and HBV resulted in significantly higher liver transaminases levels as well as HBV DNA copies, indicating that co-infection with C. sinensis and HBV infection may aggravate the disease state. Combination treatment with antiviral and anti-C. sinensis drugs in co-infected patients could contribute to a reduction in viral load and help with liver function recovery. Furthermore, excretory-secretory products (ESPs) of C. sinensis may have a potential role in promoting HBV viral replication. This may explain, at least in part, the higher HBV DNA copies observed in co-infected patients. Additionally, a mixture of ESP and HBV positive sera could induce PBMCs to mainly produce Th2 cytokines such as IL-4, IL-6 and IL-10 compared to HBV alone. A possible reason for higher HBV DNA copies and a weaker response to antiviral therapies in co-infected patients was the shift in cytokine production from Th1 to Th2 that may inhibit viral clearance.
| Clonorchiasis, caused by Clonorchis sinensis (C. sinensis), is one of the parasitic zoonosis. It is estimated that approximately 35 million people are infected in Asia, among which approximately 15 million infected people were in China [1–3]. Previous epidemiological data showed that clonorchiasis is endemic in the southeast of China, especially in Guangdong province [4]. The people become infected with C. sinensis by the consumption of raw or undercooked fish that contains C. sinensis metacercariae [5]. The adult worms of C. sinensis, located in the small bile ducts of the liver, lead to mechanical damage, while excretory-secretory products (ESPs) of C. sinensis cause chemical damage. Both mechanical damage and chemical damage play a key role in causing hyperplasia and adenomatous changes in the bile ducts[6]. The ESPs are known to be involved in parasite-host interaction and have clinical significance in the diagnosis and pathogenesis [7–10].
Hepatitis B virus (HBV) infection is a major public health issue that may develop into cirrhosis, hepatic decompensation and hepatocellular carcinoma (HCC)[11]. An estimated 2 billion people have been infected, and more than 350 million are chronic carriers of the virus despite the availability of a prophylactic vaccine[12]. HBV infection situation is even more severe in China where approximately 170 million people are chronically infected with HBV [13,14]. Th1 responses seem to be involved in the clearance of HBV, while chronic HBV infection elicits very weak T cell responses [15,16].
A relatively smaller population is infected with C. sinensis compared with HBV[17]; concurrent infection with HBV and C. sinensis is often observed in areas where C. sinensis is endemic[18]. In some epidemiological studies, the positive rate of HBV surface-antigen (HBsAg) was significantly higher in areas endemic with C. sinensis than non-endemic areas [19–21]. However, C. sinensis infection and chronic HBV infection are two different causes of liver disease in China, and limited data are currently available as to whether there is any association between HBV and C. sinensis infections. There has been no clear discussion as to whether the C. sinensis infection may impact upon HBV infection or vice versa. The aim of our study is to evaluate the impact of C. sinensis infection on HBV infection as well as the response to antiviral therapy in co-infection with C. sinensis and HBV. Our results showed that co-infected individuals presented weaker liver function and higher HBV DNA titers. In co-infected patients, the efficacy of anti-viral treatment was better in patients who were prescribed ETV and PZQ than ETV alone. A possible reason for higher HBV DNA copies and a weaker response to antiviral therapies in co-infected patients was the shift in cytokine production from Th1 to Th2 that may inhibit viral clearance.
The Institutional Review Boards of the Third Affiliated Hospital and Zhongshan School of Medicine, Sun Yat-sen University, approved this study as an exempt study for which informed consent did not need to be sought from subjects. Informed consent was not sought for this study as all information was obtained from the existing medical record, and data were analyzed anonymously.
HBV positive sera were obtained from chronic HBV patients, and peripheral blood mononuclear cells (PBMCs) were collected from healthy donors or chronic HBV patients from the Third Affiliated Hospital of Sun Yat-sen University. Samples were anonymously coded in accordance with local ethical guidelines (as stipulated by the Declaration of Helsinki), and written informed consent was obtained from patients and healthy volunteers. The work was conducted in strict accordance with the study design as approved by the Clinical Research Ethics Committee of the Third Affiliated Hospital and Zhongshan School of Medicine, Sun Yat-sen University, Guangzhou, China.
First, we evaluated HBV patients by screening all consecutive patients at the Third Affiliated Hospital of Sun Yat-sen University between July 2014 and February 2015. The inclusion criteria for patients who were mono-infected with HBV were the following: men and women aged 18 years and older; HBV surface-antigen (HBsAg)-positive; and HBV DNA >20 IU/mL. The inclusion criteria for patients who were co-infected were the following: men and women aged 18 years and older; HBV surface-antigen (HBsAg)-positive; HBV DNA >20 IU/mL; and having C. sinensis eggs in the stools. The inclusion criterion for patients who were mono-infected with C. sinensis was having C. sinensis eggs in the stools. The inclusion criteria for healthy subjects were negative for both HBV and C. sinensis. Furthermore, we compared co-infected patients who were treated with entecavir (ETV, 0.5 mg once daily) alone with those who received a combination treatment of ETV (0.5 mg once daily) and praziquantel (PZQ, 210 mg/kg, 3 times a day, for 3 days) to determine the impact of C. sinensis on antiviral therapies. The following data were collected from electronic medical records by computer-assisted chart review: age, gender, date of prescription of antiviral drug ETV and PZQ drug, HBV DNA copy numbers, serial liver function test once in 2 weeks, including aspartate aminotransferase (AST), alanine aminotransferase (ALT), and total bilirubin (TB). Patients with the following concomitant conditions were excluded: those co-infected with HIV, hepatitis A, C, D and E, those with type I and type II diabetes, those co-infected with Schistosoma japonicum, or Schistosoma mansoni or other parasites, and those with alcoholic liver, autoimmune diseases, cholestasis, serious heart diseases and pregnant women. Due to the retrospective nature of the study, written informed consent could not be obtained from all patients. All data were de-identified prior to analysis.
Biochemical tests were performed using routine automated analyzers. HBsAg was detected by electrochemiluminescence immunoassay with COI >1.00 (COBAS). Serum ALT, AST, and TB levels were determined using commercial kits (Maccura, China). Serum levels of HBV DNA were measured by real-time PCR with a lower detection limit of 20 IU/mL (COBAS).
ESPs were obtained as previously described [22]. Briefly, the living adult C. sinensis parasites were cultured in DMEM, and the supernatant was collected at 48 h.
Fresh whole blood (5 ml) was obtained from health adult volunteers and chronic HBV patients. The blood was mixed with the same volume of phosphate-buffered saline (PBS) and layered on 5 ml of lymphocyte separation medium (TBD, China). The sample was centrifuged at 2000 rpm for 20 min at RT. PBMCs from the upper portion of the Ficoll layer were collected, washed with PBS and centrifuged at 1500 rpm for 10 min at RT. The PBMCs were suspended in RPMI 1640 (Gibco) with 10% fetal bovine serum (Gibco) at a concentration of 2 × 106 cells/ml in endotoxin-free tubes.
PBMCs, seeded into 12-well cell culture clusters at a density of 1.0 × 106 viable cells per 200 μl of culture medium, were incubated with one of the following conditions: 1) HBV DNA-positive sera (2.0 ×106 HBV DNA IU/mL), 2) HBV DNA-positive sera (2×106 HBV DNA IU/mL) and ESPs (20 μg/mL), or 3) ESPs (20 μg/mL) alone for 48 h at 37°C in a humidified 5% CO2 incubator. Finally, cells from each culture and their corresponding supernatants were analyzed cytokine mRNA expression and HBV DNA copies, respectively.
Total RNA of PBMCs was extracted using Trizol reagent (Life Technologies, USA) according to the manufacturer’s protocol. cDNAs were synthesized using a cDNA Synthesis Kit (TransGen, China). Quantitative real-time PCR was performed using a Bio-Rad CFX96 Real-Time System (Bio-Rad, USA) to measure SYBR Green (TRANSGEN BIOTECH, China) incorporation into double stranded amplicons. Reactions were performed in 20 μl volumes containing forward and reverse primers at a final concentration of 100 nM. Primer sequences are listed in Table 1. The PCR reaction conditions included a denaturation step at 94°C for 30 sec, then 40 cycles of a three-step cycling reaction as follows: 94°C for 5 sec, then 55°C for 15 sec and 72°C for 10 sec. Melting curve analysis revealed a single peak for each primer set. IL-2, IL-4, IL-6, IL-10 and IFN-γ were measured and normalized relative to β-actin expression. The changes in mRNA expression were analyzed by calculating 2-ΔΔCt. The accession numbers for genes mentioned in the text are listed in S1 Table.
Serum samples were obtained by centrifugation at 3000 rpm for 5 min. Serum samples were immediately stored at -80°C and thawed prior to analysis. Cell culture supernatants and serum concentrations of IL-2, IL-4, IL-6, IL-10 and IFN-γ were analyzed by ELISA with commercially available kits (Elabscience, China) according to the manufacturer’s instructions. The concentration of each cytokine was determined using a standard curve according to the kit instructions.
All data were presented as the mean values±standard error or mean values. Data analyses were carried out using the GraphPad Prism software 5.0. For comparison with more than two groups, one-way ANOVA test was conducted, and if the data were nonparametric, a Kruskal-Wallis test with a confidence interval of 95% was employed. p<0.05 was considered statistically significant.
During the study period, there were 701 patients who met the selection criteria, of whom 51 were C. sinensis/HBV co-infected, 53 were C. sinensis mono-infected and 520 were HBV mono-infected. In addition, 77 healthy individuals with a mean age similar to the patient population were included. The patients' characteristics are reported in Table 2. Patients in the infected groups were predominantly men (94% for C. sinensis/HBV co-infected, 73% for mono-HBV, 81% for C. sinensis infected). Liver function was assessed by measurement of TB, ALT and AST in plasma. All 3 infected groups of patients showed higher levels of ALT, AST and TB than the healthy control subjects. ALT, AST and TB levels were significantly higher in the co-infected group (p<0.0001, p<0.0001, p<0.0001, respectively) compared to HBV mono-infected patients. Furthermore, HBV DNA log copies were also significantly higher in the co-infected patient group (p <0.05, Table 2). Taken together, these data indicate that patients co-infected with C. sinensis and HBV had weaker liver function than mono-HBV infected and that the presence of C. sinensis may aggravate the disease state.
Given that the presence of C. sinensis may aggravate HBV infection disease state, we further investigated whether inhibition of C. sinensis could influence the efficacy of antiviral treatment in co-infected patients clinically. There were 51 co-infected patients, 21 of whom were prescribed ETV and PZQ drugs and 30 of whom were prescribed antiviral ETV drugs only. Because 9 out of 30 patients had not been checked for liver function after treatment, they were excluded from this part of the study. There were no significant differences in HBV DNA copies and the levels of ALT, AST and TB between the two groups before treatment (Fig 1). After one program of treatment, the level of ALT, and AST and HBV DNA copies were significantly decreased compared to the pretreatment values in both groups, but no significant differences were observed in the levels of ALT and AST between the two groups (Fig 1A and 1B). There was no obvious change in TB level between pre- and post-treatment in C. sinensis/HBV-NONPZQ groups. However, C. sinensis /HBV-PZQ patients demonstrated significantly lower levels of TB than C. sinensis /HBV -NONPZQ patients after one program of treatment (p<0.01, Fig 1C). Additionally, C. sinensis /HBV-PZQ patients had lower levels of HBV DNA log copies compared to C. sinensis /HBV–NONPZQ (p <0.05, Fig 1D). Patients who took PZQ showed C. sinensis eggs negative by Kato-Katz thick stool smear technique (S1 Fig). Together, these results suggested that combined antiviral and anti-clonorchiasis drugs in co-infected patients could contribute to a reduction in viral load and help with liver function recovery.
Co-infected patients not only showed weaker liver function but also had significantly higher HBV DNA copies clinically. We reasoned that some metabolites of C. sinensis may directly enhance HBV replication. To address this possibility, we tested HBV DNA copies in the supernatants after co-cultured of PBMCs with ESP and HBV positive patient sera. HBV positive patient serum alone and ESPs alone served as controls. HBV DNA was measured by real-time PCR with a lower detection limit of 20 IU/mL. As expected (Fig 2), HBV DNA copies were significantly higher in the culture supernatants from the PBMCs co-cultured with ESP and HBV mixture than the control groups (p <0.01). These data suggested that ESPs may, in some ways, promote viral replication.
To define and compare the secretion of Th1 cytokines (IL-2 and IFN-γ) and Th2 cytokines (IL-4, IL-6 and IL-10) following in vitro stimulation, we used quantitative RT-PCR to analyze the mRNA levels of each cytokine secreted by PBMCs, which were stimulated with ESPs alone, HBV positive patient serum alone or the mixture of the two, respectively. Data have been normalized for β-actin transcript expression. As showed in Fig 3, the levels of different cytokine mRNAs varied. IL-4 and IL-10 cytokine mRNA levels were higher in ESP-stimulated PBMCs than in HBV-stimulated PBMCs (Fig 3B), whereas there is no significant difference in IL-2 and IFN-γ cytokine mRNA levels between these two groups (Fig 3A). In particular, IL-6 cytokine mRNA levels were significantly higher in HBV positive sera stimulated PBMCs than in ESP-stimulated PBMCs (Fig 3B). This finding suggested that both stimulators could induce PBMCs to produce Th1 and Th2 cytokines in vitro. Furthermore, in response to the treatment with a mixture of HBV and ESPs, IL-4 and IL-10 mRNA level increased two-fold, and IL-6 mRNA level increased three-fold compared with PBMCs stimulated with HBV positive sera alone, while there were no significant changes in IL-2 and IFN-γ levels. These data suggested that PBMCs stimulated by a mixture of ESP and HBV produced higher level of Th2 cytokines including IL-4, IL-6 and IL-10 compared to HBV alone, it seems that due to presence of ESP, the cytokine production shift towards Th2.
To validate the real-time PCR results, the protein levels of those cytokines were examined in cell culture supernatant both from stimulated PBMCs from healthy donors and chronic HBV patients (S1 Fig). ESP/HBV stimulated PBMCs secreted a higher level of IL-4, IL-6 and IL-10 than in the HBV group, and IFN-γ was lower in ESP/HBV than in HBV (S2 Fig). However, there was no significant difference between ESP/HBV and HBV groups with regard to IL-2 level.
A similar pattern of cytokine expression could be observed in the stimulated PBMCs from chronic HBV patients (S3 Fig). Note that there is no significant difference in IL-4 and IL-6 in stimulated PBMCs with those from chronic HBV patients. These data confirm that PBMCs stimulated by a mixture of ESPs and HBV mainly produced Th2 cytokines.
To verify whether there are changes in serum cytokine levels in C. sinensis/HBV co-infected patients, HBV mono-infected patients and C. sinensis mono-infected patients, we performed cytokine ELISA to examine the levels of IL-2, IL-4, IL-6, IL-10 and IFN-γ. C. sinensis/HBV co-infected patients had both lower IFN-γ and IL-2 levels than both HBV mono-infected and C. sinensis mono-infected patients (S2 Table). All 3 groups of infected patients had higher IL-4, IL-6 and IL-10 levels than healthy control subjects. IL-6 levels were further increased in C. sinensis/HBV co-infected patients, compared with those in HBV mono-infected patients (p <0.05). C. sinensis/HBV co-infected patients had higher IL-10 levels than HBV mono-infected patients (p <0.05).
In this study, we investigated the relationship between C. sinensis infection and HBV infection in humans and further studied the impact of C. sinensis infection on the efficacy of antiviral treatment.
In our study, we provide strong evidence for the existence of an association between HBV infection and C. sinensis infection. Our data showed that a correlation of co-infection C. sinensis and HBV developed predominantly in males. This finding is supported by previous reports that C. sinensis infections in male individuals are usually higher than that in female individuals [23–25]. Additionally, our results showed that co-infected patients had significantly higher liver transaminases levels as well as HBV DNA copies, indicating that concomitant C. sinensis infection aggravated the liver disease. PZQ is known to be very effective and the drug of choice against trematode and cestode infections. Mesan et al. demonstrated that oral PZQ to patients co-infected with schistosomiasis and hepatitis C virus (HCV) could help the response to HCV treatment [26]. Patients who received C. sinensis infected liver transplantation who took PZQ experienced improved liver function[27]. The present study provides ample evidence of significant decreases in the levels of TB and HBV DNA copies by using the combination of PZQ during antiviral therapies in co-infected patients. It is theorized that the beneficial effects are likely related to the clearance of C. sinensis worms and subsequent reduction of metabolites of C. sinensis. This result further indicated that the efficacy of HBV antiviral treatment was related to the removal worms in co-infected patients.
On the other hand, previous studies suggested that Schistosoma mansoni soluble egg antigens (SEA) have the potential to enhance HCV propagation [28] and SEA of Schistosoma Haematobium induces HCV replication in PBMCs [29], which indicates that some components of trematode could enhance viral replication. ESPs of C. sinensis could cause chemical damage to the host [30,31] and induce cell proliferation in vitro [32]. Other findings suggested that in vitro infection of PBMCs with human sera contain HBV particles may be a suitable model to study the early steps of the viral life cycle [33,34]. M. Cabrerizo et al showed that the viral DNA can be detected after incubation of PBMCs with human sera containing HBV particles and that HBV is able to infect, replicate and release viral particles in the medium in in vitro infected PBMCs[35]. Therefore, we have used cultures of PBMCs from a healthy donor to test the impact of ESPs on HBV particle replication. We have demonstrated that HBV DNA can be detected in the supernatant 48 h after incubation of PBMCs with human sera containing HBV particles. Additionally, the results revealed that significantly higher level of HBV DNA in the supernatant from cells co-cultured with HBV positive sera and ESPs. This may explain, at least in part, the higher HBV DNA copies observed in co-infected patients. Unfortunately, which components of ESPs are involved were not identified in this study. Additional studies are required to determine which components of ESPs are the key role players.
Studies on bile and serum of patients indicated that infection with C. sinensis correlated with Th2 type responses, emphasizing the decrease in concentration in IL-2 and an increase in IL-4. It was suggested that ESPs are immunogenic, stimulate inflammation and promote proliferation, and suppress apoptosis [32]. Most proteins belonging to ESPs, including CsRNASET2, CsLAP2, and CsNOSIP, contributed in eliciting Th2 immune response in mice [36–38]. Cytokines are important mediators in the regulation of the immune response. However, it was not known yet whether ESPs of C. sinensis could stimulate PBMCs to produce cytokines in vitro as well as whether they have an influence on cytokine expression produced by HBV positive sera stimulated PBMCs. Thus, we examined the levels of cytokine specific mRNAs to clarify the cytokine response. We observed that PBMCs stimulated by either ESPs or HBV positive sera could prompt the secretion of Th1 cytokines by enhancing the expression of IL-2 and IFN-γ and Th2 cytokines by enhancing the expression of IL-4, IL-6 and IL-10. Additionally, in response to mixtures of HBV positive sera and ESPs, the levels of Th2 cytokines were notably higher compared to HBV positive sera alone, whereas there was no significant change in Th1 cytokine expression level, indicating that HBV/ESP predominantly produced Th2 cytokines. IL-6 exhibits both pro- and anti-inflammatory functions in innate immunity [39] and several studies have shown that IL-6 serum levels are increased in HBV positive patients, significantly higher in patients with severe and acute infections [40,41]. Wang et al. suggested that IL-6 is involved in the activation of natural killer cells and cytotoxic T lymphocytes induce the killing of hepatocytes, indicating that IL-6 plays an important role in liver cell necrosis and apoptosis[42]. Combined with our observation, these results suggested that a significantly higher level of IL-6 in response to a mixture of HBV and ESP stimulation, which may explain why the liver function is damaged severely in co-infected patients. In addition, compared to HBV infection alone, the levels of IL-4 and IL-10 were significantly increased, but IFN- γ was not changed in mixed HBV and ESP stimulation, indicating that cytokine production may shift from Th1 response to Th2 response. However, Th1 (including IL-2 and IFN- γ) cytokines have been identified to participate in the viral clearance while Th2 cytokine IL-10 serves as a potent inhibitor of Th1 effectors cells[43]. Therefore, one possible reason for the weaker response to antiviral therapies in co-infected C. sinensis/HBV patients was that the shift in cytokine production from Th1 to Th2 that may inhibit viral clearance. Cytokines participate in the induction and effector phases of the immune and inflammatory responses based on the protein level rather than the mRNA level; thus, we determined the protein levels of the above cytokines. Our results suggested that the cytokine pattern were similar between mRNA and protein levels. However, when we evaluated the response of PBMCs from chronic HBV patients to the same stimulators, the pattern of cytokine expression was similar to health subjects’ PBMCs, except for IL-4 and IL-6 level. We reasoned that PBMCs from chronic HBV patients may consist of antigen specific T and B cells, which may interfere with cytokine production.
IL-6 plays an important role in host defense against pathogens and mediates anti-parasite protective responses[44]. Additionally, IL-6 may participate in pathological complications of HBV [45]. Our results are consistent with these findings; we found that serum IL-6 were significantly higher in C. sinensis mono-infected patients than both HBV mono-infected and health control subjects; that IL-6 levels were higher in C. sinensis/HBV co-infected patients than in the mono-infected group as well. IL-10 is mainly involved in the regulation of inflammatory response. IL-10 can antagonize Th1 cell responses by inhibiting Th1 cell differentiation and IFN-γ production [46]. Increased levels of IL-10 are relevant to the degree of liver inflammation and lead to disease progression [47,48]. Serum IL-10 levels have been reported in patients with chronic hepatitis and cirrhosis [49] and IL-10 can restrain the host’s anti-HBV activity. In this study, the level of IFN-γ in the co-infected patient group was significantly lower than in mono-HBV patients and mono-C. sinensis patients. IFN-γ not only is a cytokine produced by Th1 and NK cells but has a critical role in the suppression of HBV replication [50,51]. Studies have shown that chronic HBV patients have a lower level of IFN-γ; so, patients may fail to develop an efficient anti-viral immune response [51,52]. In this study, the serum level of IFN-γ in the co-infected patient group was significantly lower than in the mono-HBV patients and mono-C. sinensis patients. Therefore, we assume that co-infection with C. sinensis in HBV infection may suppress the immune response by stimulating IL-10 production as well as inhibiting IFN-γ secretion as a result. In addition, our results showed that C. sinensis seems to induce Th2-related cytokines, with an increase in serum levels of IL4, IL-6 and IL-10. Given that serum cytokine levels were fluctuating in C. sinensis/HBV patients, this could partly suggest that C. sinensis infection could exacerbate the imbalances of Th1/Th2 cytokine in HBV patients; however, further analysis is warranted.
In conclusion, this study is the first to provide strong evidence for the association between C. sinensis infection and HBV infection. Co-infected individuals presented weaker liver function and higher HBV DNA titers. In co-infected patients, the efficacy of anti-viral treatment was better in patients who were prescribed ETV and PZQ than ETV alone. C. sinensis/HBV co-infection could exacerbate the imbalance of Th1/Th2 cytokine, which may lead to the chronicity of HBV infection, and C. sinensis may play a role in the unresponsiveness to antiviral therapy in co-infected patients. Further investigations are required to address this point.
|
10.1371/journal.pgen.1005435 | Comparative Study of Regulatory Circuits in Two Sea Urchin Species Reveals Tight Control of Timing and High Conservation of Expression Dynamics | Accurate temporal control of gene expression is essential for normal development and must be robust to natural genetic and environmental variation. Studying gene expression variation within and between related species can delineate the level of expression variability that development can tolerate. Here we exploit the comprehensive model of sea urchin gene regulatory networks and generate high-density expression profiles of key regulatory genes of the Mediterranean sea urchin, Paracentrotus lividus (Pl). The high resolution of our studies reveals highly reproducible gene initiation times that have lower variation than those of maximal mRNA levels between different individuals of the same species. This observation supports a threshold behavior of gene activation that is less sensitive to input concentrations. We then compare Mediterranean sea urchin gene expression profiles to those of its Pacific Ocean relative, Strongylocentrotus purpuratus (Sp). These species shared a common ancestor about 40 million years ago and show highly similar embryonic morphologies. Our comparative analyses of five regulatory circuits operating in different embryonic territories reveal a high conservation of the temporal order of gene activation but also some cases of divergence. A linear ratio of 1.3-fold between gene initiation times in Pl and Sp is partially explained by scaling of the developmental rates with temperature. Scaling the developmental rates according to the estimated Sp-Pl ratio and normalizing the expression levels reveals a striking conservation of relative dynamics of gene expression between the species. Overall, our findings demonstrate the ability of biological developmental systems to tightly control the timing of gene activation and relative dynamics and overcome expression noise induced by genetic variation and growth conditions.
| Embryonic development necessitates a delicate balancing act. On one hand, precise regulation of the expression of developmental genes is crucial for the maintenance of morphology and function. On the other hand, these same regulatory networks must allow normal development to proceed through genetic variation and environmental changes. To learn how regulatory circuits operate robustly within natural variation, we study the temporal expression profiles of key regulatory genes in the Mediterranean sea urchin, Paracentrotus lividus, and compare them to those of its Pacific Ocean relative, Strongylocentrotus purpuratus. These species shared a common ancestor about 40 million years ago and show highly similar embryonic morphologies. Our studies reveal highly reproducible gene initiation times that show lower variations than the variations in maximal mRNA levels within the species (Pl). We observe high interspecies conservation of the temporal order of gene activation within regulatory circuits and some cases of divergence. This conservation was even more profound when expression levels were normalized and scaled to the different developmental rates between the species. Our findings highlight that, despite genetic variations and different growth conditions, expression dynamics in developmental gene regulatory networks are extremely conserved over 40 million years of evolution.
| Normal development requires precise temporal control of differential gene expression, yet development must be robust to natural genetic variation and environmental changes. [1–3]. This resilience of developmental systems is important for keeping a wide genotypic pool adaptable in changing environmental conditions and thus, for the survival of the species [4,5]. Identifying how the control systems overcome genetic and environmental changes is important to the mechanistic understanding of developmental processes and their evolution [1,3,4]. Specifically, comparing different aspects of expression dynamics between individuals within the species and between closely related species can illuminate the range of variation in temporal expression that can still produce similar embryonic structures [1,6–8].
Comparative studies of interspecies differences in the kinetics of gene regulatory circuits can provide predictions for trans and cis evolutionary changes in circuit connectivity. The timing of gene expression depends on the temporal expression profiles of the inputs (trans) and the logic applied on the inputs by the cis-regulatory modules [9,10] (S1A–S1C Fig). For example, if two inputs are activated sequentially and the target cis-regulatory element requires both of them (necessary inputs, AND logic), the target gene will turn on only after the activation of the later input gene (S1B Fig) [9]. If the two inputs are additive (OR logic), the target gene will turn on immediately after the activation of the earlier input gene [11] (S1C Fig). Thus, evolutionary changes in cis-regulatory logic, e.g. from AND to OR, could result in changes in gene expression timing. Comparing the expression profiles of both input and target genes between two species can provide predictions for changes in input dynamics and in the target's cis-regulatory logic.
Comparative studies of temporal variation of gene regulatory circuits between related species must rely on detailed experimentally-based models of the gene regulatory networks in these organisms. The current models of the gene regulatory networks that drive ectoderm, endoderm and mesoderm specification in the sea urchin embryo are among the most comprehensive of their kind and are based on experimental studies in a few main species. [12–16]. The purple sea urchin, Strongylocentrotus purpuratus (Sp) inhabits the Pacific coasts of North America while the sea urchin Paracentrotus lividus (Pl) inhabits the eastern Atlantic Ocean and the Mediterranean Sea. These species shared a common ancestor about 40 million years ago and the average similarity in their coding sequences is about 85%, which is similar to that found between human and mouse. The growth temperature of these two species is different, reflecting their different environments; Pl embryos will successfully develop over a temperature range that is higher than Sp (standard lab temperatures 18°C versus 15°C, respectively). These species show apparent similarities in size, morphology, spatial gene expression patterns and gene regulatory networks, despite their genomic divergence and geographic distance (Fig 1A) [14–25]. High resolution studies of the temporal expression profiles of more than a hundred regulatory and differentiation genes that operate at different embryonic territories were performed for Sp [13,26], but equivalent information for Pl is still limited [18].
Here, we perform high-resolution quantitative analysis of the transcriptional expression profiles of key regulatory genes in Pl, asses the temporal expression variation within the species and compare gene expression dynamics to those measured in Sp [26]. For these studies, we selected regulatory circuits that operate in five embryonic territories and contain common network motifs found in many other gene regulatory networks, such as positive feedback and feedforward structures. The positive feedback circuitry locks down a specification state within a cell (intracellular, S1D Fig) or within an embryonic territory (intercellular, S1E Fig) and is important for cell fate decision [15,27–29]. Coherent and incoherent feedforward motifs are used for the sequential activation of genes in a cell (S1F Fig) [30–32]. Our results portray a tight control of timing of gene activation that is highly conserved between the species despite their genetic and geographic distance. The developmental rates of the two species scale linearly, in agreement with the species’ different growth temperatures. When we scale the developmental rates of the two species, we reveal a remarkable conservation of relative expression dynamics. Thus our study illuminates the dynamic properties of biological regulatory systems and their ability to control relative dynamics accurately despite genetic and growth condition differences.
Comparing gene expression profiles between Pl and Sp can identify both conserved and diverged expression patterns and suggest similarity and changes in circuits’ connectivity. High resolution time courses in Sp were measured by nanostring up to 48 hpf in this species [26], which includes the time interval 0–30 hpf in Pl (Fig 1A). While comparing actual mRNA levels between species is difficult due to the different methods used [26], comparison of initiation times and relative gene expression levels is possible. In Fig 3, we present comparative expression profiles of the studied genes separated into five regulatory circuits that initiate the specification of the skeletogenic mesoderm (Fig 3A–3C), the aboral non-skeletogenic mesoderm that form pigment cells (Fig 3D–3F), the endoderm (Fig 3G–3I), the aboral ectoderm (Fig 3J–3L) and the oral ectoderm (Fig 3M–3O). Sp expression profiles are taken from Materna et al, 2010, [26] (running averages of two biological replicates measured by nanostring technique, the data is available at http://vanbeneden.caltech.edu/~m/cgi-bin/hd-tc/plot.cgi). The circuit diagrams are based on experimental validations that include perturbation and cis-regulatory analysis in Sp [13,15,27,33,34,36]. Below we discuss the level of conservation of each circuit between the species in the light of our temporal expression comparison and previous studies.
Embryo development generates similar morphologies despite natural genetic variation and within broad environmental conditions. This flexibility of the developmental program is essential for the survival of the species and keeping a wide genotypic pool adaptable in a changing environment. Understanding the properties of the regulatory control system that underlie cell fate specification is a key to the mechanistic understanding of this developmental stability. Here we studied the reproducibility and conservation of expression dynamics of regulatory circuits in two sea urchin species that shared common ancestor about 40 million years ago and inhabit distinct geographic habitats. Embryo size, cell types and morphologies of these two species are highly similar despite their genomic and geographic distance (Fig 1). Our studies illuminate tight control of gene activation timing within the species (Fig 2) and a striking similarity of relative dynamics revealed by scaling the developmental rates of the two species and normalizing gene expression levels (Fig 5). The regulatory systems that enable this reproducibility and conservation are the underlying mechanisms of morphological similarity amidst genetic and environmental variation.
The high resolution of our studies reveals tight control of initiation times that show lower variation than the variations in maximal mRNA levels between different individuals in the same species (Fig 2). Interestingly, lower variations of initiation time compared to the variation of expression levels were also detected in a comparative study of the developmental transcriptomes of two Xenopus species [35]. Previous studies in yeast provide a possible mechanistic explanation of these findings [56,57]. These studies show explicitly that the initiation of gene activation is highly similar for different levels of the activating input once the input level is above a certain threshold for long enough time [56]. On the other hand, once the gene is on, the level of gene expression is highly dependent of the level of the activating input. The molecular explanation for the different behavior of initiation timing and expression level was suggested by the same group several years before [57]. Their measurements and modeling of expression kinetics indicated that the timing of gene initiation is controlled by the slow rate of nucleosome removal from the DNA. Once the nucleosomes are removed, the level of gene expression depends on the affinity of the transcription factor binding sites and the concentration of the activating transcription factor that define the binding site occupancy and the rate of mRNA generation. Thus, the ability to buffer variations in expression level and still tightly control the timing of gene activation, possibly by using nucleosomal positioning as a threshold mechanism, could be a general property of eukaryote gene regulatory networks.
Our interspecies comparison of temporal expression profiles of key regulatory circuits revealed a high conservation of the temporal order of gene activation within the circuits but also some cases of divergence (Fig 3). Integrating the differences in temporal profiles with available perturbation and spatial expression data provides predictions for specific cis-regulatory changes within the ectodermal circuits. The highest interspecies conservation of temporal ordering and the timing of gene activation are observed in the endoderm circuit (Figs 2G–2I and 5C). This degree of conservation supports the conservation of both the architecture and the cis-regulatory logic of this circuit. The endodermal circuit is one of the most conserved circuits within echinoderms, with a similar architecture detected in the sea star that shared a common ancestor with the sea urchin ~500 mya [58,59]. The mesodermal and ectodermal networks show higher variation of circuit connectivity between the sea urchin and sea star [60–62], emphasizing the strong developmental constraints on the endoderm circuit. The constraints that define this high degree of temporal conservation could be the requirement to initiate gastrulation and the invagination of the gut at the right developmental time. Thus, high-resolution comparison of circuits’ dynamics is a good tool for the prediction of conservation and changes in circuit connectivity when the general circuit structure is known at least in one of the species.
We used gene initiation times measured in the two species to estimate a ×1.3 ratio between the molecular developmental rates in Pl and Sp (Fig 4). Apparently, a major contribution to the accelerated developmental rate in Pl is its higher culture temperature compared to the culture temperature of Sp (18°C in Pl vs. 15°C in Sp). A recent study had shown that when Sp embryos are cultured in 18°C their developmental rate increases by about ×1.24 fold based on morphological comparison, close to the ratio we obtained [3]. This is in agreement with recent studies in invertebrate and vertebrate embryos that show morphological and molecular scaling with temperature of diverse species [35,63]. A recent morphological comparison of ten Drosophila species shows that the rate of embryogenesis scales with temperature within a wide range of temperature (17.5°C-32°C) [63]. A comparative study of two Xenopus species grown in different culture temperature (28°C vs. 22°C) shows that the rate of embryogenesis scales with temperature based on morphology and on the timing of gene activation for most studied genes [35]. Thus, the ability to adapt to different temperatures by scaling the developmental rates without distinct morphological phenotypes is a common property to both vertebrate and invertebrate species.
Our studies reveal remarkable interspecies conservation of expression dynamics when the developmental rates of the two species are scaled and gene expression levels are normalized (Fig 5). This demonstrates an impressive ability of biological developmental systems to tightly control gene activation timing and relative expression dynamics despite genetic and growth conditions differences. This raises the question: Is the observed conservation an outcome of a strong negative selection against genetic changes of regulatory circuits or due to the structure of regulatory circuits that buffers genetic and environmental changes? We tend to support the second option and the ability of the regulatory system to overcome expression noise. This could be achieved by noise filtration mechanisms, e.g., the threshold activation suggested above, or by the use of network motifs that define different levels of sensitivity to upstream variation. For example, computational studies show that positive feedback circuitry is more efficient than other architectures in buffering noise in the inducing signal while keeping high responsivity to the level of the signal [64,65]. On the other hand, incoherent feedforward motifs can generate consistent response to activating input that depends mostly on fold changes in input and not on noisy absolute protein levels [64,66–68]. Apparently, this flexible design of gene regulatory circuits enables them to conserve similar expression dynamics and specify similar cell types while allowing the species to keep a broad genotypic variance and survive through changing environmental conditions.
Adult sea urchins were supplied from a mariculture facility of the Israel Oceanographic and Limnological research in Eilat. Sea urchin eggs and sperm were obtained by injecting adult sea urchins with 0.5M KCl. Embryos were cultured at 18°C in artificial sea water. Total RNA was extracted using Qiagen mini RNeasy kit from embryos at indicated time points. 1 μg of total RNA from each time point of each three independent biological replicates was used to generate cDNA using BioRad i-script kit and subsequently used for QPCR.
Initiation times, t0, for both Fig 2D and Fig 4 were estimated by the use of the sigmoid function: Log(mRNA(t)) = a − b/(1 + exp(c(t − t0)) as in Yanai et el, 2011 [35]. The sigmoid was fit using Matlab’s Curve Fitting Toolbox, using the nonlinear least-squares method. For all genes R2>0.94, except from PlDlx repeat #3 that had R2 = 0.88. To calculate Pearson correlation between Sp and Pl time course, we first scaled the developmental rates in the two species using a factor of ×1.3 between Pl and Sp. The exact time points we compared are given in S1 Table. Then we calculated Pearson correlation between the averaged expression levels in Sp and Pl using the CORREL function in excel.
|
10.1371/journal.pcbi.1000816 | Using Entropy Maximization to Understand the Determinants of Structural Dynamics beyond Native Contact Topology | Comparison of elastic network model predictions with experimental data has provided important insights on the dominant role of the network of inter-residue contacts in defining the global dynamics of proteins. Most of these studies have focused on interpreting the mean-square fluctuations of residues, or deriving the most collective, or softest, modes of motions that are known to be insensitive to structural and energetic details. However, with increasing structural data, we are in a position to perform a more critical assessment of the structure-dynamics relations in proteins, and gain a deeper understanding of the major determinants of not only the mean-square fluctuations and lowest frequency modes, but the covariance or the cross-correlations between residue fluctuations and the shapes of higher modes. A systematic study of a large set of NMR-determined proteins is analyzed using a novel method based on entropy maximization to demonstrate that the next level of refinement in the elastic network model description of proteins ought to take into consideration properties such as contact order (or sequential separation between contacting residues) and the secondary structure types of the interacting residues, whereas the types of amino acids do not play a critical role. Most importantly, an optimal description of observed cross-correlations requires the inclusion of destabilizing, as opposed to exclusively stabilizing, interactions, stipulating the functional significance of local frustration in imparting native-like dynamics. This study provides us with a deeper understanding of the structural basis of experimentally observed behavior, and opens the way to the development of more accurate models for exploring protein dynamics.
| As more protein structures are solved, we are able to perform a more critical assessment of the relationship between protein structure and dynamics, and to gain a deeper understanding of the major determinants of structural dynamics. Here we perform a systematic study on a set of proteins structurally determined by NMR spectroscopy. The dynamics are analyzed using elastic network models and a novel method based on entropy maximization to demonstrate that properties such as contact order and secondary structure do play a role in defining the experimentally observed covariance data. Most importantly, an optimal description of observed cross-correlations requires the inclusion of destabilizing, as well as stabilizing, interactions, stipulating the functional significance of local frustration in imparting native-like dynamics.
| Associated with each protein fold is a set of intrinsically accessible global motions that arise solely from the 3-dimensional geometry of the fold and involve the entire architecture. For a number of systems it has been shown that these intrinsic motions play an important role in protein function [1], facilitating events such as recognition and binding [2], [3], catalysis [4]–[6] and allosteric regulation [1], [7], [8]. The time scales of these cooperative motions are usually beyond the reach of conventional MD simulations. They are modeled instead with coarse-grained techniques that omit the finer details of atomic interactions.
The elastic network model (ENM) is an example of a coarse-grained model that has enjoyed considerable success in predicting global dynamics of proteins and other macromolecules. The central idea behind the ENM is that, in the vicinity of a minimum, the potential energy landscape of a biomolecular system can be approximated by the sum of pairwise harmonic potentials that stabilize the native contacts. In the simplest ENM, the Gaussian network model (GNM) [9], each node of the network is identified by an amino acid, and each edge is a spring that provides a linear restoring force to deviations from the minimum-energy structure. The system's dynamics is therefore expressed in terms of the normal modes of vibration of the many-bodied system about its equilibrium state; and dynamical information about the protein, such as the expectation values of residue fluctuations or cross-correlations, is uniquely defined by the network topology.
A few prevalent methods are used for constructing ENMs, but most have at their hearts two underlying assumptions: The springs are all at their rest lengths in the equilibrium (native) conformation, and the force constants decrease with the distance between nodes, among other variables. In the earliest models [9], [10] and the anisotropic network model (ANM) [11]–[13], force constants were taken to be uniform for all nodes separated by a distance less than a specified cutoff distance and zero for greater distances. In parallel, models were proposed in which the force constants decay exponentially [14], [15] or as an inverse power of distance [16], [17], or where stronger interactions are assigned to sequentially adjacent residues [8], [16], [18]. Although such modifications can lead to modest improvements in the agreement between ENM predictions and certain experimental data, there is still no clear “best” method for assigning force constants in an ENM.
A common approach for assessing the performance of ENMs or estimating their force constants has been to compare the ENM-derived autocorrelations of residue motions to the corresponding X-ray crystallographic B-factors or the mean-square fluctuations (MSFs) in residue coordinates observed between NMR models. Because the slow modes have the largest amplitudes, often the focus of study has been a narrow band of the slowest modes. The ENM slow modes have indeed been shown to agree well with those predicted by detailed atomic-level force fields and with experimentally determined dynamics [19], [20]. However, the majority of the dynamical information conveyed by the ENM is contained in the residue cross-correlations, and this information has been largely overlooked during comparisons of ENM results to experimental data. Further, the subtle and complex dynamics of the structures that lie beneath the gross global motions are ignored when only the slowest modes are considered. Mid- and high-frequency modes are predicted with relatively lower confidence by ENMs, but these modes may be important for coordinating the finer motions of the molecule while the slower modes orchestrate its global rearrangements [21]. Finally, while the ENM-based studies have shown that the network topology is the dominant factor that defines the collective modes, especially those in the low frequency regime, there may be other structural properties (e.g. secondary structure, hydrogen bond pattern, distance along the sequence/chain between pairs of interacting residues) that are not accounted for by ENMs but which may provide a more realistic description of equilibrium dynamics, if accurately modeled.
Here we examine the ensembles of structural models determined by NMR for 68 proteins and evaluate for each ensemble the covariance in the deviations of residue-positions from their mean values. We present a technique for optimizing ENM force constants within a pre-defined network topology so as to provide the most accurate representation of the experimentally observed covariance data. Our method is based on the concept of entropy maximization: Briefly, when inferring the form of an unknown probability distribution, the one that is least reliant on the form of missing data is that which maximizes the system's entropy subject to constraints imposed by the available data [22], [23]. This method has been applied to a variety of biological problems, including neural networks [24], gene interaction networks [25], and protein folding [26].
The resulting auto- and cross-correlations in residue fluctuations are used to build an ENM-based model with optimal force constants (OFCs). It can be shown (see [25] and Methods) that when the constraints of the maximization are pair correlations, the probability distribution takes a Gaussian form. Further, the only terms that contribute to the probability distribution are those that correspond to pairs with correlations that are explicitly considered as constraints on the entropy maximization. In terms of the ENM, this means that for a given network topology, there exists a unique set of force constants that exactly reproduces the experimentally observed cross- correlations between all pairs of interacting residues, along with their autocorrelations (or MSFs).
Notably, our technique captures the physical significance of factors such as sequence separation and spatial distance which have been empirically found to influence force constant strengths. Sequence separation is expressed in terms of contact order, i.e., the number of residues along the sequence between two residues that are connected by a spring in the ENM. Further, our analysis benchmarked against a test set of 41 NMR ensembles of proteins suggests additional factors, including hydrogen bond formation and secondary structure type, which should also be incorporated in the ENMs for a more accurate description of experimental data. It also identifies factors that are of little consequence insofar as the collective dynamics near equilibrium conditions are concerned. Amino acid specificity turns out to be one of them; diffuse, overlapping distributions of OFCs are obtained for different types of amino acids, precluding the assignment of residue-specific OFCs. A modified version of the GNM, mGNM, that accounts for these factors is proposed and is verified to perform better than existing models especially in reproducing cross-correlations. Finally, the study highlights the importance of higher modes and the role of frustration in protein dynamics, the implications of which are discussed with regard to model development and protein design.
The training set of 68 proteins structurally characterized by NMR and deposited in the Protein Data Bank (PDB) [27] (Table S1) contains a total of 252,775 possible pairwise interactions (based on the combination of all pairs of residues), of which 43,118 (17.1%) fall within the 10Å cutoff. Upon optimization, a mean force constant of 6.23 kcal/mol/Å2 was found, averaged over all pairs and all proteins. Notably, this value is on the same order as typical uniform ENM force constants [8], [28], and provides an estimate of the strength of generic inter-residue interactions in native folds. To eliminate environment-specific effects and allow for the compilation and comparative analysis of the results for all proteins, we normalized the force constants such that the average force constant magnitude in each protein is unity. The resulting normalized OFCs range from −10.0 to 31.1, in dimensionless units, with a mean of 0.430 and a standard deviation of 1.831. Most (71%) of the force constants have absolute magnitude less than 1.0. Figure 1A displays the distribution of OFCs as a function of the distance dij between the interacting pairs of residues i and j, and colored by contact order k. k designates the sequential separation between residues i and j, k = 1 corresponding to bonded pairs. The inset in Figure 1A displays the dependence of the average magnitude <|γij|> on distance.
A closer examination of the influence of contact order on the OFCs yields the histograms displayed in Figure 1B. Whereas most OFCs are generally small and distributed evenly around zero, those associated with bonded interactions tend to be positive and large, with a mean value of 2.898 and standard deviation of 3.009 (see Figure 1, black dots). These large positive values reflect the almost rigid 3.8Å distance restraints on the backbone pseudo-bonds (virtual Cα-Cα bonds), consistent with the fact that the peptide bond dihedral angle ω is confined to the trans state, and consequently, in the absence of rotatable bonds the distance between the consecutive α-carbons is almost fixed.
Second-neighbor (k = 2) interactions tend to be negative, with mean −0.211±1.436 (red dots in Figure 1A and red histogram in Figure 1B). They obey a unique distance dependence (Figure 1C, red curve), suggesting that 2nd neighbors closer than a certain distance are generally too strained. Likewise, those stretched beyond a certain separation exhibit negative force constants. These interactions add frustration to the system: They tend to favor conformational changes away from the equilibrium structure, but only in a manner that does not violate the more magnanimous k = 1 restraints. Taken together, the k = 1 and k = 2 interactions suggest a flexibility of virtual bond angles, which allows adjacent (first neighboring) residues along the sequence to retain almost rigidly their separation while second neighbors tend to move with respect to each other.
The k = 3 interactions (blue dots in Figure 1A), on the other hand, are positive (0.385±1.366) indicating a dynamic correlation between adjacent virtual bond angles. More detailed analysis shows that in this case there is a weak tendency of 3rd neighbors to be destabilized when their distance approaches 10Å (Figure 1C, blue curve). A similar trend is observed in the case of 2nd neighbors, when they approach their maximal separation (∼7.4 Å) allowed by chain connectivity. These observations point to the instability of the conformations that strain the backbone.
The k = 2 interaction type and strength depend on the distance between residues i and i+2 (Figure 1C). If the residues are separated by 6Å or less, γij tends to be strong and negative, and the correlation between k = 1 and k = 2 force constants is −0.386; for distances of more than 6Å, the correlation with k = 1 drops to −0.100. This suggests the importance of secondary structure in protein dynamics, which will be our focus next.
In helices, second neighbors tend to be separated by about 5.47±0.20Å, compared to 6.66±0.41Å in strands. As can be seen from the red curve in Figure 1C, the former separation coincides with the minimum (i.e., largest negative value) in the OFC curve, which is also consistent with the red histogram displayed in Figure 2B for α-helices. The positioning of α-carbons i and i+2 along an α-helical turn requires the dihedral angles φ and ψ on both sides of Cαi to assume narrowly distributed values in the Ramachandran space and entails relatively tight packing of side chains, which may not be sufficiently stable per se, unless stabilized by hydrogen bonds formed between the adjoining residues on both sides. No such effect is discerned in 2nd neighboring residues on β-strands, given that the corresponding dihedral angles are more broadly distributed, and the backbone conformation allows for favorable interactions between every other side chain.
Notably, 3rd neighbors on β-strands tend to exhibit negative OFCs (Figure 2C). The Cαi-Cαi+3 distance of 8.796±1.408 Å falls in the regime of negative force constants (see the blue curve in Figure 1C). In the case of helices, third neighbors are located at a distance of 5.230±0.531 Å, and experience favorable interactions on a local scale (Figures 1C and 2C). The flexibility of the β-strand k = 3 contacts and the rigidity of the β-strand k = 1 and k = 2 contacts suggests that strands have a propensity for twisting motions.
Hydrogen bond formation is also found to have a strong influence on the OFCs. Using the DSSP [29] algorithm, we determined secondary structures for residues in our dataset and found that the interactions between hydrogen-bonded residues tend to be larger than those between residues that are not hydrogen-bonded (see Figure 2D), which strongly supports the physical realism of the derived OFCs. In α-helices, the average OFC for k = 4 interaction representative of hydrogen-bonded residues on consecutive turns is 0.962±1.341, compared to 0.137±1.008 for all other k = 4 interactions. Similarly, interactions between hydrogen-bonded partners in extended strands or isolated β-bridges have values around 1.801±2.321, compared to 0.412±1.817 for other interactions, thus more than counterbalancing the destabilizing interactions between 3rd neighbors. In both cases, the distributions for hydrogen-bonded and non-hydrogen-bonded interactions overlap significantly but are distinct, with Kolmogorov-Smirnov [30] probabilities of less than 10−44. This sensitivity to atomic-level details is missing in many coarse-grained ENMs, but it is an essential component of the potential energy.
Clearly, despite the existence of destabilizing interactions on a local scale, the overall structure is stable, i.e., the native structure is a global energy minimum (as also confirmed mathematically; see Methods) because these destabilizing pairwise interactions are more than counterbalanced by other stabilizing interactions. For example, there is a weak (−0.274) anti-correlation between the k = 1 and k = 2 force constants, and more significant anti-correlations between k = 2 and k = 3 (−0.689) and between k = 4 and k = 5 (−0.614) (See Table 1). In particular, when residues i and i+2 are in helices, the force constants corresponding to the interactions between first and second neighbors exhibit a correlation of −0.641 (see also Figure S1). The third and fourth neighbors on α-helices, on the other hand, are distinguished by their strong stabilizing interactions (Figure 2C and D). Similar effects occur between 2nd and 3rd neighbors in β-strands, and in all cases hydrogen bonds appear to make significant contributions to the overall stability. The presence of these (anti)correlations suggests that on a local scale there is a subtle balance between favorable and unfavorable interactions that is instrumental in determining the marginal stability of the molecule as well as its collective motions about the equilibrium structure.
We analyzed the dependence of the OFCs on amino acid type and coordination number. The distribution of force constant strengths exhibit some variations by amino acid type as can be seen from the heights and widths of the distributions in Figure S2, but there is no specific correlation of force constant values with amino acid type. Although each amino acid has a unique distribution of force constant strengths, all of these distributions overlap to a large extent, so that accurately predicting interaction strength based on amino acid type is not possible. This observation agrees with the longstanding argument that the global dynamics of solvated proteins are structure-based, and not sequence-based. We note that the insensitivity of force constants to amino acid type does not imply that all contacts contribute equally to the free energy, but that the deviations from their equilibrium positions experience comparable resistance. In terms of energy function, the depths of the energy minima may dependent on amino acid types, but the curvatures of the energy profiles near the minima do not exhibit residue-specific features at this coarse-grained level of representation.
As was seen through the large values of the bonded interactions, physical constraints directly impact the interaction values. We therefore expect the OFCs to be greatest in magnitude for the spatially constrained residues in the protein interior, and the mean-square fluctuations to decrease with the coordination number. Indeed, there is a modest (0.508) correlation between the magnitudes of the bonded interactions and the coordination numbers of the nodes they join. There is a stronger (−0.582) (anti)correlation between the coordination number and self-interaction, and a very strong (−0.909) one between a residue's self-interaction and the sum of its interactions with its first neighbors. The weight of the node, defined as the sum of the magnitudes of its edges, relates inversely to its MSF in much the same way as the degree of a node in GNM relates to its MSF (Figure S3).
Although the force constants vary in value at all distances, we were curious to examine in more detail whether there exists an underlying trend that describes the force constant magnitude as a function of distance between residues. We calculated the average absolute magnitude of the force constants as a function of residue separation (see Figure 1A, inset) and examined the functional form of this distance dependence. Using a function of the form as proposed by Hinsen [14], we find the highest correlation of only 0.339 when the distance r0 is 6.805Å, which is about twice the proposed value of r0 = 3.0Å for non-bonded force constants. Fitting the average magnitude to a function of the form , we find the best fit (cc = 0.356) using an exponent of α = 1.953, which is remarkably close to the exponent α = 2 suggested by Jernigan and coworkers [17]. Although the trend is for the average magnitude of force constants to decay with distance between nodes, the correlations are not very strong and the abundance of noise in the force constants prohibits the identification of a definitive function with which they universally decay. Figure 1C shows that the distance dependence also varies with contact order.
We compared the collective dynamics calculated with GNM to those found via OFCs (shortly referred to as OFC-GNM), with regard to the level of agreement achieved with experimental data. The computed covariance matrix contains three types of elements: diagonal, interacting (nodes joined with an edge) and non-interacting. Diagonal elements are representative of the MSFs of individual residues, and off-diagonal terms represent the cross-correlations between the fluctuations of pairs of residues. Table 2 summarizes the level of agreement of the two methods with the experimentally observed covariances. Notably, the optimized model provides a more accurate description of not only MSFs and cross-correlations between connected nodes, but also the cross-correlations between pairs of residues that are located farther apart in the structure. As shown in Table 2, experimental covariances between non-interacting residues have a correlation of 0.759 with the covariances predicted by OFC-GNM, compared to −0.014 for GNM.
One attractive feature of GNM is its ability to provide results that are robust against minor changes in structure or network topology. To test the resilience of OFC-GNM dynamics, we set small force constants identically to zero and re-calculated the covariance matrix. When the smallest 5% and 10% of the interactions are discarded, the correlation between OFC-GNM and experiment drops from 0.967±0.020 to 0.407±0.443 and 0.238±0.347, respectively. Unlike the GNM, the optimized model is therefore quite sensitive to the existence or loss of weak interactions. We also examined the robustness of the modes in the low frequency regime. The values in parentheses in Table 2 shows that the top ranking five modes computed with the OFC-GNM yield good agreement with their experimental counterpart, whether the GNM cross-correlations exhibit a considerable decrease in their level of agreement with experiments.
We briefly investigated whether the trends observed in the optimized force constants can be used to create a more effective ENM. Using a separate set of 41 proteins (Table S2), we tested the effects of incorporating bonded interactions, second neighbor interactions and hydrogen bonding into the ENM. The results, summarized in Table 3 and Table S3, indicate that including these properties mildly improves the agreement of the ENM with observed covariances for the test set. We obtained the best agreement when bonded interactions and hydrogen bonded interactions are increased in magnitude and second-neighbor force constants are negative. One set of parameters for this model, which we refer to as modified GNM or mGNM, is given in Table 3.
At present, there are copious NMR and X-ray data available from which we can extract information on protein equilibrium dynamics, and the current state of molecular dynamics is such that one can likewise approximate equilibrium ensembles of small proteins in silico. By developing coarse-grained models that reproduce these dynamics, we are able to deepen our understanding of the factors that influence protein folding and function.
In the present analysis we selected to use NMR data that provide conformational ensembles based directly on experiments, but any covariance data could have been used, in principle. The REACH algorithm [31] identifies effective ENM force constants through an inversion of a covariance matrix derived from MD simulations. Similarly, the heteroENM [32] utilizes an iterative algorithm to similarly fit the force constants with MD-derived covariances. The advantages to using MD-derived covariances are precision and flexibility. Because the locations of all atoms in an MD run are known to machine precision in each simulation frame, the covariance between even the most distant atoms, such as those separated by several nanometers, can be exactly calculated within the context of the simulation. Further, MD simulations permit in silico alterations to the system under study, allowing one to find effective force constants that are specific to any environment that can be simulated. This is a boon in particular to those who wish to study the global dynamics and interactions of multiple large molecules. On the other hand, there are some shortcomings of MD that make it an unattractive option for developing an ENM. First, MD is itself a theoretical model, and the performance of any MD-based ENM is limited by the accuracy of the force field: Inaccurate MD results beget inaccurate ENM results. Second, MD is stochastic in nature, insofar as simulations of identical systems starting from different initial states may produce different results due to sampling inaccuracies. Finally, MD is generally applicable only for short (<1µs) simulations. Covariances calculated over a short time should not be assumed to remain valid when the timescale is increased by several orders of magnitude.
Amino acid covariances are calculated here from experiments, specifically NMR structural data. A few well-studied proteins have been crystallized in multiple states – such as those bound to different ligands – allowing residue covariances to be calculated from X-ray data. Although a growing body of work suggests that functional states assumed by the proteins under different conditions are captured in multiple crystal structures [33]–[36], such multiple X-ray crystallographic structures have been determined for a few well-studied proteins only, and in most cases proteins crystallized in diverse states may not be representative of the native ensembles of conformations accessible to the protein. A more abundant source of protein conformational ensembles is NMR data. The use of various NMR techniques in determining solution dynamics of proteins has been reviewed extensively (see, for example, [37], [38]), and a number of techniques have been proposed for inferring native-state protein ensembles from NMR data [39]–[43]. Covariances calculated from NMR ensembles have been shown to agree well with MD [44], X-ray B-factors [45], [46] and covariances between multiple crystal structures [33]–[36]. NMR data are not, however, without their shortcomings: NMR ensembles may be affected by the sparsity of data and conformational variations found in solution, and as such they necessarily contain noise and do not purely reflect the native state ensemble. As the NOE intensities that are used to define structures decay rapidly with interatomic distance, long-ranged interactions are a likely source of noise in NMR covariance data. Force constant optimization methods that rely on full covariance data [31], [32] retain this noise. We were able to identify the major determinants of the effective force constants that describe the collective dynamics of proteins by resorting to a rigorous entropy maximization procedure that addresses such uncertainties.
Strikingly, a subtle interplay between stabilizing and destabilizing interactions has been disclosed, which depends on contact order, secondary structure and hydrogen-bond-formation properties. Although all of the proteins that we have analyzed are relatively small, the physical basis of the factors impacting force constant strength leads us to believe that our results hold for larger proteins as well.
The OFCs are derived from existing structural data, and in this respect our work is similar in spirit to the extraction of knowledge-based potentials from known structures [47]–[53]. The present study differs, however, in four ways: First, previous studies aimed at evaluating the effective potentials of mean force that determine the equilibrium state/energetics of native structures, and they were used in evaluating folded or docked conformations. Here, the goal is to assess the effective force constants that determine the collective fluctuations away from the equilibrium state, which are used in evaluating the equilibrium dynamics. Second, the training dataset consists of distinct proteins' structures in the former approach, whereas here ensembles of conformations corresponding to a given protein are analyzed. Third, the former group of studies counts the probabilistic occurrences of inter-residues pairs (or pair radial distribution functions) to derive potentials of mean force using inverse Boltzmann law; here, the departures in coordinates from their mean values are examined, and optimal spring constants are evaluated from an entropy maximization scheme, which is appropriate for sparse data. Fourth, the knowledge-based potentials evaluated in previous studies are residue-specific, whereas the OFCs show no significant dependence on amino acid type. This final observation is in accord with the concept that amino acids influence the fold, and the fold influences the dynamics.
In our calculations we intentionally used a slightly longer cutoff distance (10Å) than those determined to optimally reproduce B-factors (7–8Å) [19], [54]. Our reasoning was that, if a shorter cutoff distance is better, then force constants for residues that are far from each other will tend to be close to zero. Although we find that the average magnitude of the force constants decays with distance, we do not find that the force constants all drop sharply to zero after some distance. GNM consistently predicts global protein motions that agree with experimental observations, using a uniform force constant. It would therefore not have been unexpected to find that the OFCs tend to cluster about a single non-zero value. Instead, we find that the OFCs adopt a range of values centered about zero, and that the strongest indicators of force constant strengths are contact order and backbone hydrogen bond formation propensities.
The difference between the predictions of the GNM and observed protein motions is illustrated in the three examples of Figure 3, selected from the test set (Table S2). The three curves therein represent the MSFs of residues based on five slowest modes derived from NMR data (black, solid), predicted by the GNM (red, dashed), and predicted by the mGNM (blue curve). As the GNM is based entirely on the protein's folded topology, it tends to instill the most motion in the least connected nodes, e.g., chain termini or the most exposed loop regions. However, the size of the motion may depart from those indicated by NMR models, and mGNM tends to yield a better agreement with NMR data. Application to the complete test set of NMR ensembles confirmed that the correlation with experiments is improved even when contact order, distance dependence and hydrogen bonding are incorporated into the GNM without laboriously optimizing the force constants (Table 3). The fact that these physically meaningful effects emerged independently from our entropy maximization calculations validates our approach to some extent. Less expected was the prominence of negative force constants.
Overwhelmingly, the methods of ENM construction rely on two assumptions that guarantee physically plausible behavior, but which may be unwarranted. The first is that all springs are at their rest lengths in the equilibrium conformation, and the second is that all spring constants are positive. Taken together, these assumptions are sufficient, but not necessary, to guarantee that any deformations will increase the system's energy. Our optimization procedure naturally produces interactions that are physically equivalent to springs of negative force constant, but so long as the interaction matrix remains nonnegative definite, the system is in a stable equilibrium and negative force constants are acceptable. The existence of negative force constants reflects the implicit frustration of folded proteins; the backbone restrains the protein to certain compact folds, and not all native state contacts are guaranteed, nor should be expected, to be favorable. Negative force constants make the structure prone to certain deformations that may not be preferred when all force constants are positive. Frustration in proteins results in a rough free-energy landscape that gives rise to folding intermediates and alternative conformations [55]–[58], and calculations involving Go-like potentials, or knowledge-based potentials [49] reveal the requirement to include both stabilizing and destabilizing interactions for an accurate assessment of the folding behavior or stability of proteins. The balance between attraction and repulsion endows proteins with both the sensitivity and the stability that are prerequisite for proper function [59]. We find that the (i, i+2) interactions are the most likely to be at a local maximum, promoting a change in the angle between (i, i+1) and (i+1 i+2) pseudobonds.
When we include factors such as hydrogen bonds and negative k = 2 force constants in the GNM, the improved agreement comes in the off-diagonal components of the predicted covariance matrices. Cross-correlations are often overlooked when assessing ENM predictions, but they are essential because they carry information on how the molecule moves as a whole. The autocorrelations that indicate how much individual residues move are each the sum of positive terms and are necessarily dominated by the slower modes. The cross-correlations, on the other hand, are sums of positive and negative terms and are therefore susceptible to the influence of higher modes. Slight modifications to the GNM, such as those that we have introduced in mGNM, do not perturb the network enough to significantly alter the slow modes (Figure 3), but their effects are captured in the higher modes.
Although the slowest modes get the most attention because of their prevailing role in determining the molecule's global motions, the high-frequency modes have shown to be important for identification of conserved residues and folding cores [60]–[63]. Mid- to high-frequency modes are also crucial to all aspects of protein behavior. Allosteric transitions have been shown to occur largely along the slowest modes, but higher modes are essential for the complete transition [64]. Similarly, a protein's response to external perturbations [28] is dependent on all modes, not only the slowest few. An ENM that accurately captures all modes has an enhanced ability to predict large-scale conformational changes, and our technique opens the door to developing better ENMs based on experimental data.
Figure 4 shows pairwise comparisons of the eigenspaces spanned by the slowest modes of various models. Panel A shows the correlation of mobilities as a function of the fraction of modes used in the comparison, and panel B shows a similar plot of the overlap of the eigenspaces (see Methods). The green and black curves relate the GNM and mGNM, respectively, to the experimental covariance matrices. The average mobility correlation of GNM with the experimental covariances peaks at 0.76 when 12% of the modes are considered and then falls as more modes are taken into account, indicating that the predicted modes in the mid-to-high frequency range introduce errors manifested by departures from experimental data. The modified GNM does not exhibit this decline, but remains steady even as higher modes are considered, indicating that the higher modes of the mGNM do not adversely affect the predicted mobility of the system. Comparison of GNM to mGNM (blue curves) shows that the slowest 2% of modes of these models are highly overlapping, but that the similarity decreases as more modes are considered. The modifications of mGNM therefore do not affect the slowest mode, which is presumably determined by the fold topology, but they change the shapes of higher modes.
Interestingly, the overlaps of the GNM and the mGNM with the modes of the covariance matrix are almost identical (compare green and black curves, panel B), suggesting that, despite the improved agreement in mobility, the modifications that we have made to the mGNM still fail to precisely capture the system's overall dynamics. Although some additional improvement may be gained by fine-tuning the parameters of the mGNM (last line, Table 2), the similarity in slow modes of GNM and mGNM once again indicates that fold topology has the dominant influence on the mode shapes.
For our training set, we start with a set of 68 proteins (Table S1), each of which has at least 40 NMR structures available. The proteins in our set have between 43 and 151 residues. For each protein we calculate the mean structure from the NMR ensemble, and we select as a representative structure the NMR model that has lowest root-mean-square deviation (RMSD) from the mean. The test set consists of 41 proteins (Table S3), each having at least 40 NMR models and no fewer than 50 residues.
We seek to determine the pairwise interactions that optimally describe observed covariances between residues while minimizing the assumptions about the form of missing data. For this, we turn to the principle of maximum entropy, which states that when inferring the form of an unknown probability distribution from a limited number of samples drawn from the distribution, the method that is minimally reliant on the form of missing data is entropy maximization. Here the central idea is outlined in terms of the GNM.
Consider a protein of N residues for which m structures are known (e.g., m models deposited in the PDB for a given protein resolved by NMR spectroscopy). The position of residue i in structure k is given by the vector, , the average position of residue i in all structures that have been optimally superimposed (to eliminate external degrees of freedom) is defined as , and the vector displacement of residue i in structure k from the average is . In the GNM, we replace the vector displacement ΔRi with the scalar displacement Δri, which is defined such that and .
Now define the set π of q pairs of residues such that for all pairs we know the covariances , but for pairs we do not know . We seek the probability distribution that produces the known covariances while remaining minimally presumptive about the form of missing information. According to Jaynes [22], [23], this is the distribution that maximizes entropy subject to the constraints that some pair covariances are known and must be reproduced.
Defining the N-component vector, , the probability distribution that we seek is ρ(Δr), and it has the properties(1)(2)We define the entropy , and impose the above constraints as Lagrange multipliers:(3)Maximizing ζ with respect to ρ(Δr), we find(4)or, defining Z = e1+λ. and the matrix K with elements Kij = μij,(5)Direct integration leads to the result(6)which is the well-known relationship between covariances and pair interactions. The probability distribution in Equation 5 is of the same Gaussian form as the probability distribution from GNM [9], but with the interaction matrix K replacing the product of the spring constant γ and the Kirchhoff matrix Γ. Thus, the off-diagonal elements of K correspond to the negative spring constants: Kij = −γij, where γij is the force constant of the interaction between residues i and j. We are claiming knowledge for the covariance information of only the q residue pairs in the set π, so K cannot be found through the simple inversion of the covariance matrix. The matrix K has a well-defined form: the elements are the Lagrange multipliers that have imposed the above constraints on the covariance and may therefore be different from zero; the elements are identically zero. Mathematically, this means that there are no constraints on the covariances of pairs . We then have partial information for both K and K−1: The elements and are known, and the elements and are to be determined. The solution can be found through an N-dimensional minimization as follows. Consider the function(7)of two symmetric square matrices K and C. Differentiation with respect to each element of K reveals that there exists a single minimum at(8)Because Cij is undefined for all , we can allow , automatically satisfying the minimization condition for elements not in π. The remaining elements of K can be found by starting with a matrix of the general form of K and iteratively adjusting the non-zero elements against the gradient given in Eq. 8 until the minimum is reached. Optimization is achieved when for all interactions. This criterion appears to be sufficiently strict: Reducing the optimization constant from 0.01 to 0.005 changes the spring constants by less than 1%, on average. The optimization is somewhat computationally intensive: Each step requires an O(N3) matrix inversion, and the minimization completes after about 104 steps, making this technique best-suited for small proteins.
It is noteworthy that only those interactions corresponding to known covariances are optimized, and the rest remains zero. This result stems from the application of entropy maximization. Whereas many networks are capable of exactly accounting for the covariance information in the q known interactions, this is the only one that does so without prior assumptions about other covariances. Each pair interaction carries information on the covariance of two of the N nodes, so a network of more than q interactions carries information on more than q covariances. Nevertheless, all covariances can be calculated with the resultant network. Those covariances that are not known a priori and included in the calculation simply result from the optimized interactions. The matrix C is nonnegative definite by construction, and its inverse K is therefore also nonnegative definite. As a result, no deviation from the native state conformation can lower the system's energy.
The interaction matrix K has the dimensions of Å−2, and physical values for the force constants can be determined by multiplying by 3kBT, where kB is the Boltzmann constant and T is the temperature. Using this conversion, the OFCs vary between −1686 kcal/mol/Å2 and 3868 kcal/mol/Å2, with a mean of 6.23 kcal/mol/Å2. When K is scaled by a scalar constant, γ, its corresponding covariance matrix is scaled by γ−1. Thus, the mean element magnitude of the covariance matrix affects the magnitudes of the elements of the interaction matrix, such that large covariances tend to produce weak interactions. The experimental conditions under which the structures are solved influence the magnitudes of the covariances, and therefore also influence the magnitudes of the effective force constants. To reduce the bias on force constants caused by environmental specificity, the OFCs for each protein are scaled by the mean magnitude of the non-zero off-diagonal interactions in that protein.
In the GNM, each residue is a node of the network and is represented by its Cα atom. Nodes that are within a cutoff distance, Rc, are considered connected via an elastic edge. Typical values of Rc are between 7Å and 10Å. Using the N-dimensional column vector, , of displacements of the nodes from their equilibrium positions, the potential energy is found to be , where γ is a uniform force constant assigned to all interactions, and Γ is the Kirchhoff adjacency matrix, with off-diagonal elements Γij = −1 if nodes i and j are in contact and Γij = 0 otherwise. The diagonal elements of Γ are such that the sum over all elements in any row or column is identically zero. The elements of the covariance matrix predicted by the GNM are related to Γ as .
If U and V are two sets of normal modes for an N-dimensional system under different models, then we define the overlap of the first m modes of the models as , where u(k) and v(p) are the kth and pth slowest modes of U and V, respectively. Qm ranges from 0, if none of the space spanned by the slowest m modes of U can be projected onto the first m modes of V, to 1, if the two spaces overlap exactly.
The force constant between residues i and i+k is . The correlation coefficient between force constants corresponding to different contact orders is calculated as follows. First, for a contact order n<k, we define as the average force constant for all pairs between i and i+k that have a contact order of n:(9)The correlation between force constants and is then(10)Table S2 lists such correlations for contact orders in the range 1≤k≤5.
|
10.1371/journal.ppat.1006505 | HIV-1 infection depletes human CD34+CD38- hematopoietic progenitor cells via pDC-dependent mechanisms | Chronic human immunodeficiency virus-1 (HIV-1) infection in patients leads to multi-lineage hematopoietic abnormalities or pancytopenia. The deficiency in hematopoietic progenitor cells (HPCs) induced by HIV-1 infection has been proposed, but the relevant mechanisms are poorly understood. We report here that both human CD34+CD38- early and CD34+CD38+ intermediate HPCs were maintained in the bone marrow (BM) of humanized mice. Chronic HIV-1 infection preferentially depleted CD34+CD38- early HPCs in the BM and reduced their proliferation potential in vivo in both HIV-1-infected patients and humanized mice, while CD34+CD38+ intermediate HSCs were relatively unaffected. Strikingly, depletion of plasmacytoid dendritic cells (pDCs) prevented human CD34+CD38- early HPCs from HIV-1 infection-induced depletion and functional impairment and restored the gene expression profile of purified CD34+ HPCs in humanized mice. These findings suggest that pDCs contribute to the early hematopoietic suppression induced by chronic HIV-1 infection and provide a novel therapeutic target for the hematopoiesis suppression in HIV-1 patients.
| Multi-lineage hematopoietic abnormalities generally occur during chronic infection which results in a disorder of human leukocyte development and differentiation, contributing to human immunodeficiency virus-1 (HIV-1)-infection induced immune-pathogenesis in AIDS patients. Although successful antiretroviral therapy can reduce plasma viral loads to undetectable levels and ameliorate HIV-1-associated hemato-suppression, immune cell development is only partially restored. The mechanism for the abnormal hematopoiesis occurring during chronic HIV-1 infection remains unclear. HIV-1 infection may directly or indirectly functionally impair hematopoietic progenitors by either viral products or induction of persistent inflammatory responses, leading to hematopoiesis obstacles. Here, we show that HIV-1 infection significantly depleted and functionally impaired human hematopoietic progenitors in the bone marrow of both HIV-1-infected patients and humanized mice through a plasmacytoid dendritic cell (pDC)-dependent mechanism, as depletion of pDCs significantly recovered cell numbers and functions and gene expression profiles of hematopoietic progenitor cells in humanized mice in vivo. Our study clarifies a novel mechanism underlying hemato-suppression induced by chronic HIV-1 infection and provides a novel strategy to halt HIV-1 disease.
| Human immunodeficiency virus-1 (HIV-1) infection in patients leads to multi-lineage hematopoietic abnormalities, including anemia, granulocytopenia and thrombocytopenia [1,2]. Abnormalities in fetal hematopoiesis have also been reported in aborted fetuses from HIV-1 seropositive women [3]. The defect in hematopoietic progenitor cells (HPCs) or hematopoiesis induced by HIV-1 has been proposed [1,4,5]. In addition, the degree of the hematopoietic pathology correlates with the stage of disease progression [6], and end-stage disease is characterized by pancytopenia [1]. Long-term bone marrow (BM) cultures from HIV-1-infected patients exhibit low CD34+ progenitor cell growth and differentiation [7,8], indicating functional impairment of early hematopoietic progenitors. Although the successful highly active antiretroviral therapy (HAART) clearly ameliorates HIV-1-associated hemato-suppression, it does not completely restore blood cell development [9]. These observations indicate that hematopoietic failure is an important aspect of HIV-1 infection-induced pathogenesis [10].
HPCs are comprised of diverse populations, including both early and intermediate progenitors. Each subpopulation expresses distinct sets of cell surface antigens, although they all express the cell surface antigen CD34 [11,12]. Early and intermediate populations can be distinguished by the expression of CD38, with the former being negative for CD38 and the latter being positive for this antigen. Functionally, intermediate progenitors include common myeloid progenitors that can give rise to all myeloid, erythroid and megakaryocyte lineages. Due to limited access to the BM in humans, properties of human HPC subsets and their alterations in healthy and HIV-1 disease states have been difficult to characterize.
The mechanisms underlying abnormal hematopoiesis in HIV-1 infection remains unclear due to the paucity of robust animal models that mimic human hemato-suppression in vivo. Although previous studies failed to detect HIV-1 infection of HPCs, recent reports indicated that HIV-1 could directly infect HPC subsets and lead to their impairment [13–16]. In addition, HIV-1 proteins such as Nef [17] and prolonged treatment with antiretroviral drugs could also compromise hematopoietic progenitors [18]. Although these studies investigated a litany of direct and indirect causes of HIV-1-associated hemato-suppression, how HIV-1 affects hematopoiesis in vivo remains unclear.
There is emerging evidence that certain cytokines induced during inflammation have significant effects on HPCs in the BM. Type I and II interferon (IFN) [19–23], tumor necrosis factor (TNF) [24–26] and lipopolysaccharide (LPS) [27,28] directly stimulate HPC proliferation and differentiation, thereby increasing the short-term output of mature effector leukocytes. However, chronic inflammatory cytokine signaling can lead to functional exhaustion of HPCs [19,22,28]. Our previous study demonstrated that plasmacytoid dendritic cells (pDCs), the major type I interferon (IFN-I)-producing cells during acute or chronic HIV-1 infection, could inhibit viral replication while significantly contributing to HIV-1 infection-induced immune-pathogenesis, including increased immune cell death and reduced immune reconstitution of human CD45+ cells in humanized mice in vivo [29]. These findings suggest that pDCs play a pivotal role in the hemato-suppression induced by chronic HIV-1 infection.
In this study, we sought to understand the role of pDCs in HIV-1-associated hemato-suppression in a humanized mouse model in vivo. We discovered that HIV-1 infection depleted CD34+CD38- early HPCs and functionally impaired human CD34+ HPCs in the BM of patients and humanized mice with HIV-1 infection. This phenomenon was further found to be dependent on pDCs, as depletion of pDCs significantly recovered HPC cell numbers and multi-lineage colony-forming functions. Our present study therefore reveals a novel mechanism for hematopoiesis suppression induced by chronic HIV-1 infection and provides a new strategy to rescue HPC function and halt HIV-1 disease progression.
By gating on live human CD45+ cells (VLD-mCD45-) with a lymphoid morphology that lacked common markers (Lineage−) for T cells (CD3), B cells (CD19 and CD20) or NK cells (CD56 and CD16), we identified human BM-derived HPCs as CD34+ cells, which included early CD34+CD38- and intermediate CD34+CD38+ subpopulations (S1 Fig). We then analyzed the human HPCs from the BM of humanized mice at various time points after human CD34+ cell transplantation. The early and intermediate HPCs could be detected significantly at both 16 weeks and 50 weeks after CD34+ cell transplantation with relatively stable levels (Fig 1A).
To determine if human HPCs derived from humanized mice were functional, human CD34+ cells were isolated from the BM of humanized mice and human fetal livers, and colony-forming unit (CFU) assays were performed by culturing CD34+ HPCs in a complete methylcellulose medium system. Two weeks later, HPCs derived from the BM of humanized mice produced 50 colonies on average for every 500 CD34+ cells plated, similar to that of human fetal liver-derived CD34+ cells (Fig 1B). All hematopoietic lineages were generated in cultures from the BM of humanized mice. In addition, HPCs from humanized mice could proliferate and differentiate into various blood lineage cells in vitro at a similar frequency to that of CD34+ cells from human fetal livers, including colony-forming unit-granulocyte and macrophage (CFU-GM), colony-forming unit-erythroid (CFU-E) and colony-forming unit-granulocyte, erythroid, macrophage, megakaryocyte (CFU-GEMM) (Fig 1B). We then measured the proliferation capacity of HPCs by BrdU labeling in vivo and found that 8.9% of human CD34+ cells showed proliferation (Fig 1C). Notably, the CD34+CD38- early HPCs were much more proliferative, with an average of nearly 25% of cells being BrdU positive, which was significantly higher than the relatively quiescent CD34+CD38+ intermediate HPCs with 1.6% BrdU labeling (Fig 1C and 1D). These data suggest that the human CD34+CD38- early HPCs and CD34+CD38+ intermediate HPCs were both functionally developed and maintained in the BM of humanized mice.
Utilizing the robust animal model, we were able to investigate whether chronic HIV-1 infection affected human HPCs. HIV-1 infection was established in humanized mice, as measured by plasma HIV-1 RNA (copies/mL, S2 Fig). On termination, we also measured HIV-1 gag p24 expression in both T cells and CD34+ HPCs by flow cytometry. Although a previous study suggested that HIV-1 has the potential to infect intermediate CD34+CD38+ HPCs [13], we found that p24 expression was absent in BM CD34+ HPCs from humanized mice with HIV-1 infection; in contrast, CD3+ T cells showed high levels of p24 expression (10.5%) (Fig 2A). Further analysis indicated that the frequency of CD34+CD38- early HPCs was largely decreased in humanized mice with chronic HIV-1 infection, while the proportion of intermediate CD34+CD38+ HSCs was relatively expanded (Fig 2B).
Summarized data further demonstrated that chronic HIV-1 infection significantly reduced CD34+CD38- early HPCs by nearly 8-fold as compared to the non-infected animals; meanwhile, the proportion of CD34+CD38+ intermediate HPCs was increased from 76% to nearly 100% as shown in the stacked bar graph (Fig 2C). When the absolute cell counts of early and intermediate HPCs were calculated, the number of CD34+CD38- early HPCs was dramatically reduced in the BM of chronically infected animals (Fig 2D), while CD34+CD38+ intermediate HPC counts were affected mildly (Fig 2E). These results suggest a depletion of early HPCs during chronic HIV-1 infection. Importantly, we observed a similar depletion of BM CD34+CD38- early HPCs in HIV-1-infected patients. As shown in Fig 2F, the percentage of CD34+CD38- early HPCs was significantly decreased within total CD34+ HPCs in an HIV-1-infected patient when compared to a healthy control (HC). The depletion of CD34+CD38- early HPCs in human BM from HIV-1 infection was strengthened by the addition of more patients (n = 5) (Fig 2G). In contrast, the proportion of intermediate CD34+CD38+ HSCs was increased within total HPCs in HIV-1-infected patients relative to those of HC subjects (Fig 2H). These data indicated that early CD34+CD38- HSCs were preferentially depleted by chronic HIV-1 infection, and the humanized mouse is a highly relevant animal model that mimics HIV-1-induced hemato-suppression conditions in patients.
We next analyzed the effect of chronic HIV-1 infection on the homeostatic proliferation of human HPCs in humanized mice. The results indicate that proliferation of human CD34+ cells in the BM was inhibited by approximately 3-fold in chronic HIV-1 infection compared to the mock animals (Fig 3A). Consistent with the preferential reduction of CD34+CD38- HPCs, BrdU-positive CD34+CD38- early HPCs were significantly decreased by chronic HIV-1 infection; meanwhile, the proliferation of CD34+CD38+ intermediate HPCs was only mildly reduced by chronic HIV-1 infection (Fig 3B and 3C). In terms of cell numbers, the early HPC counts were significantly decreased by chronic HIV-1 infection compared with mock treatment in mice (Fig 3D), while intermediate HPC counts were only slightly reduced (Fig 3E). Thus, while both cell types were less proliferative in the presence of HIV-1, the more marked difference was observed in CD34+CD38- cells. Therefore, chronic HIV-1 infection appeared to suppress human hematopoiesis by inhibiting homeostatic proliferation of human early HPCs in the BM.
In order to assess the quality of in vivo human HPCs during chronic HIV-1 infection, we measured CFU activity of purified human Lin-CD34+ HPCs from mock- or HIV-1-infected humanized mice, including GM, E and GEMM (S3 Fig). As shown in Fig 4, HPCs isolated from uninfected mice consistently produced over 50 colonies per 500 CD34+ cells on average, whereas HPCs derived from HIV-1-infected mice produced less than 30 colonies per 500 CD34+ HPCs. Moreover, the ability of HPCs to generate all lineages, including GM, E and GEMM, was also suppressed to some extent in chronically infected mice. Therefore, the results indicated that chronic HIV-1 infection leads to the impaired differentiation of HPCs in vivo.
Increasing reports have demonstrated that chronic inflammation could lead to the functional exhaustion of BM HPCs [19,22,28]. Our recent study indicated that depletion of pDCs efficiently rescued human CD45 cell reconstitution in humanized mice with chronic HIV-1 infection [29]. Thus, we hypothesized that pDCs may be responsible for the depletion of CD34+CD38- early HPCs and their functional impairment during chronic HIV-1 infection.
To address the role of pDCs in the impairment of CD34+CD38- HPCs in HIV-1 infection, humanized mice with chronic HIV-1 infection were treated with a pDC-depleting antibody (15B) as in our previous report [29]. Similarly, the plasma viral load was increased upon the depletion of pDCs and maintained at a higher level until termination (S4A Fig). Notably, the depletion of pDCs significantly changed the percentage of CD34+CD38- early HPCs in the BM from humanized mice with chronic HIV-1 infection (Fig 5A). Pooled data further confirmed that both the percentages and cell counts of CD34+CD38- early HPCs were restored by depletion of pDCs in HIV-1-infected mice (Fig 5B and 5C). In contrast, pDC depletion did not influence the percentages of CD34+CD38- early HPCs (CD38 expression) and their proliferation in vivo indicated by BrdU expression in the BM in the absence of HIV-1 infection (S4B Fig). In addition, the cell counts of CD34+CD38+ intermediate HPCs showed only minor recovery by the depletion of pDCs during chronic HIV-1 infection, which is consistent with the slight decrease in proportion of CD34+CD38+ intermediate HPCs (Fig 5B and 5C).
Notably, only half of the animals showed a recovery in the proportion of CD34+CD38- early HPCs after pDC depletion in Fig 5B, which prompted us to investigate whether the extent of pDC depletion affected the effectiveness of rescue of CD34+CD38- early HPCs in humanized mice with chronic HIV-1 infection. We therefore divided the 13 animals into two groups; animals with less than the median percentage of CD34+CD38- HPCs were placed in the "non-rescued" group (n = 6), while the others were included in the "rescued" group (n = 7). The rescued group was found to have significantly more CD34+CD38- early HPCs and fewer CD34+CD38+ intermediate HPCs than the non-rescued group (S4C and S4D Fig). Importantly, the rescued mice showed a marked lack of pDCs in the BM and lower levels of IFN-α in plasma compared with non-rescued mice (S4E and S4F Fig). Accordingly, the rescued mice were also characterized by a higher level of HIV-1 replication than that of non-rescued mice (S4G Fig). Thus, we observed that the CD34+CD38- early HPCs was negatively correlated with pDC percentages within CD45+ cells in BM of these HIV-1 infected humanized mice with pDC depletion (S4H Fig). These data indicated that the extent of rescue of CD34+CD38- early HPCs was closely linked to in vivo pDC depletion rather than other potential causes such as HIV viral load in humanized mice with chronic HIV-1 infection.
To further qualify HPCs after pDC depletion, Lin-CD34+ cells were purified for colony-forming assays ex vivo. Cell colonies including GM, E and GEMM were found in culture (S3 Fig). The results demonstrated that pDC depletion could dramatically enhance CFU activity of the Lin-CD34+ cell population as well as increase the quantity of each colony type individually as compared with HIV-1-infected mice (Fig 5D).
To understand how pDCs contribute to the impairment of HPCs during chronic HIV-1 infection, we analyzed gene expression of human HPCs in BM from HIV-1 chronically infected humanized mice. Human Lin-CD34+ cells from BM of humanized mice were isolated and submitted for gene expression analysis by cDNA array (Fig 6A), as described in a previous study [29]. A total of 3114 genes were significantly up-regulated, and 2994 genes were down-regulated spontaneously in CD34+ HPCs from HIV-1-infected humanized mice as compared to mock-treated mice (fold change > 2, Fig 6B and S5 Fig). Astonishingly, pDCs depletion during chronic HIV-1 infection in mice restored most of the interferon-stimulating genes (ISGs) to levels found in non-infected animals (S6A and S6B Fig). Along with the recovery of ISG gene expression, only 924 genes were significantly up-regulated, and 364 genes were down-regulated in BM CD34+ HPCs with pDCs depletion as compared to mock-treated mice (fold change > 2, Fig 6B and S5 Fig). Thus, pDCs depletion resulted in a restoration of a total of 5664 genes among 6108 genes (> 92.7%) that changed during HIV-1 infection in humanized mice, drawing back the whole gene expression profile to a pattern quite similar to that in mock-infected mice (Fig 6B and S5 Fig).
We then searched the GeneGo database to identify potentially relevant pathways in the genes influenced by chronic HIV-1 infection. The hematopoietic cell lineage was the most affected pathway induced by chronic HIV-1 infection among the top 15 pathways, whereas those dysregulated genes were also significantly restored to mock levels in pDC-depleted BM CD34+ HPCs of humanized mice chronically infected with HIV-1 (Fig 6C). We comprehensively analyzed the 88 genes in the hematopoietic cell lineage pathway (S7A Fig) and found that HIV-1 infection induced a significant up-regulation of 27 genes and down-regulation of 28 genes in humanized mice relative to mock controls (S7B Fig). However, depletion of pDCs during HIV-1 infection only induced a significant change in expression of one gene in relation to mock controls (S7B Fig). Summarized data further indicated that although HIV-1 infection led to a significant change of most of the genes with more than a 2-fold change, pDC depletion could attenuate the up-regulation of genes and restored the down-regulated genes to normal levels during chronic HIV-1 infection (Fig 6D and S7C Fig). In particular, some genes related to the HPC quiescent state (DTX3L and CXCR4) [30], colony forming capacity (CD34, CD38, FLT3 and TGFBR1-3) [31], self-renewal and expansion capacity (Notch1-2, Jagged1-2 and Hes-1) [32,33], as well as several important cell death genes were significantly altered by chronic HIV-1 infection, while pDCs depletion in HIV-1-infected humanized mice largely restored the abnormal expression of these genes to a similar pattern as seen in mock-infected mice (Fig 6E and 6F, S1 Table).
Simultaneously, we also performed mRNA expression analysis using spleen-derived CD45+ cells (S8A Fig). A total of 4757 genes were significantly up-regulated or down-regulated in splenic CD45+ cells from HIV-1-infected humanized mice as compared to mock-infected mice (fold change > 2, S8B Fig). The depletion of pDCs resulted in a restoration of about 44.0% of genes (2092/4757) of splenic cells changed by HIV-1 infection in humanized mice to a pattern quite similar to that in mock-infected mice (S8B Fig). The pathway analysis indicated that systemic lupus erythematosus (SLE) was the most affected pathway induced by chronic HIV-1 infection (S8C Fig), whereas the dysregulated genes expressed by splenic CD45+ cells were not significantly restored to mock levels in pDC-depleted humanized mice chronically infected with HIV-1 (S8D Fig). We also analyzed 127 genes in the SLE pathway changed by HIV-1 infection (S8E Fig) and found that HIV-1 infection induced significant changes of 48 genes relative to mock controls. However, depletion of pDCs during HIV-1 infection also induced significant changes in expression of 115 genes in relation to mock controls (S8E Fig), indicating that pDCs depletion failed to restore the changes in gene expression by spleen CD45+ cells to normal levels during chronic HIV-1 infection (S8F Fig).
These data strongly suggest that pDCs has relatively unique effects on the numerical reduction and functional impairment of BM CD34+ HPCs in HIV-1 chronically infected humanized mice, contributing to the suppression of hematopoiesis characterized by dysregulation of gene expression profiles.
We finally tested whether IFN-I directly up-regulates CD38 expression on HPCs in vitro. As shown in S9A and S9B Fig, neither IFN-α nor IFN-β showed any significant effect on CD38 expression on CD34+ cells in vitro. In addition, IFN-I culture did not affect CD34+ HPC expansion (S9C Fig). These data indicate that IFN-I did not directly affect CD38 expression and expansion of CD34+ HPC cells in vitro.
The present study demonstrates, for the first time, that human CD34+CD38- early HSCs are subject to preferential depletion and functional impairment in vivo in the BM of humanized mice with chronic HIV-1 infection, in a pDC-dependent fashion. This study thus reveals a new target for the development of novel drugs targeting pDC activity to treat hematopoietic disorders during chronic HIV-1 infection and also demonstrates the utility of humanized mice to investigate important questions on HIV-1-mediated hematological abnormalities in the BM in vivo.
Previous studies have attempted to delineate the mechanism by which HIV-1 infection induces the impairment of HPCs, but it remains an intractable question due to the lack of a suitable experimental animal model that closely mimics human hematopoiesis during an ongoing HIV-1 infection in vivo. Here, we provide evidence that both early and intermediate HPCs are functionally developed and maintained for long-term self-renewal in the BM of humanized mice in vivo. This animal model has been demonstrated to be persistently infected by various strains of HIV-1 and develop major immune pathogenesis induced by acute or chronic HIV-1 infection as observed in HIV-1 patients [29,34–36]. Importantly, Nixon et al. made initial advances toward adopting a humanized mouse model to investigate an HIV-1 infection-induced hematopoiesis disorder [13]. Our present study underscores the rationale for utilizing a humanized mouse model to study the impact of HIV-1 infection on hematopoiesis in vivo. Taken together, these studies support the humanized mouse model as an experimentally amenable in vivo system for investigating HIV-1-associated pathogenesis.
Accumulating evidence has demonstrated that patients with long-term HIV-1 infection exhibit a deficiency in hematopoiesis [1,4,5], although the stage at which HPCs are impaired by HIV-1 infection is unclear. Nixon et al. showed that HPCs were susceptible to HIV-1 infection in vitro and in vivo in humanized mice and concluded that direct infection of intermediate CD34+CD38+ HPCs by HIV-1 adversely affected their hematopoietic potential and correlated with the observed pancytopenia in HIV-1 infected patients [13]. In this study, we found that CD34+CD38- early HPCs were preferentially depleted during chronic HIV-1 infection, which correlated with the depression of hematopoiesis development and dysregulated gene expression in bulk Lin-CD34+ HPCs. However, our study failed to detect significant productive infection of Lin-CD34+ HPCs in the BM in vivo even with pDC depletion, as the depletion of pDCs led to dramatically increased viral replication. One possible explanation for this finding is that different HIV-1 strains may exhibit discrepancies during infection of HPCs in vivo. Future studies should examine the effect on HPC subsets by infection with other HIV-1 strains with different tropisms.
The mechanisms leading to abnormal hematopoiesis have not been clearly addressed in HIV-1 infection. Aside from the direct infection of HPC subsets, pDCs are possibly critical factors leading to hematopoietic suppression during HIV-1 infection, as the depletion of pDCs rescued early HPCs and their hematopoiesis. Available lines of evidence have also demonstrated that pDCs substantially mediate detrimental effects during chronic HIV-1 infection in vivo, even while they inhibit viral replication [29,37–39]. In addition, pDCs can secrete other pro-inflammatory cytokines, including TNF-α and IL-6. These chronic inflammatory cytokines can lead to exhaustion of hematopoiesis [19,22,28]. Most importantly, pDCs are the major IFN-I-producing cells during HIV-1 infection [29,40,41]. Currently, IFN-I is perhaps the primary contributor, since depletion of pDCs completely abolished IFN-I responses in humanized mice with chronic HIV-1 infection [29]. IFN-I has been recently demonstrated to be actively involved in immune pathogenesis of chronic virus infection [39,42–45]. Our data indicate that IFN-I did not directly affect CD38 expression on HPCs in vitro during short-term culture although it possibly promote the maturation of embryonic hematopoietic stem cells [46]. Other factors associated with chronic HIV-1 infection may contribute to IFN-induced HPC depletion and will be investigated in future study. Of course, we could not exclude the possibility that HIV-1 products are involved in the dysregulation of hematopoietic development. For example, HIV-1 Nef has been found to be responsible for hematopoietic defects of the BM in HIV-1 infection, dependent on the presence and activation of the PPARγ signaling pathway [17]. Thus, pDCs may contribute to abnormal hematopoiesis during chronic HIV-1 infection directly through viral infection or indirectly via diverse cytokines. Taking these studies into consideration, we also propose that HIV-1 infection may affect HPC function through multiple mechanisms (S10 Fig). Future studies should investigate in detail the individual factors responsible for compromising hematopoietic activity.
In summary, pDCs play a pivotal role in the immune-pathogenesis and hematopoiesis depression induced by chronic HIV-1 infection. This study, therefore, provides new insight into HIV-1-induced dysregulation of hematopoiesis and provides a novel strategy for treating abnormal hematopoiesis during chronic HIV-1 infection.
Approval for animal work was obtained from the University of North Carolina Institutional Animal Care and Use Committee (IACUC ID: 14–100). The study protocol on human samples was approved by the Institutional Review Board and the Ethics Committee of Beijing 302 Hospital in China. The written informed consent was obtained from each subject. Human BM samples were obtained from adult donors with liver transplantation as healthy controls and from HIV-1-infected adult patients for pathological diagnosis. Human fetal livers and thymuses (gestational age 16 to 20 weeks) were obtained from medically indicated or elective termination of pregnancies through a non-profit intermediary working with outpatient clinics (Advanced Bioscience Resources, Alameda, CA). Written informed consent from the maternal donor was obtained in all cases under regulations governing the clinic. All animal studies were conducted following NIH guidelines for housing and care of laboratory animals. The project was reviewed by the University’s Office of Human Research Ethics, which determined that this submission does not constitute human subjects research as defined under federal regulations [45 CFR 46.102 (d or f) and 21 CFR 56.102(c)(e)(l)].
We constructed humanized Balb/c rag2-γc (DKO) mice and Nod-rag1-γc (NRG) mice (The Jackson Laboratory) in a similar manner as previously reported [36]. Briefly, human CD34+ cells were isolated from 16- to 20-week-old fetal liver tissues (Advanced Bioscience Resources, Alameda, CA). Tissues were digested with liver digest medium (Invitrogen, Frederick, MD). The suspension was filtered through a 70-μm cell strainer (BD Falcon, Lincoln Park, NJ) and centrifuged at 150 × g for 5 minutes to isolate mononuclear cells by Ficoll. After selection with the CD34+ magnetic-activated cell sorting (MACS) kit, CD34+ HPCs (0.5 × 106) were injected into the liver of each 2- to 6-day-old DKO or NRG mice, which had been previously irradiated at 300 rad. More than 95% of the humanized mice were stably reconstituted with human leukocytes in the blood (60–90% at 12–14 weeks). Each cohort had similar levels of engraftment. All mice were housed at the University of North Carolina at Chapel Hill.
An R5-tropic strain of HIV-1, JR-CSF, was used for chronic HIV-1 infection. All viruses were generated by transfection of 293 T cells (SIGMA-ALORICH, Cat#12022001-1VL) with pYK-JRCSF (NIH AIDS reagents program, Cat# 2708). Humanized mice with stable human leukocyte reconstitution were infected with JR-CSF at a dose of 10 ng p24/mouse, through intravenous (i.v.) injection. Humanized mice infected with 293 T mock supernatant were used in control groups (S2 Fig). Viral genomic RNA in plasma was extracted using the QIAamp Viral RNA Mini Kit (QIAGEN, Cat# 52904) according to the manufacturer’s instruction. HIV-1 replication (genome copies/ml plasma) was measured by real-time PCR (ABI Applied Biosystem) or by p24-FACS detection of productively infected human T cells.
A monoclonal antibody specific to blood dendritic cell antigen-2 (BDCA2), 15B, was used to treat humanized mice through intra-peritoneal (i.p.) injection (4 mg/kg) as previously reported [29]. Briefly, 15B was applied to mice at 7 weeks post-infection by injecting twice every week for another 4 weeks.
For chronic JR-CSF infection, mice were terminated at 12 week post-infection. On termination, total leukocytes were isolated from mouse lymphoid organs as previously described [29,34–36]. Lymphoid tissues, including peripheral blood (PB), peripheral lymph nodes (pLN), mesenteric lymph nodes (mLN), spleen and BM were harvested for analysis. Red blood cells were lysed with ACK buffer, and the remaining cells were stained and fixed with 1% (wt/vol) formaldehyde before FACS analysis. The total cell number was quantified by using Guava Easycytes with Guava Express software (Guava). Human BM cells were isolated by ficoll-hypaque density gradient centrifugation and collected for further analysis.
Surface and intracellular fluorochrome-conjugated antibodies or reagents from Biolegend, BD Bioscience, eBioscience and R&D Systems were used in this study. For humanized mice, live human leukocytes (Y7-mCD45-hCD45+) were analyzed for HPC subsets or phenotypic expression by using the CyAn FACS instrument (Dako). Live/dead fixable violet dead cell dye (LD7) was purchased from Molecular Probes (Eugene, OR). For intracellular p24 staining, freshly isolated cells were collected for surface staining, followed by cell permeabilization using a Cytofix/Cytoperm kit (BD Bioscience) and intracellular staining and washing. The data were analyzed using Summit Software.
5-Bromo-2’-deoxyuridine (BrdU, Cat#: B5002, Sigma-Aldrich, St. Louis, MO) was first dissolved in water at a concentration of 10 mg/mL for stock in -20°C. The BrdU stock was then diluted in 200 μL PBS and injected i.p. at 100 mg/kg body weight. Four hours later, the mice were terminated, and BM cells were collected. BrdU staining was performed according to the manufacturer’s instructions. In brief, cells were first stained for surface markers and then incubated with a working solution of the BrdU staining buffer for 15 minutes, followed by incubation with DNase I (BIO-RAD, Cat#: 7326828) for 1 hour at 37°C in the dark. Thereafter, the cells were stained with the FITC-conjugated anti-BrdU antibody for 30 minutes at room temperature in the dark and subsequently washed. The data were analyzed using Summit Software.
The EasySep human CD34+ selection kit (Cat#:18056, StemCell Tech, Canada) was used to isolate CD34+ cells from frozen BM cells. The purity of CD34+ cells was greater than 90%. The CD34+ cells were then counted and seeded in complete methylcellulose (Methocult H04034; Stem Cell Technologies) at a concentration of 500 cells/mL and plated in 35-mm grid plates, 1 mL/plate, in triplicate per mouse according to the manufacturer’s instructions. Colonies were counted 2 weeks later in a blinded fashion using a QImaging Micropublisher 3.3 CCD digital camera and QCapture software version 3.0 (QImaging, Surrey, BC).
BM cells were pooled by mouse groups for human CD45+ cells sorting. Cells were stained with human CD45, mouse CD45 and 7-Aminoactinomycin D (7-AAD). For human CD34+ HSC sorting, anti-lineage (anti-CD3, anti-CD14, anti-CD16, anti-CD19, anti-CD20 and anti-CD56) and anti-CD34 antibodies were added to the antibody mix. Cell sorting was performed by the UNC Flow Cytometry Core.
CD34+ cells were isolated from human fetal liver tissues. Then the cells were cultured in StemSpan SFEM medium (Stem Cell Technologies) with heparin (10 μg/ml, Sigma), recombinant human SCF (20 ng/ml, R&D), thrombopoietin (40 ng/ml, Cell Sciences) and CHIR99021 (GSK3 inhibitor, 250 nM, STEMGENT) for 48 hours in the presence of IFN-α or IFN-β at the dose of 20 IU/ml and 200 IU/ml, respectively. The cells were counted and collected for the detection of CD38 expression on HPCs.
RNA purification was carried out using the RNeasy Plus Mini Kit (Cat# 74134, QIAGEN, Venlo, Limburg, Netherlands) according to the manufacturer’s instructions. DNase (QIAGEN) treatment was added to the column to eliminate any potential DNA contamination during RNA preparations. Total RNA was checked for quantity, purity and integrity by capillary electrophoresis. RNA was amplified with Cy3- and Cy5-labeled CTP in separate reactions to produce differentially labeled samples and reference cDNAs. Total RNA (200 ng to 400 ng) was used as the starting material to prepare cDNA. Both samples were hybridized to the same microarray (UNC Genomic and Bioinformatics Core) using SurePrint G3 Human Gene Expression 8 ◊ 60K Microarray Kit (Agilent). Agilent Feature Extraction v18 software was used to analyze all images. Gene expression values were quantified by the log2 ratio of the red channel intensity (mean) vs. green channel intensity (mean), followed by LOWESS normalization to remove the intensity-dependent dye bias.
Data were analyzed using GraphPad Prism software version 5.0 (GraphPad software, San Diego, CA). Data from different cohorts of mice were compared using a 2-tailed unpaired T test. All results were considered significant for P values < 0.05.
|
10.1371/journal.pgen.1005955 | The MKK7 p.Glu116Lys Rare Variant Serves as a Predictor for Lung Cancer Risk and Prognosis in Chinese | Accumulated evidence indicates that rare variants exert a vital role on predisposition and progression of human diseases, which provides neoteric insights into disease etiology. In the current study, based on three independently retrospective studies of 5,016 lung cancer patients and 5,181 controls, we analyzed the associations between five rare polymorphisms (i.e., p.Glu116Lys, p.Asn118Ser, p.Arg138Cys, p.Ala195Thr and p.Leu259Phe) in MKK7 and lung cancer risk and prognosis. To decipher the precise mechanisms of MKK7 rare variants on lung cancer, a series of biological experiments was further performed. We found that the MKK7 p.Glu116Lys rare polymorphism was significantly associated with lung cancer risk, progression and prognosis. Compared with Glu/Glu common genotype, the 116Lys rare variants (Lys/Glu/+ Lys/Lys) presented an adverse effect on lung cancer susceptibility (odds ratio [OR] = 3.29, 95% confidence interval [CI] = 2.70–4.01). These rare variants strengthened patients’ clinical progression that patients with 116Lys variants had a significantly higher metastasis rate and advanced N, M stages at diagnosis. In addition, the patients with 116Lys variants also contributed to worse cancer prognosis than those carriers with Glu/Glu genotype (hazard ratio [HR] = 1.53, 95% CI = 1.32–1.78). Functional experiments further verified that the MKK7 p.116Lys variants altered the expression of several cancer-related genes and thus affected lung cancer cells proliferation, tumor growth and metastasis in vivo and in vitro. Taken together, our findings proposed that the MKK7 p.Glu116Lys rare polymorphism incurred a pernicious impact on lung cancer risk and prognosis through modulating expressions of a serial of cancer-related genes.
| Rare variants have been identified to be associated with a variety of human malignancies, which account for a considerable fraction of heredity for complex diseases. To date, however, the precise molecular mechanism of rare variants involved in tumors initiation and progression largely remains unclear. We tested the associations between rare variants in MKK7 and lung cancer risk and prognosis in two-stage retrospective studies with a total of 5,016 lung cancer patients and 5,181 controls in Chinese. We found that the rare variant from Glu to Lys in MKK7 p.116 locus exerted a detrimental effect on lung cancer risk, progression and prognosis. Further functional experiments demonstrated that lung cancer cells with p.116Lys variant accelerated the potentials of cell growth, proliferation, colony formation, migration and invasion than the cells with p.116Glu. This rare variant also promoted the xenograft growth and metastasis of nude mice in vivo through regulating a serial of cancer-related genes. Our data indicated that p.Glu116Lys rare variant in MKK7 might be a novel biomarker for lung cancer risk and prognosis.
| Ever-increasing epidemiological studies, especially the genome-wide association studies (GWAS), have extensively identified numerous genetic variants, including single-nucleotide polymorphisms (SNPs), to be associated with risk and progression of various human malignancies[1–3]. Despite these discoveries, much of the genetic contributions to complex diseases remains unclearly illuminated because of the fact that only a small proportion of cancer heritability can be explained by those common SNPs, typically with minor allele frequency (MAF) >5%, reflecting that some ‘missing heritability’ existed [4, 5]. Recently, accumulating evidence revealed that rare variants (MAF<1%) could decipher accessional disease risk or trait variability [6–8]. An example is that the rare variants located in proto-oncogenes or tumor suppressor genes may contribute to phenotypic variations through modifying their biological functions or genes expression, and thus play an important role in cancer initiations and progressions[9, 10]. These findings provide novel approaches for the exploration of cancer mechanism.
Human mitogen-activated protein kinase kinase 7(MKK7, also known as MAP2K7, MIM: 603014) belongs to the MAP kinase kinase family, and is identified as a tumor suppressor gene [11]. Evidence has demonstrated that MKK7 serves as a critical signal transducer involved in several cancer-related signaling pathways and genes, and thus participates in regulating cells proliferation, differentiation and apoptosis [12–14]. MKK7 deletion in mice caused distinct phenotypic abnormalities[15], whereas expression of MKK7 could inhibit lung cancer cells development[16]. In addition, several studies also indicated that MKK7 acts as a suppressor in tumors migration, invasion and metastasis [17–19].
Human MKK7 gene is located at chromosome 19p13.3-p13.2, a region spanning over a fragile site associated with various human diseases [20, 21]. A study reported that the somatic mutations and loss of heterozygosity at 19p13.2 commonly existed in lung cancer [22]. Furthermore, another study showed that several non-synonymous somatic mutations of the MKK7 gene also occurred and were associated with colorectal cancer predisposition [23]. Nevertheless, it is still molecularly unexplained how these rare variants implicated in cancer initiation and development. Therefore, in the current study, we test the hypothesis that the rare variants in MKK7 might be associated with lung cancer risk and prognosis by disturbing the biological functions of MKK7.
Based on three independent case-control studies, we genotyped five rare SNPs in MKK7 (i.e., rs28395770G>A: p.Glu116Lys, rs56316660A>G: p.Asn118Ser, rs56106612C>T: p.Arg138Cys, rs55800262G>A: p.Ala195Thr and rs1053566 C>T: p.Leu259Phe) and investigated their associations with lung cancer risk, metastasis and prognosis. The biological effects of those promising rare variants on lung cancer were further assessed by a series of functional experiments.
The demographic distributions of the three study populations are described in Table 1. Consistently, no significant deviations were observed in distributions of age, sex, drinking and family cancer history from the cases to controls in all the studied sets (P >0.05 for all), except for smoking status (P < 0.05). These variables were further adjusted in the multivariate logistic regression model to control possible confounding on the main effects of the studied polymorphisms. The histological types and clinical stages of the cases were also enumerated in Table 1. In addition, we recalculated the samples size based on population sources. There were 3005 cases and 3013 healthy controls in Guangzhou area, 2011 cases and 2168 cancer-free controls in Suzhou area.
Table 2 summarized the genotype distributions of the studied MKK7 rare SNPs and their associations with lung cancer risk. In the discovery set, we found a significant frequency deviation between the cases and controls (exact P = 4.12×10−12) in p.Glu116Lys rare polymorphism. Compared to individuals with 116Glu/Glu genotype, the carriers with Lys/Glu heterozygote harbored a 3.33-fold increased risk of lung cancer (odd ratio [OR] = 3.33, 95% confidence interval [CI] = 2.29–4.86), and carriers with Lys/Lys variant genotype exerted a much higher cancer risk (OR = 3.94, 95% CI = 1.09–14.3). When combined with variant genotypes, they (Lys/Glu+Lys/Lys) also contributed a pernicious impact on lung cancer risk (OR = 3.38, 95% CI = 2.35–4.85), conforming to the fitted genetic model with the smallest akaike information criterion (AIC = 4415.3). However, we did not receive any association between other rare SNPs and lung cancer risk.
We further confirmed the above associations in another two validation sets, and obtained consistent results. The p.116Lys variants genotypes (Lys/Glu+Lys/Lys) exerted a 3.52-fold increased risk of lung cancer (OR = 3.52, 95% CI = 2.54–4.89) in validation set I, and a 2.87-fold increased risk of lung cancer (OR = 2.87, 95% CI = 2.04–4.04) in validation set II. Because the homogeneity test showed that the association in the above three sets was homogeneous (P = 0.711), we then merged the three populations to increase the study power, and found that the compared with Glu/Glu common genotype, the carrier with Lys/Glu or Lys/Lys had a remarkably adverse effects on lung cancer risk (OR = 3.23, 95% CI = 2.62–3.98; OR = 3.75, 95% CI = 2.09–6.71; respectively). Similarly, the Lys (Lys/Glu+Lys/Lys) variants also had a 3.29-fold increased risk of lung cancer under the dominant model (OR = 3.29, 95% CI = 2.70–4.01). The heritability test indicated that the p.Glu116Lys rare variant could explain about 2.16% of lung cancer heritability.
In stratification analysis, as is presented in Table 3, no deviation of p.116Lys variants on cancer risk was observed in most subgroups except for the strata of clinical stages (P = 0.016). We further evaluated the relationships between MKK7 p.Glu116Lys and lung cancer progression, and found that p.Glu116Lys was significantly associated with pejorative clinical stages (P <0.001, shown in S1 Table). As is revealed in S1 Table, the patients with 116Lys variants had increased probability of progressing to IV stage (OR = 1.69, 95% CI = 1.28–2.24). Likewise, the frequency of p.116Lys adverse genotypes elevated continuously along with the risk of lymphatic metastasis extent at diagnosis (5.8% for 0, 8.5% for 1, 8.7% for 2, and 10. 8% for 3), and with the distal metastasis extent at diagnosis (7.0% for 0, 9.9% for 1). In brief, patients with 116Lys variants were more likely to have metastasis (either nodal or distal metastasis) than those with Glu/Glu genotype (OR = 1.84, 95% CI = 1.34–2.53).
We further evaluated the associations between the combined types of those selected rare SNPs and lung cancer risk. As is shown in S2 Table, the individuals with only p.Glu116Lys variant was associated with lung cancer susceptibility (exact P = 1.18×10−33), accompanying by a 3.24-fold increased cancer risk (OR = 3.24, 95% CI = 2.64–3.97), which was best fitted for the heredity model (AIC value = 13840.2). It achieved 100% study power and yielded a value of 0.000 with a 0.001 prior probability lower than the preset FPRP-level criterion 0.20, suggesting that this finding is noteworthy. Individuals with a combination of p.Glu116Lys and p.Asn118Ser variant genotypes also had an increased risk of lung cancer (OR = 3.16, 95% CI = 1.02–9.76), but it achieved only 66.7% moderate power and a 0.985 FPRP value at a 0.001 prior probability, which is higher than the preset criterion 0.20. Furthermore, we also used the SKAT method to test combined genotypes associated with lung cancer risk, and found that only those combinations containing the p.Glu116Lys rare variation had prominent relevancies with lung cancer risk(P <0.01 for all). All these results indicated that among all of the MKK7 five rare polymorphisms, the p.Glu116Lys contributed the main effect on lung cancer risk. A serial of experiments was further conducted to decipher the biological mechanisms of p.Glu116Lys on lung cancer.
The distributions of demographic and clinical characteristics in the three datasets are presented in S3 Table. The Log-rank test and univariate Cox analysis revealed that patients with characteristics including ≥60, smoking or advanced stage had a significantly shorter median survival time (MST) and an increased death risk (P <0.05 for all). In contrast, the female patients, and those patients suffering from surgical operations, chemotherapy or radiotherapy prolonged survival time and had a more benignant prognosis (shown in S3 Table).
The relevancies between the MKK7 rare SNPs and lung cancer outcomes are shown in Table 4. In the discovery set, compared with Glu/Glu genotype, the patients with p.116Glu/Lys heterozygote had a significantly shorter MST (7 months vs. 13 months; Log-rank test P = 6.19×10−5) and a higher death risk (hazard ratio [HR] = 1.69, 95% CI = 1.31–2.19). Multivariate proportional hazards regression analysis indicated that this rare variant appeared an undesirable survival of lung cancer under the additive genetic model (HR = 1.63, 95% CI = 1.31–2.03). Congruously, the 116Lys (Lys/Glu+Lys/Lys) variants exerted a poor prognosis (HR = 1.73, 95% CI = 1.35–2.21) and a shorter MST (7 months vs. 13 months; Log-rank test P = 9.61×10−5; Fig 1A), while compared to the Glu/Glu wild-genotype. However, for other rare SNPs, no significant associations with lung cancer survival were found.
The associations between MKK7 rare SNPs and prognosis of lung cancer were further verified in other two validation sets. In those two datasets, when compared with the Glu/Glu genotype, patients with Lys/Glu genotype had a decreased MST (validation set I: 9 months vs. 15 months, Log-rank test P = 0.033; validation set II: 12 months vs. 16 months, Log-rank test P = 0.031) and had shown an increased death risk (validation set I: HR = 1.43, 95% CI = 1.10–1.85; validation set II: HR = 1.41, 95% CI = 1.04–1.92); those patients carrying Lys/Lys homozygote also exerted a pernicious cancer prognosis (validation set I: HR = 1.99, 95% CI = 1.09–3.64; validation set II: HR = 2.34, 95% CI = 1.15–4.75), along with a shorter MST (validation set I: 7 months vs. 15 months, Log-rank test P = 0.048; validation set II: 9 months vs. 16 months, Log-rank test P = 0.026). Similarly, the patients with p.116Lys (Lys/Lys+Lys/Glu) variants presented a shorter MST (validation set I: 9 months vs. 15 months, Log-rank test P = 0.013, Fig 1B; validation set II: 12 months vs. 16 months, Log-rank test P = 0.006, Fig 1C) and worse survival outcomes (validation set I: HR = 1.49, 95% CI = 1.17–1.90; validation set II: HR = 1.50, 95% CI = 1.13–2.00).
Pooled analysis of the three cohorts indicated that patients with Lys/Glu or Lys/Lys variant genotype harbored reduced 4 and 7 MST months (P = 2.61×10−7), coupling with a 149% (HR = 1.49, 95% CI = 1.27–1.74) and a 194% (HR = 1.94, 95% CI = 1.31–2.89) cancer death risk, respectively, while compared to patients with Glu/Glu genotype. Also, the p.116Lys (Lys/Lys+Lys/Glu) detrimental genotypes conferred a 5-months decreased in MST compared with that of Glu/Glu genotype (9 months vs. 14 months, Log-rank test P = 1.03×10−6) and had a 53% higher death risk (HR = 1.53, 95% CI = 1.32–1.78).
As is revealed in Table 5, although the strength of relevance represented by the HR values between the p.116Lys variants and lung cancer prognosis were different across a plurality of stratums, the homogeneity test showed that the difference was only significant in subgroups of clinical stage and distant metastasis (P values equal to 0.019 and 0.010, respectively). The unfavorable influence of the p.116Lys variants on cancer prognosis was more conspicuous in advanced stages (HR = 1.88, 95% CI = 1.53–2.30). The patients in the distant metastasis stage had 62% higher death risk than those without metastasis (HR: 1.88 vs. 1.26). We also found a remarkable modification effect between the clinical stage and the p.116Lys variants on lung cancer prognosis (P = 0.039).
We further analyzed the associations between the combinational genotypes of MKK7 rare SNPs and lung cancer prognosis. As is presented in S4 Table, combined-type of only p. Glu116Lys was significantly associated with cancer outcome. Patients only with Lys variants genotypes showed a shorter MST (9 months vs. 14 months, Log-rank test P = 4.01×10−8) and a higher death risk (HR = 1.58, 95% CI = 1.35–1.85) when compared with patients without those genotypes. This noticeable result achieved 100% study power and yielded a value of 0.000 at a 0.001 prior probability, which is lower than the preset FPRP-level criterion 0.20. Although the combination of the p.Glu116Lys and p.Asn118Ser was significantly associated with survival time using Log-rank test (P = 0.025), but it obliterated the relevance in the Cox regression analysis (HR = 1.74, 95% CI = 0.99–3.08) and obtained a FPRP value of 0.991 at a 0.001 prior probability higher than the preset criterion 0.20, suggesting that this result was likely to be untrustworthy.
To explore the effects of the MKK7 p.Glu116Lys rare variant on cell biological behaviors, multitudinously functional experiments were further executed. The proliferation test showed that cells with over-expressing MKK7-116Lys displayed a higher proliferation potential than cells with over-expressing MKK7-116Glu (Fig 2A, ANOVA test P<0.001). Cells highly expressing MKK7-116Lys also had strikingly promoted abilities of colony formation in common plate, as well as in soft-agar, compared to the cells with MKK7-116Glu (Fig 2B and 2C). In addition, we further performed flow cytometry to evaluate the influence of p.Glu116Lys variants on cells cycle and apoptosis. We found that the over-expressing MKK7-116Lys in A549 cells induced a significantly reduction in the G0/G1 phase (12.9% decreased, P = 0.015) and a corresponding increase in the G2/M phase (7.4% increased, P = 0.038), while compared with the cells stably expressing MKK7-116Glu (Fig 2D). Notably, the A549 cells with MKK7-116Lys also had decreased apoptosis rate than cells with MKK7-116Glu (Fig 2E, P = 0.043). Furthermore, cells with highly expression of MKK7-116Lys showed remarkably promoted migration and invasion capabilities in comparison to cells with over-expressing MKK7-116Glu (Fig 2F and 2G). These arresting results also occurred in the L78 cells with stably over-expressing MKK7-116Lys. All these findings suggested that the MKK7-116Lys variant had a detrimental impact on promoting cell proliferation, invasion and immigration.
To further determine the effect of p.Glu116Lys on tumor growth and metastasis in vivo, cells with stably over-expressing MKK7-116Glu or MKK7-116Lys were injected into nude mice subcutaneously (both for A549 and L78 cell lines), and intravenously (for A549 cell line only), respectively. As is shown in Fig 3A, the injection of MKK7-116Lys cells resulted in tumor formation began 4 days earlier compared to the results from injection of MKK7-116Glu cells. The tumor grew faster, and after 4 weeks, the tumor size in the former group was larger than the latter group (For A549: 1246.3±102.3 mm3 vs. 846.3±78.5 mm3, P <0.001, Fig 3A; for L78: 1474.5±99.4 mm3 vs. 921.1±88.4 mm3, P <0.001, Fig 3B). Moreover, we used the MRI and histology examination to determine whether the MKK7 p.Glu116Lys could cause tumor metastases, and found that all the mice injected with A549 cells over-expressing MKK7-116Lys suffered from pulmonary metastasis, while the mice group injected with MKK7-116Glu A549 cells did not (Fig 3C, 3D and 3E). These findings demonstrated that MKK7-116Lys variant enhanced lung tumor growth and metastasis in vivo.
To decipher the potential mechanisms behind the MKK7 with p.Glu116Lys rare variant induced lung cancer risk and progression, we further performed DGE sequencing to compare gene profiles between A549-MKK7-116Glu cells and A549-MKK7-116Lys cells. We found that compared to cells with stably over-expressing MKK7-116Glu, cells with MKK7-116Lys had 192 genes differentially expression with a q value of <0.001 (S1 File). Among these genes, 128 genes were up-expressed, and 64 genes were down-expressed. We further validated the DGE results using qRT-PCR assay, with detecting differently expressed genes including 5 up-expression genes(STC2, SLC1A3, MSMO1, BCL10 and HMGCR) and 5 down-expression genes(SAA1, SBK2, CDH5, COL4A2 and BCL9L). The results were in concordance with those findings through DGE sequencing. Furthermore, we conducted Gene Ontology (GO) analysis using these differentially expressed genes. The GO results indicated that these 192 differentially expression genes were annotated to be associated with cell cycle process, cell proliferation, apoptosis, tissue development, tumor invasion, and metastasis et al. Above results suggests that alteration from 116Glu to Lys in MKK7 might influence its downstream targets expression and thus facilitate lung cancer initiation and development.
In the current study conducted among southern and eastern Chinese with a total of 5,016 lung cancer patients and 5,181 controls, we estimated the relationships between rare variants in MKK7 gene and lung cancer risk and prognosis, and found that the p.Glu116Lys rare variant was significantly associated with an increased lung cancer risk, progression and prognosis. The individuals with 116Lys variants had promotional cancer risk and higher probability of metastasis at diagnosis. The harmful role of the 116Lys variants also resulted in a poorer lung cancer prognosis in the patients than in the patients with Glu/Glu genotype. Further functional assays demonstrated that lung cancer cells with p.116Lys variant accelerated cell growth, proliferation, colony formation, migration and invasion. They also promoted the xenograft growth and metastasis of nude mice in vivo through regulating a serial of cancer-related genes. However, no conspicuous evidence was obtained to prove any significant association between other rare SNPs and lung cancer risk and prognosis. To the best of our knowledge, this is the first study to investigate the associations between the genetic rare variants in MKK7 and lung cancer risk, as well as metastasis and prognosis.
Accumulating evidence indicated that ‘missing heritability’ in complex human diseases caused increasing attention over the past a few years because the findings from the GWAS and other epidemiological studies did not completely explained the genetic heritability [4, 5, 24]. With rapid advances in high-throughput sequencing technologies, unnoticed genetic components such as low-frequency (1%≤MAF< 5%) and rare genetic variants (MAF< 1%) are being thoroughly assessed and investigated for their associations with complex human diseases [6, 7, 25]. These approaches highlight an unparalleled opportunity to decipher unexplained genetic contributions in forming complex traits [26, 27], especially in human malignancies [10].
MKK7 has been identified to be a tumor suppressor gene constitutive activation of JNK signaling pathway to induce cell apoptosis [28, 29]. Recently, a study reported that tissue-specific inactivation of the stress signaling kinase MKK7 in ras-driven lung carcinomas and NeuT-driven mammary tumors markedly accelerates tumor onset and reduces overall survival through directly coupling oncogenic and genotoxic stress to the p53 stability[11]. Lin HJ et al. identified that MKK7 could negatively regulate the expressions of MMP-2 and MMP-9 and thus inhibited cancer cell migration and invasion [17]. In addition, another report showed that ectopic expression of the MKK7 suppresses the formation of overt metastases by inhibiting the ability of disseminated cells to colonize the lung[14]. Furthermore, several studies display an intimate linkage with germline mutations in MKK7 and cancer onset and progression [30–32]. In the present study, we found that p.Glu116Lys rare variant in MKK7 contributed a pernicious impact on lung cancer risk and prognosis. We also observed a remarkable interaction between clinical stage and the rare variant on cancer survival. The p.Glu116Lys variant located at kinase activity domain of the MKK7 gene, which might influence the structure and functions of the MKK7 based on the bioinformatics analysis (http://snpinfo.niehs.nih.gov/). Our biological assays demonstrated that the 116 locus alteration from Glu to Lys in MKK7 could promote cells proliferation, migration, invasion, and reduce cells apoptosis in vitro; the adverse role of 116Lys variant was also found to facilitate the xenograft growth and metastasis in vivo. The 116Lys variant further altered the expression of downstream genes modulated by MKK7 as the DGE results showed, which might be closely related with lung cancer initiation and development. Among these differentially expressed genes, there were ones annotated as cell-regulated genes, cell apoptosis, cancer-related genes, tumor invasion and metastasis. For example, as the DGE results had indicated, STC2, YEATS4 and SLC1A3 genes were up-expressed in the cells with stably over-expressing MKK7-116Lys compared with the cells with MKK7-116Glu. A study had reported higher mRNA and protein expressions of STC2 in lung cancer tissues compared to the adjacent normal tissue. Knockdown of STC2 slowed down lung cancer cell growth progression, colony formation and metastasis [33]. Another article showed the proof that overexpression of YEATS4 abrogated senescence in human bronchial epithelial cells, while RNAi-mediated attenuation of YEATS4 could conversely reduce lung cancer cells proliferation and tumor growth, impair colony formation, and induce cellular senescence[34]. In addition, several genes such as CDH5 and UBA7 (also known as UBE1L) were significantly down-regulated in the cells with MKK7-116Lys. A previous study had reported a downregulation of CDH5 in Bulgarian patients with early-stage non-small cell lung cancer [35]. Loss of UBE1L is a common event in lung carcinogenesis, and the UBE1L gene suppressed lung cancer growth by preferentially inhibiting cyclin D1 [36, 37]. All the above public evidence was in accordance with our findings in the current study, which convincingly supported our ultimatums that MKK7 p.Glu116Lys rare variant exerted adverse effects on lung cancer risk, progression and prognosis by modulating a number of cancer-related genes.
Our study has several strengths and limitations. Based on three independent case-control studies, we have obtained consistent results of the association between the MKK7 p.Glu116Lys rare variant and lung cancer risk and prognosis, with a compellingly strong study power of 100% (two-sided test, α = 0.05) to detect an OR of 3.29 for the 116Lys variant genotypes (which occurred at a frequency of 2.7% in the controls), and with a 100% statistical power for HR with a value to 1.53, while compared with the 116Glu wild-genotype. A serial of functional experiments further sustained the results that the p.116Lys variants conferred noxious effects on lung cancer risk, progression and prognosis. However, there are also some limitations. The selection bias is unavoidable on account of the hospital-based retrospective studies. Also, with restriction to a Chinese Han population, it is uncertain whether our findings could be generalized to other populations. Furthermore, due to the technological limitations, we did not promulgate any direct target genes of MKK7 with respect to the p.Glu116Lys rare polymorphism, which might help us to understand the precisely molecular mechanism of this rare SNP on influencing cancer risk and progression.
In conclusion, our findings indicated that the p.Glu116Lys rare variant of MKK7 was associated with an increased lung cancer risk and worsened prognosis in Chinese, which was likely to be related to modulation of a serial of cancer-related genes. These results suggested that the MKK7 p.Glu116Lys may be a useful predictive biomarker for lung cancer susceptibility and prognosis. Validations through larger population-based studies in different ethnic groups, and functional assay to reveal target gene of the p.Glu116Lys rare SNP in MKK7 are warranted.
Each participant was scheduled for an interview to collect individual information on smoking status, alcohol use, and other selected factors, and to obtain a donated 5 mL of peripheral venous blood under his or her informed consent. The study was approved by the institutional review boards of Guangzhou Medical University (Ethics Committee of Guangzhou Medical University: GZMC2007-07-0676) and Soochow University (Ethics Committee of Soochow University: SZUM2008031233). All experiments and procedures involving animals were conducted in accordance with guidelines approved by the Laboratory Animal Center of Guangzhou Medical University.
In this study, three two-stage independently retrospective studies with a total of 5,015 lung cancer patients and 5,181 healthy controls were performed in southern and eastern Chinese populations. In brief, 1,559 lung cancer cases and 1,679 cancer-free controls as the discovery set, which included southern Chinese with 1,056 primary lung cancer cases and 1,056 healthy subjects recruited from the Guangzhou city, and eastern Chinese with 503 patients and 623 controls enrolled from Suzhou city, have been previously described[1, 38, 39]. In the validation set I, 1,949 lung cancer patients that were continuously recruited from Guangzhou between April 2009 and June 2014 with a 90% response rate and 1,957 sex and age (± 5 years) frequency matched cancer-free controls who were randomly selected from about 3,000 individuals participating in health community programs with an 83% response rate were used. Moreover, the other population from Suzhou city was used as validation set II, in which 1,508 lung cancer cases were enrolled between December 2009 and March 2014 with an 85% response rate and 1,545 controls were randomly selected from 8000 participators in the annual healthy checkup programs with a response rate of 91%. All the participants were genetically unrelated ethnic Han Chinese and none had blood transfusion in the last six months. Definitions of smoking status, pack-years smoked, drink status, family history of cancer and family history of lung cancer have been previously described[38, 39].
As was previously reported, clinical information and characteristics of patients were also collected [40]. Patient follow-ups were performed through telephone calls every three months from time of enrollment to the last scheduled follow-up or death. Survival time was calculated starting from the day the patients first received confirmed diagnoses to the date of the last follow-up or death, and dates of death were acquired from medical records or information provided by family members through telephone follow-ups. Patients that were lost to follow-ups or had no accurate data on clinical information were excluded. In the finalized study, 908 patients from the discovery set, 1027 patients from validation set I and 971 patients from validation set II that have completed the follow-up and had intact survival data were included in this study. In addition, to eliminate the bias in patient selection, we analyzed the differences in clinical features, as well as in survival data, between the included and excluded groups, and no deviated results were observed.
Because no published data reveal potentially functional variants in MKK7, we only selected those exon variants in gene coding region causing amino acid change that are supposed to be with most functional potential. Through the strategy of searching for the rare polymorphisms located in the MKK7 gene exons region based on the public dbSNP database (http://www.ncbi.nlm.nih.gov/snp/, access to 1/1/2014), we found that 5 SNPs of MKK7 gene (i.e., rs28395770G>A: p.Glu116Lys, rs56316660A>G: p.Asn118Ser, rs56106612C>T: p.Arg138Cys, rs55800262G>A: p.Ala195Thr and rs1053566C>T: p.Leu259Phe) were rare with MAF<1% in Chinese population. We then re-sequenced the whole cDNA of MKK7 in 100 normal Chinese Hans randomly picked from the controls, and no newfound rare variants outside of those 5 SNPs were obtained. Therefore, we chose these above rare SNPs in the current study.
Genomic DNA was extracted from 2 mL peripheral blood using the routine method. Genotypes of all the selected SNPs were determined by direct DNA sequencing. A fragment of a total of 1,102 bp from the whole genomic DNA templates with the forward primer 5′-CCCAGCATTGAGATTGACCAGA-3′ and reverse primer 5′- TGCCATGTAGGCGGCACA-3′, which comprises the 5 studied SNPs was amplified. The PCR program for the amplification was as follows: 95°C for 5 minutes and then 40 cycles of denaturation (95°C for 45 seconds), annealing (61°C for 1 minute), and extension (72°C for 1 minute and 30 seconds), and a final polymerization step at 72°C for 7 minutes. The products were then separated by a 1% agarose gel and extracted. Finally, the PCR products were sequenced by an automated sequencing system (ABI Prism 3730 Genetic Analyzer; Applied Biosystems, Foster City, USA) operating according to the manufactures’ protocols (S1 Fig).
The cDNA sequence of human MKK7 gene with a wild-type (p.116Glu) was synthesized by the Sangon Biotech Company (Shanghai, China) and cloned into pLVX-IRES-neo expression vector (Clontech Laboratories Inc., San Francisco, CA, USA). The mutated pLV-MKK7-116Lys plasmid was induced by site-directed mutagenesis using the Quick Change XL site-directed mutagenesis kit (Stratagene, La Jolla, CA, USA). The resulting constructs were verified by direct sequencing. The lentiviral production and transduction were performed abiding by protocol described elsewhere [40]. In brief, replication-defective VSV-G pseudotyped viral particles were packaged using a 3-plasmid transient cotransfection method (Lenti-T HT packaging mix, Clonetech, San Francisco, CA, USA). Viruses were then harvested and concentrated. For transfection, two human lung cancer cell lines, A549 (a human lung adenocarcinoma cell line) and L78 (a human lung squamous carcinoma cell line) were infected with control lentivirus (an “empty” vector without the MKK7 fragment inserted), pLV-MKK7-116Glu lentivirus and the pLV-MKK7-116Lys lentivirus, respectively. The cells were stably selected with G418 at 100 μg/ml (Gibco, Lyon, France), and the drug-resistant cells were confirmed by qRT-PCR and western blotting assays (S2 Fig).
Cells infected with different allele lentivirus (pLV-MKK7-116Glu and pLV-MKK7-116 Lys) were seeded into 96-well flat-bottomed plates. 1,000 cells per 100 μl of cell suspension were used to add in each well. After a certain time of cultivation, cell viability was measured by MTT assay as is previously described [40]. In brief, 20 μl MTT solutions (5 mg/mL, Sigma, USA) per well were added for 4 h before the end of the experiment. After that, the supernatant fluid was removed and 150 μl of DMSO was added to each well. The absorbance was then measured at 490 nm wavelength using a Plate Reader (Bio-Tec Instruments, Inc.) after shaking the plate for 15 min at room temperature.
For cell cycle analysis, cells with stably expressing MKK7-116Glu or MKK7-116Lys were collected, washed with PBS and fixed by 70% ethanol for at least 1 h. Subsequently, the cells were stained with 0.5 mL propidium iodide (PI) staining solution, and cellular DNA content was analyzed using a flow cytometry (BD Biosciences, CA, USA). For cell apoptosis, an annexin v-fluoresce-in isothiocyanate (V-FITC)/PI double staining assay was conducted according to the manufacturer instructions. In brief, the cells were harvested and stained with annexin V-FITC and PI for 20 min at room temperature in the dark. The cells were then washed twice with PBS, and the fluorescence of the cells was measured by flow cytometry.
Cells with stably over-expressing MKK7-116Glu or MKK7-116Lys were seeded into a 6-well plate (100 cell/well) with RPMI 1640 medium supplemented with 10% fetal bovine serum (FBS), and allowed to grow until visible colonies formed (approximately 2 weeks). After washing with PBS, the cell colonies were fixed with 4% paraformaldehyde and stained with crystal violet (Invitrogen) for 30 min, then washed, air dried, photographed and counted. Furthermore, colony formation assay in soft-agar was also executed to detect the effect of MKK7 Glu116Lys rare variant on cell malignant transformation. The detailed procedures were previously described [40]. Briefly, cells suspended with DMEM medium containing a concentration of 0.35% soft agar were poured onto 6-cm tissue culture dishes coated with 5 ml of 0.75% bottom agar. At the end of the experiment, the colonies were then stained, photographed and counted.
Cell migration and invasion abilities were appraised by Corning transwell insert chambers (8-uM pore size; Costar, USA) and BD BioCoat Matrigel Invasion Chamber (Becton Dickinson Biosciences, USA), respectively. 2×104 (migration assay) or 2×105 (invasion assay) transfected cells in 200μl serum-free RPMI 1640 medium were seeded in the upper chamber, and 800 μl medium with 10% FBS were added to the lower compartment. After 24 h for migration assays or 48h for invasion assays at 37°C in a 5% CO2 humidified atmosphere, cells in the upper chamber were carefully scraped off using a cotton swab, and the cells that had migrated to or invaded the lower surfaces of the membrane were fixed with 4% paraformaldehyde solution and stained with crystal violet (Invitrogen), imaged and counted. Assays were independently conducted for three times.
Female BALB/c nude mice that were 4–5 weeks of age were purchased from the Laboratory Animal Center of Guangdong province (Guangzhou, China). Cells with MKK7-116Glu or MKK7-116Lys were diluted to a concentration of 5×107/ml in physiological saline. 0.1 ml of the cells suspension was injected subcutaneously into the dorsal flank of mice to construct tumor growth model (both for A549 and L78 cell lines), or injected intravenously into the caudal vein of mice to construct tumor metastasis model (for A549 cell line only). Six nude mice were used for each group. When a tumor was palpable in the growth model, tumor size was measured every other day using a caliper along two perpendicular axes and calculated according to the following formula: Volume = 1/2×length×width2. The tumor metastases were evaluated by magnetic resonance imaging (MRI) and histology examination.
MRI was performed proximately 10 weeks post-injection using Philips Gyroscan Intera 1.5T ultraconducted MRI scanner (Netherlands) and incorporating a removable gradient coil insert. The details of MRI imaging were conducted as suggested by the public literature [41]. In brief, mice were placed prone on an MR-compatible sled within a carrier tube and positioned in the magnet. Induction and maintenance of anesthesia during imaging was achieved through inhalation of 10% chloral hydrate. MRI examination of coronal T2-weighted (T2WI) scanning was conducted with the following variables: Repetition time (TR) = 4000ms, echo time (TE) = 111ms, field-of-view (FOV) = 3 cm, number of slices = 20, slice thickness = 1.0 mm, matrix = 256×256. Following image acquisition, raw image sets were transferred to a processing workstation and processed using the medical imaging software. Tumor metastatic burden were calculated from manually traced regions-of-interest (ROI).
The animals were euthanized and their tumor masses were harvested and fixed with 10% neutral formalin solution, embedded in paraffin, and sectioned at 5 μm. The sections were then stained with hematoxylin-eosin (HE) staining and examined by light microscopy at 20× magnification.
The total RNAs from different A549 transfectant cells were extracted using the TRIzol reagent (Invitrogen) in accordance with the manufacturer instructions. RNA quantity and quality were assessed using a NanoDrop 2000 spectrophotometer (Thermo Scientific, MA, USA). The gene expression profiling both in A549-MKK7-116Glu cells and A549-MKK7-116Lys cells were conducted using Illumina NlaIII digital gene expression (DGE) sequencing. Analyses were performed according to the manufacturer recommendations [42]. Briefly, DGE sequence libraries were sequenced using Illumina HiSeq 2000 platform. Differentially expressed genes between the two groups of cells were identified using the reads per kilobase of transcript per million mapped reads (RPKM) method. The q value ≤ 0.001 and the absolute value of log2 ratio ≥ 1 were as the threshold to judge the significance of gene expression differences.
On the basis of genes profiling sequencing results, the expression levels of 10 selected genes (includes 5 up-expression genes and 5 down-expression genes) in the A549-MKK7-116Glu cells and A549-MKK7-116Lys cells were verified by the quantitative real time PCR (qRT-PCR) assay described elsewhere [39]. The relative levels of RNA were detected using the ABI Prism 7900HT sequence detection system (Applied Biosystems) and with the SYBRPremix Ex Taq (Perfect Real Time, TaKaRa, China) and β-actin as the internal reference. Each assay was performed in triplicate and independently repeated three times. All the primers used for PCR amplification are listed in S5 Table.
The chi-square test was used to assess differences in the distributions of demographic characteristics between cases and controls. The distributions of genotypes between cases and controls were analyzed with Fisher’s exact test. Unconditional logistic regression model with or without adjustment for surrounding factors was used to evaluate the associations between the MKK7 rare SNPs and lung cancer risk and metastasis. The correlations between MKK7 rare genotypes and lung cancer clinical features were tested using Spearman rank correlation. The sequence kernel association test (SKAT) was used to estimate the combined effect of multiple variants in MKK7 and lung cancer risk using R software (version 3.0.2; The R Foundation for Statistical Computing) with the SKAT package[43]. The REML model was used to assess the heritability explained by the genetic variants [44]. Breslow-Day test was used to test the homogeneity between the subgroups. The statistical power was calculated using the PS Software. The false-positive report probability (FPRP) test was applied to detect false-positive association findings [45]. The associations between clinical variables, as well as genotypes, and overall survival time were estimated using the Kaplan-Meier method and Log-rank test. The Cox proportional hazards regression model with or without adjustment for confounders was used to evaluate the effect of rare polymorphisms on lung cancer prognosis. Multiplicative interactions were assessed by logistic regression or Cox regression [38]. The differences in gene expression, colonies number levels, and cells’ ability to invade and migrate were analyzed using the student’s t-test. Repeated measure ANOVA test was performed to analyze the deviation of cell proliferation and tumor growth in different groups. All tests were two-sided using the SAS software (version 9. 3; SAS Institute) and P <0.05 was considered statistically significant.
|
10.1371/journal.pcbi.1002301 | Coherent Conformational Degrees of Freedom as a Structural Basis for Allosteric Communication | Conformational changes in allosteric regulation can to a large extent be described as motion along one or a few coherent degrees of freedom. The states involved are inherent to the protein, in the sense that they are visited by the protein also in the absence of effector ligands. Previously, we developed the measure binding leverage to find sites where ligand binding can shift the conformational equilibrium of a protein. Binding leverage is calculated for a set of motion vectors representing independent conformational degrees of freedom. In this paper, to analyze allosteric communication between binding sites, we introduce the concept of leverage coupling, based on the assumption that only pairs of sites that couple to the same conformational degrees of freedom can be allosterically connected. We demonstrate how leverage coupling can be used to analyze allosteric communication in a range of enzymes (regulated by both ligand binding and post-translational modifications) and huge molecular machines such as chaperones. Leverage coupling can be calculated for any protein structure to analyze both biological and latent catalytic and regulatory sites.
| What are the molecular mechanisms of allosteric communication in proteins? We base our analysis on the hypothesis that a folded protein has a number of conformational degrees of freedom, which describe fluctuations around the native conformation and switching from/to functional states. Transitions between the protein states involved in function and its regulation are based on coherent conformational degrees of freedom. Motion of one part of a protein along such a degree of freedom, implies a correlated motion in other parts of the protein. By determining which binding sites are simultaneously affected by the same motion we find sites that are allosterically coupled, i.e. where binding at one site can cause a change in ligand-affinity at another. Leverage coupling, the quantity introduced to measure this type of connection, reflects allosteric communication between different binding sites. We show how it can be used to understand allostery in enzymes of different sizes as well as in large protein complexes such as chaperones. Analysis of leverage coupling provides guidance in targeting native and latent regulatory sites.
| The concept of allostery was originally formulated to describe cooperative ligand binding in oligomeric proteins. The first model of positive cooperativity in binding of oxygen to hemoglobin was proposed by Linus Pauling in 1935 [1], but the term allostery was coined in connection with the phenomenological MWC (Monod-Wyman-Changeux) and KNF (Koshland-Némethy-Filmer) models, developed in the 1960s [2], [3], [4]. Since then, there have been numerous studies of the mechanisms of allosteric regulation [5], [6], applying different experimental [7], [8] and computational approaches [9] to proteins as different as small single-domain enzymes, motor proteins [10] and chaperones [11], [12]. Although much progress has been made, the dichotomy between the original MWC and KNF models, or their modern counter parts, conformational selection and induced fit, dominates the discussion of allostery to this day [6]. The two models do however not describe mutually exclusive scenarios [13], [14], [15]: in both cases there is a shift in the population of different functional states upon effector binding. The main difference between the two is whether binding precedes conformational change or not [14]. Transition pathway analysis is primarily a matter of kinetics, whereas the shift in conformational equilibrium is one of thermodynamics: the conformational states involved determine which binding sites are allosterically connected, and their relative stability before and after binding determines the effect of regulation [6]. The major task therefore is to use this understanding to find structural determinants and molecular mechanisms of allosteric communication between distant binding sites [16].
Recently we developed the concept of binding leverage to measure the ability of a generic ligand, binding at different sites, to couple to conformational transitions, and thus its potential to have an allosteric effect [15]. We showed that in the majority of the studied cases, known allosteric and active sites had high binding leverage. We treated each site individually under the assumption that a site that has high binding leverage is connected to the global dynamics of the protein, without any specification of what other sites could be connected. Here we move on to investigate how allosteric communication takes place between specific pairs of sites. We introduce the concept of leverage coupling, which provides a quantitative characteristic of allosteric communication. We will also demonstrate how binding leverage and leverage coupling can be used to analyze allosteric communication mediated by metal binding and phosphorylation, as well as the function of three chaperones (GroEL-GroES, CCT and thermosome).
In this paper we will develop a molecular model of allosteric communication based on the concept of binding leverage (described in Methods). We recently showed that binding leverage could identify key binding sites, and also potentially latent allosteric sites, in a wide range of proteins [15]. Here we investigate how specific pairs of sites are allosterically connected via leverage coupling.
To study site-site communication, we make the following assumption: sites that have high binding leverage for the same motion are more likely to be allosterically coupled than sites that only have high binding leverage for motion along independent degrees of freedom.
To represent a set of independent degrees of freedom we will use low frequency normal modes, which describe coherent motion involving the whole protein, and thus allow communication across large distances. We do not propose that protein dynamics is best described by global harmonic motion, but recognize the fact that the modes have repeatedly been shown to describe functional conformational change for proteins [17], [18]. We therefore use them as a set of basis vectors describing the allowed directions of motion around the folded state of a protein, and explore the possibility that movement along a given mode can have an independent functional relevance.
We have illustrated the role of independent degrees of freedom in allostery for a toy protein in Figure 1. This protein has four binding sites W, X, Y and Z, and we have included two normal modes in the illustration, indicated by red and green arrows. The green mode causes closing of site Z and opening of site X, and only slight deformations of the other two sites. The red mode causes opening of site X and closing of site Y. Small red and green arrows indicate the deformation at each site for either mode. Site X and Y both have high binding leverage under the red mode and sites X and Z have high binding leverage under the green mode. This means that the pairs X and Y and Z and X are allosterically coupled, whereas the other pairs of sites are only weakly coupled (indicated by the thickness of the lines crossing the protein, connecting the corresponding sites). In practice, X could be a catalytic site, Z an activator site and Y an inhibitor site. There is only indirect competition between the effects of Z and Y, i.e. if an activator is present at Z the effect of an inhibitor at Y might be weaker, and vice versa. With other patterns of communication, there can of course also be cases where activator and inhibitor binding are mutually exclusive. Alternatively, if this protein was an oligomer, X, Z and Y could be identical sites with positive or negative binding cooperativity.
To quantify the strength of communication between two sites P and Q, as described in the previous paragraph, we introduce the leverage coupling DPQ. In the following, lower case roman indices (i, j) will number residues, lower case greek indices (μ, ν) normal modes, and upper case roman indices (P, Q) sets of residues, such as probe locations (see Methods) or biological binding sites. We denote the binding leverage of probe location P due to normal mode μ as LPμ (see Methods). The symbol ΔiP is 1 if residue , and 0 otherwise. The leverage λiμ for a given residue and normal mode is then
This calculation is done because our simulations generate a highly redundant set of probe locations, i.e. the denominator above can be large. Similarly, for an arbitrary set of residues P, we writewhere the norm of P is the number of elements in the set. Next, we introduce the vector , where n is the number of modes considered. The scalar productis large only if the sets P and Q have high leverage for the same normal modes. We will call the quantity DPQ the leverage coupling between the two sites. For example, for the two normal modes in Figure 1, DXY and DXZ are large, and DXW, DZY, DZW and DYW are small. Similarly, the matrix measures the normalized leverage coupling and has the range 0≤CPQ≤1. Since DPQ is based on normal mode vectors that represent infinitesimal motion, and depends on the size of the probe used in the calculation of LPμ, the scale of leverage coupling values is arbitrary and unique to each protein. We therefore always compare the leverage coupling of specific sites to the average coupling between the residues not belonging to any sites, i.e. the background leverage coupling for a given structure.
The leverage coupling DPQ gives a measure of the strength of site-site coupling, but depends directly on the magnitude of conformational change at the different sites. In molecular machines like the chaperones, the conformational change at binding sites is small compared to the large-scale functional motions. Here, the measure CPQ can be used instead of DPQ to see how binding sites are correlated with different modes of functional motion. In this case we are interested in comparing the values between different sites and look for the most correlated pairs of sites for a given protein. The range of color bars in all figures containing CPQ matrices is from 0 to 1, which reflects the span of CPQ values. Finally, the special case where one of the sets only has one residue can be used to see how one site couples to the rest of the protein. We will denote this variant DPi, where P is the studied site and the index i runs over all residues.
We study 15 enzymes regulated by ligand binding, 14 of which were studied in our papers on binding leverage [15] and local closeness [19]. The addition is the 20-meric enzyme GTP cyclohydrolase I (GTPCHI), which is both activated and inhibited allosterically by different substances [20], [21]. These 15 enzymes are supplemented by 5 additional proteins, to generalize the analysis to other types of regulation and non-enzymes: Glycogen phosphorylase (GP) is allosterically regulated by both phosphorylation and ligand-binding [22]. The serine-protease thrombin is allosterically regulated by sodium binding [23]. The type I (GroEL-GroES) and type II (CCT, thermosome) chaperones are molecular machines regulated by ATP binding and hydrolysis [24]. The simulation parameters for the proteins discussed in the main text are summarized in Table 1. The binding leverage was calculated using the ten lowest frequency normal modes [15]. The analysis of all the other proteins in this paper is based on the calculations described in the above paper.
To begin with, we will briefly try to give the reader some intuition of what the leverage profiles can look like and how they relate to each other. The leverage profile similarity (defined in Methods) for the 10 lowest frequency normal modes, excluding the trivial first six modes, is plotted in Figure 2A for four different proteins. A value of 1 indicates that the two corresponding modes affect the exact same sites, and 0 that there is no overlap. Also included in the same panel is the importance of each of these normal modes, Λμμ (see methods). Like for leverage coupling, the scale of leverage profiles is arbitrary and only relative values are relevant. For adenylate kinase (AdK), the most significant leverage profiles correspond to modes 1, 2 and 3. Of these profiles, Λ1 and Λ2 are very similar. Figure 2B shows that these two leverage profiles peak at the same position, whereas the third is spread over more residues. That the leverage profiles are similar means that binding leverage is high for the same sites under the corresponding normal modes, even though these modes are orthogonal. Also included in the figure is the total binding leverage along the sequence, which is the sum of λiμ over all modes μ. Almost all active site residues (involved in ATP and AMP binding) are located at peaks in the total binding leverage.
Having verified that different sites have their highest binding leverage for different normal modes, we move on to the analysis of leverage coupling. Supplementary Figure S1 contains plots of the leverage coupling matrix DPQ for the proteins not discussed in detail in the main text. The figure illustrates that, with the exceptions of ATCase and PTP1B, which we showed were difficult to analyze with binding leverage [15], there is generally a stronger coupling between at least some of the allosteric and active sites (including homotropic communication) than between these sites and the rest of protein. One can also see that some sites are more strongly coupled than others are. We will however not analyze these proteins in detail; instead, we will focus on a couple of noteworthy cases.
The tetrameric enzyme phosphofructokinase (PFK) in Bacillus stearothermophilus has one regulatory site where it is activated by ADP binding and inhibited by phosphoenolpyruvate (PEP) binding. The individual low frequency normal modes for this protein are less similar to each other than for AdK and there are also more modes that contribute significantly to binding leverage (Figure 2A). In Figure 3 we display the leverage coupling DPQ for the four effector sites (P = 1–4, ADP/PEP), the four active sites (P = 5–8, F6P) and the remaining residues of the four chains (P = 9–12, BG). As indicated by the color bars, the figure displays values from 0 to the maximal value of leverage coupling measured, in each matrix. Interactions between the effector sites dominate the matrix, and interactions between effectors and active sites are also strong, whereas interactions between the four active sites are weak. The latter indicates that there could be cooperative binding of effector but not of substrate. Experiments have shown that substrate binding is only cooperative in the presence of PEP [25].
The normalized leverage coupling CPQ is high if the sites P and Q have their peaks in binding leverage for the same modes. The CPQ matrix in Figure 3 for PFK indicates that different sets of modes affect the effector and active sites – the correlations are strong within the two groups of sites, but weaker between them. To demonstrate the validity of this interpretation we also included the DPQ-matrices for four of the individual modes. The modes were chosen from the dominating ones in Figure 2A. Modes 1 and 2 primarily affect the effector sites. Mode 1 also involves some connections between effectors and substrate. Mode 4 essentially only affects the active sites, and is probably responsible for any (weak) substrate binding cooperativity. Mode 10 provides relatively strong connections between the active site and the allosteric site, and Figure 2A shows that this is the second most important mode.
To illustrate the communication between sites we color the surface of the protein by the leverage coupling between one site and each residue of the protein, DPi (see Methods) in Figure 4, the raw data can be found in Figure S2. The coloring in this figure, and in similar ones below, uses cyan for DPi = 0, and magenta for the maximal value of DPi over all residues i for a given site P, i.e. the coloring gives the pattern of communication for a given site, but no indication of coupling strength compared to other sites P. The studied effector site in PFK communicates most strongly with the other effector sites (Figure 4B), whereas the active site is connected with the other active sites, as well as the allosteric site (Figure 4C). This apparent asymmetry comes from the fact that the interaction between effector sites is stronger than between anything else, but the connection between the active site and the effector site has approximately the same strength as the connections between active sites. Noteworthy is also the fact that neither site has any strong connections to sites other than the functional ones.
GTPCHI catalyzes the first step in the production of tetrahydrobiopterin (BH4) from GTP. It has positive cooperativity with respect to GTP binding. Allosteric regulation depends on the presence of the GTPCHI feedback regulatory protein (GFRP). In combination with phenylalanine, GFRP reduces the cooperativity of GTP binding, increasing the activity at low GTP concentrations [20]. The GFRP-GTPCHI complex can also be inhibited by BH4 [21]. Both BH4 and phenylalanine bind at similar locations at the GTPCHI-GFRP interface. The architecture of the GFRP-GPTCHI complex is illustrated in Figure 5B. GTPCHI is a homodecamer arranged in two pentameric rings, and the regulatory GFRP pentamers bind one to each ring.
We analyze three sites in the GFRP-GTPCHI complex, the BH4 site (BH2 in the crystal structure), the phenylalanine site (PHE) and the catalytic site (CAT). We define the catalytic site as all residues interacting with the catalytic Zn, and also His-134 and His-201 as defined in the catalytic site atlas [26]. The two allosteric sites have overlapping locations at the GFRP-GTPCHI interface and therefore have large mutual leverage coupling, as can been seen in Figure 5A, but both also couple strongly to the active site. The coupling between catalytic sites is not very strong in this complex, which is consistent with the fact that GFRP and Phe reduce cooperativity. To test the role of GFRP in modifying cooperativity in terms of binding leverage we removed GFRP from the structure and redid the calculations. The bottom two panels of Figure 5A show the coupling between the 10 different catalytic sites with and without GFRP. The effect is not very strong, but it is clear that the GTPCHI catalytic sites in the structure without GFRP are more strongly coupled compared to the background, than in the structure with GFRP.
The connections DPi between one of the allosteric BH4-sites and the rest of the protein are illustrated in Figure 5C (raw data in Figure S3). Similarly, the coupling to one of the active sites, with and without GFRP present, is shown in Figure 5D and E. In the GFRP-GTPCHI complex the regulatory sites and their surrounding residues have the strongest leverage coupling, as was also seen for the site-site coupling matrix DPQ. This figure however clearly illustrates that communication with the “background” only involves the surroundings of the effector binding sites, and does not involve any other distinct sites.
The concepts of binding leverage and leverage coupling can be generalized to study other forms of allosteric communication. Therefore, we consider cases of regulation involving metal binding and phosphorylation.
We study glycogen phosphorylase (GP) as a case of allosteric regulation via covalent modification. Glycogen phosphorylase has two main conformations: the inactive dimeric T state and the active tetrameric R state [22], [27]. In addition, it has two forms, GPa and GPb, where the former is phosphorylated at Ser14. Crystal structures are available for both R and T state forms of GPa and GPb, but the R state is favored for unliganded GPa, and the T state for unliganded GPb. Both GPa and GPb are heterotropically activated by AMP, and inhibited by ATP and other metabolites. Upon phosphorylation, residues 1–20 become more ordered and move to a new position, 30 Å or so away, as can be seen in Figure 6B.
In our calculations we use crystal structures of rabbit muscle GP. PDB entry 1gpa, representing unliganded GPa, is used for normal mode calculations; in addition we use T state GPb (1a8i) and AMP-activated R state GPb (3e3n) to define different binding sites. To be able to analyze phosphorylation using binding leverage, we treat residues 10–20 as a peptide ligand binding at two different sites, P1 (T state GPb) and P2 (R state GPa), and calculate the normal modes without the 20 first residues. Figure 6A shows DPQ for P1 and P2, and also the active (PLP) and allosteric sites (AMP). It is clear that the connections are strongest between P1 and the PLP site. There is an unexpectedly weak interaction between the AMP and PLP site. Since P1 seems more important than P2 we hypothesize that release of residues 1–20 upon phosphorylation from P1 is more important for allostery than binding to P2. The role of P1 is however somewhat uncertain given that residues 1–20 are relatively disordered in GPb. The connections are more or less symmetric between chains indicating that phosphorylation of one chain can trigger a global conformational change. To illustrate the connections between the active site and the rest of the protein we have drawn DPi for the active site in Figure 6C and D (raw data in Figure S4). This figure clearly shows strong connections between the active sites themselves and with P1, but also towards one side of the dimer interface, opposite to P2, which could contain latent allosteric sites.
We also analyzed yeast glycogen phosphorylase (yGP), which is structurally very similar to rabbit muscle GP, but differently regulated. The N-terminal strand in yGP is 40 residues longer than in rabbit muscle GP. In the GPb form the strand binds to the active site instead of P1, and in the GPa form it folds at the dimer interface, at a position similar to P2 above [28]. The differences in regulatory mechanism between these two proteins are thus primarily due to the differing length of the N-terminal strand. This strand is excluded in our calculations and we therefore do not expect any qualitative differences between the two variants. We analyzed yGP using the same parameters as above, based on PDB entry 1ygp, having removed all residues before position 22 (using the 1ygp numbering). We found that the leverage coupling between the active site and the rest of the protein is essentially identical to that of rabbit muscle GP, indicating that P1 is a latent allosteric site in yGP (data now shown).
As an example of metal binding-induced allostery we study the serine protease thrombin which is allosterically regulated by sodium binding [23]. It is also controlled by two other allosteric sites: exosite I (EX1) interacts with several different protein partners, and exosite II (EX2) interacts with several polyanionic substrates [23]. We divide the active site into three groups, the catalytic triad (CAT) and two of the substrate recognition pockets P2 and P4. The leverage coupling of this protein is shown in Supplementary Figure S5. The binding leverage of the sodium site is very low, and coupling to other sites weak. The sodium-induced conformational change primarily involves side-chain rearrangements, which are not modeled by our procedure. The concept of binding leverage could be expanded to include side-chains at a significant computational cost. Single side-chain rearrangements are however not expected to be modeled by low frequency normal modes, which means that a more refined description of motions would probably also be required to model the sodium regulation.
Above, we analyzed a set of enzymes, some of them very large with up to 3 000 residues (GTPCHI-GFRP, ATCase and GDH), and found that leverage coupling gives an understanding of allosteric communication in these enzymes. To push the envelope even further we will now move to the chaperonins, molecular machines with about 8 000 residues. These large molecules are quite challenging to study, the main bottleneck in our analysis being the time required to generate the very large number of probe locations needed, and the calculations took roughly 30–40 CPU hours for each chaperone on a modern desktop PC.
Chaperonins represent a different type of allostery compared to the homo- and heterotropic regulation seen in enzymes. These molecular machines cycle through a set of conformations to provide a protected chamber for protein folding. ATP binding and hydrolysis cause large conformational changes to facilitate substrate capture, folding and release [24]. We will analyze and compare the bacterial group I chaperonin (GroEL-GroES) and eukaryotic and archaeal group II chaperonins (CCT, Thermosome) to investigate differences in regulatory mechanisms.
The concepts developed in this paper were designed to analyze coupling between distinct ligand binding sites in enzymes, but, given a regulatory site, we can detect which parts of the protein are likely to have conformational change coupled to binding at that site. When a domain is deformed, the domain itself does not have high binding leverage, but many of the domain's hypothetical binding sites do. In this context binding leverage is therefore rather a measure of the degree of deformation of a section of the protein. By computing the leverage coupling DPQ for a site P and a domain Q, we can see how binding at the site P couples to conformational change in domain Q, making it possible to analyze allosteric communication in molecular machines such as chaperones.
The GroEL-GroES chaperone consists of two heptameric rings (GroEL) and a heptameric lid (GroES) attached to one of the GroEL rings (see Figure 7). The ring closest to GroES is called the cis-ring and the other the trans-ring. Each GroEL ring provides a folding chamber. The functional cycle roughly goes through the following steps [24], [29]: After substrate has bound to one of the open GroEL rings, ATP binds cooperatively to the GroEL ring [30] and increases affinity for GroES [31]. GroES binding causes a large conformational change increasing the volume of the cis folding chamber and changing it from hydrophobic to hydrophilic [24], allowing folding to take place [32]. ATP hydrolysis weakens the affinity for GroES and when substrate and ATP have bound to the trans ring GroES and substrate are released from the cis subunit. In addition to intra-ring communication, there is also inter-ring signaling, which (i) adjusts the trans ring to accept substrate after cis ATP hydrolysis; (ii) leads to the ejection of cis substrate as a result of trans ATP binding [33]; (iii) accelerates the ejection of cis substrate by simultaneous binding of ATP and polypeptide to the open trans ring [34]. According to cryo-EM analysis, the equatorial domains play a key role in the inter-ring signaling [35]. Here, we will study the allosteric communication between the cis ATP sites and the rest of the protein.
Conformational changes in GroEL involve the equatorial, intermediate and apical subdomains (see Figure 7). ATP binds to the equatorial domain and GroES to the apical domain. ATP binding controls the expansion of the folding chamber which takes place when the intermediate domain swings away from the equatorial domain. The apical domain follows the intermediate domain in this motion, largely as a rigid body. ATP hydrolysis mainly induces an increased flexibility of the intermediate and apical domains [29], which probably explains the looser attachment of GroES to GroEL-ADP7 than to GroEL-ATP7. ATP binding and hydrolysis is positively cooperative within each ring and negatively cooperative between the rings, providing tight ATP binding to only one ring at a time [36]. Figure 8A shows the leverage coupling DPQ and the normalized CPQ, for the ATP sites, the three subdomains of the cis ring, the trans ring and GroES. The strongest connections are between the chains of GroES. Second in strength are the connections between the apical and intermediate domains and GroES, and between the apical and intermediate domains themselves. The ATP site is also only weakly connected to the protein, a result of the fact that the equatorial domain and the ATP site undergo much smaller conformational change than the other two domains. The normalized leverage coupling CPQ however shows that the ATP site is more correlated with the apical and intermediate domains than with the equatorial domain to which it belongs. Correspondingly, there are strong correlations within the trans ring, where the magnitude of leverage coupling is much lower. The high degree of symmetry of the subsquare of the CPQ matrix describing interactions between the ATP site and the intermediate domain, and partly also the apical domain, is consistent with the positive cooperativity observed for ATP binding within one ring. Finally, there is a weaker correlation between the trans equatorial and intermediate domains, and the cis ring, particularly between the equatorial domains of either ring. These connections could be involved in the negative cooperativity between the two rings.
We also analyzed the coupling DPi for one of the ATP sites, two views of this measure are provided in Figure 7A (raw data in Figure S7). The coloring indicates that the inside of the cis cavity, the GroEL-GroES interface and the interface between apical and intermediate domains are most strongly communicating with the cis ATP site. There is hardly any connection to the trans ring. These findings should be related to the fact that the main function of ATP is to regulate the cis cavity and the interactions with GroES, and also to the positive cooperativity of ATP binding.
The human chaperone CCT has a similar function to GroEL, but does not utilize an analog to GroES. It consists of octameric rings, with similar but non-identical chains, instead of heptameric ones. It is also regulated by ATP binding and hydrolysis with steps similar to those of GroEL [24]. ATP binding is not cooperative, regulation has been described as sequential rather than concerted [37]. This is also reflected in the fact that only a fraction of the 16 ATP pockets were populated in crystal structures (13 in the one we use). The leverage-coupling matrix in Figure 8B shows that some of the apical domains are strongly coupled to each other, but coupling between intermediate domains is weaker. The normalized leverage coupling matrix in the same figure, CPQ, indicates that ATP has a weaker correlation with the apical and intermediate domains in CCT than it does in GroEL. In this plot the chains are ordered alphabetically, i.e. the first eight elements along either axis for each domain (apical, intermediate, equatorial) belong to the same ring, and the last eight to the other. This means that for CCT, interactions between the rings are as strong as within them, which is clearly different from what we saw for GroEL where the trans ring was only weakly connected to the rest of the protein. On the other hand, in CCT there is a greater asymmetry in the allosteric connections within one ring than in GroEL-GroES, in particular between the ATP site and the intermediate domain. This asymmetry is seen from the anisotropy of the different subsquares of the CPQ matrix, and is consistent with the sequential regulation of this chaperone [37].
Figure 7B shows the leverage coupling DPi for one of the ATP sites of CCT (raw data in Figure S7). As for GroEL-GroES (Figure 7A), the ATP site is more strongly connected to the inside of the cavity than the outside, but in this case the pattern is relatively symmetric between the rings. The strongest deviation from symmetry, and also the strongest visible leverage coupling, is to a nearby interface between intermediate and apical domains (magenta area in the middle panel of Figure 7B).
The archaeal thermosome is homologous to CCT, but has a higher degree of symmetry than CCT [38]. The results of the analysis of this protein can be found in Figures S6 and S7. The leverage coupling DPQ and the normalized CPQ in Figure S6A shows a pattern similar to CCT; the communication between apical domains is strong, and ATP is more strongly connected to the intermediate domain than the equatorial domain. The thermosome however displays a higher degree of symmetry (as indicated by the uniformity of subsquares in the matrices). The DPi surfaces for one of the ATP sites in Figure S6B also shows a higher degree of symmetry than for CCT; in particular, the coupling to the neighborhood of the studied site is not stronger than to the rest of the protein. The difference in the symmetry of DPi is especially clear when comparing the two corresponding curves in Figure S7. Symmetry is usually associated with positive cooperativity: the difference in symmetry between CCT and the thermosome might therefore reflect a difference in cooperativity, within the rings.
Comparing to previous computational works [11], [12], [29], [39], [40], we analyze allosteric communication between subunits in complete structures of both group I and group II chaperonins. It allows us to detect symmetry in the interactions between subunits of the cis ring of GroEL-GroES and its absence in CCT. We show that leverage coupling helps to understand positive cooperativity in the cis ring and negative cooperativity in the inter-ring communication in GroEL-Gro-ES, non-cooperative mechanism in human CCT, as well as positive intra-ring cooperativity in archaeal thermosome.
Despite the almost half-century long studies of allostery, the majority of the works represents analysis of individual proteins (or groups of homologs) and mechanisms of allostery characteristic for individual structures. In this work, we sought a structural characteristic that can be used to understand allosteric communication in proteins of different types and sizes, from small single-domain proteins to large multi-chain oligomers and chaperones. We resort here to the thermodynamic aspect of allosteric regulation, where the conformational equilibrium between different structural states and their relative stability determine allosteric communication between sites and effect of regulation. We developed the concept of leverage coupling based on the idea that long-range communication between allosteric sites can be mediated by coherent motion along independent conformational degrees of freedom. We have studied the allosteric regulation of a number of proteins controlled by ligand binding, phosphorylation, or metal binding. The analysis has provided new insight into the allosteric mechanisms involved. Two approaches to the problem have been applied, first an analysis of known biological sites, to see how they are connected to each other, and how coupling between them compares to the background. Second we have selected specific sites and analyzed how these are coupled to the rest of the protein, thus being able to identify important functional regions of the protein, that are communicating with these specific sites, and in some cases see how different sites are coupled to different parts of the protein.
We began our analysis by showing that leverage coupling largely captures the important connections in a number of enzymes, and exemplified this for phosphofructokinase (PFK) and GTP cyclohydrolase I (GTPCHI). We also showed that the role of GFRP in regulating homotropic cooperativity in GTPCHI was described well by leverage coupling. In the case study of allostery by phosphorylation in glycogen phosphorylase, we found indications that the active sites had high leverage coupling with the site where the unphosphorylated N-terminal segment binds (in a low temperature crystal structure), and hypothesize that the release of this segment upon phosphorylation causes the functional regulation. Allosteric regulation by metal binding in thrombin can however not be explained by leverage coupling, at least not in the coarse-grained version employed here. Finally, we have demonstrated that leverage coupling can be used to analyze allosteric communication in three different chaperones, and captures the differences in cooperativity between CCT and GroEL-GroES. We were able to describe allosteric communication between structural subunits providing positive cooperativity within each ring and negative cooperativity between the rings via inter-ring communication.
The concept of allosteric communication mediated by collective degrees of freedom, as presented here, is based on our understanding of the physical principles determining protein dynamics. Using normal modes and coarse-grained docking simulations is a crude approximation of these principles – a complete description of the processes involved requires a statistical mechanics analysis based on a reliable energy function and proper conformational sampling. However, our analysis is successful in identifying communicating pairs of sites in the majority of the studied proteins, supporting our assumption that allosteric regulation relies on coherent conformational changes of oligomeric proteins and their domains. We have furthermore demonstrated that different regulatory sites have different patterns of communication (see for instance the difference between active and allosteric sites in PFK), which are determined by motion along independent structural degrees of freedom, in our case different normal modes. This finding gives strong support to the idea that the ability of particular sites to couple to certain modes of motion, and not others, as illustrated in Figure 1, can provide directed and differential allosteric communication and regulation. We have thus moved beyond the framework defined by the classical KNF and MWC models, both in that we propose a molecular mechanism for connecting different sites, and in that we are able to predict and identify many functional sites. Using normal modes to represent independent conformational degrees of freedom, we find that these motions can be used not only to describe the allosteric transition geometrically – as many have done before – but also to explain allosteric connections between different binding sites and to identify latent allosteric sites. Novel allosteric connections predicted by leverage coupling can be used as targets in experimental inhibitor/activator design.
The calculation of binding leverage involves two main steps, generation of possible ligand conformations through coarse-grained Monte Carlo simulations, and analysis of the generated binding sites with respect to motions deduced from one or more crystal structures [15]. Probe conformations in which the probe is highly stressed, under a given protein motion, have high binding leverage. Binding leverage models allostery based on the assumption that binding to sites where ligand-protein interactions are connected to important degrees of freedom can affect the conformational equilibrium. We used binding leverage to rank probe locations (defined below) and found that high-ranking probe locations matched active and allosteric sites in a wide range of proteins. Here, we will give a brief overview of the procedure, which was described in detail previously [15].
Ligand binding is simulated with a completely fixed Cα-representation of the protein chain and a freely moving probe ligand in the form of a peptide with one or more Cα-atoms. The probe and protein interact via a square well potential which is attractive for Cα-Cα distances between 5.5 and 8 Å. Distances shorter than 4.5 Å are forbidden. Potential binding sites, called probe locations, are generated by running a number of short docking simulations. A probe location is defined as the residues interacting with the probe at the end of a given simulation. Binding leverage measures the ability of a probe ligand to resist a given motion, for example that of a normal mode. A spring is placed between all residue pairs in a probe location whose interconnecting lines pass through the ligand. The binding leverage of a probe location is then calculated as the total change in spring potential energy U due to a given motion, i.e.where summation is over all springs, and the additional index μ numbers the motion vectors used, i.e. one leverage is calculated for each vector. If more than one motion is considered the binding leverage can be summed to a total binding leverage for the probe location.
Cα normal modes were calculated using MMTK with default parameters for all cases [41]. For the large proteins GTPCHI, GroEL-GroES, CCT and the thermosome we used the Fourier-basis approximation [42], in all other cases vibrational modes are used.
The binding leverage of residues i under mode μ (defined in the main text) can be grouped into leverage profiles , where m is the number of residues. We write the scalar product between two profiles as
The magnitude Λμμ of a leverage profile indicates the importance of the corresponding normal mode in the total binding leverage, and the normalized scalar productis close to one when the corresponding modes involve the same binding sites, and close to zero when the overlap is small.
|
10.1371/journal.pcbi.1002420 | Conformational Control of the Binding of the Transactivation Domain of the MLL Protein and c-Myb to the KIX Domain of CREB | The KIX domain of CBP is a transcriptional coactivator. Concomitant binding to the activation domain of proto-oncogene protein c-Myb and the transactivation domain of the trithorax group protein mixed lineage leukemia (MLL) transcription factor lead to the biologically active ternary MLL∶KIX∶c-Myb complex which plays a role in Pol II-mediated transcription. The binding of the activation domain of MLL to KIX enhances c-Myb binding. Here we carried out molecular dynamics (MD) simulations for the MLL∶KIX∶c-Myb ternary complex, its binary components and KIX with the goal of providing a mechanistic explanation for the experimental observations. The dynamic behavior revealed that the MLL binding site is allosterically coupled to the c-Myb binding site. MLL binding redistributes the conformational ensemble of KIX, leading to higher populations of states which favor c-Myb binding. The key element in the allosteric communication pathways is the KIX loop, which acts as a control mechanism to enhance subsequent binding events. We tested this conclusion by in silico mutations of loop residues in the KIX∶MLL complex and by comparing wild type and mutant dynamics through MD simulations. The loop assumed MLL binding conformation similar to that observed in the KIX∶c-Myb state which disfavors the allosteric network. The coupling with c-Myb binding site faded, abolishing the positive cooperativity observed in the presence of MLL. Our major conclusion is that by eliciting a loop-mediated allosteric switch between the different states following the binding events, transcriptional activation can be regulated. The KIX system presents an example how nature makes use of conformational control in higher level regulation of transcriptional activity and thus cellular events.
| CBP (CREB-binding protein) is a transcriptional regulator of RNA polymerase II-mediated transcription. KIX is a domain of CBP. KIX binding to the activation domain of proto-oncogene protein c-Myb and the transactivation domain of the trithorax group protein mixed lineage leukemia (MLL) transcription factor forms the biologically active ternary MLL∶KIX∶c-Myb complex. This complex has been shown to play a key role in Pol II-mediated transcription. Experimental data show that the binding of the activation domain of MLL to KIX enhances c-Myb binding. Here, we studied the c-Myb∶KIX∶MLL ternary structure and models based on this structure, KIX-only, KIX-MLL, and c-Myb-KIX, by explicit solvent molecular dynamics (MD) simulations. MD simulations can help in figuring out allosteric events and thus functional mechanisms. Our in silico analysis of the dynamic behavior indicated that when MLL is bound, there is allosteric communication between the two KIX binding sites. We observed a shift in the conformational ensemble of KIX upon the binding of the MLL activation domain, leading to a pre-organization of the KIX∶c-Myb binding site and explaining the enhanced c-Myb binding observed experimentally. On the other hand, this is not the case if c-Myb binds to KIX, which suggests that KIX regulation is under conformational control.
| Allostery plays a crucial role in biological processes on the molecular and cellular levels [1]–[3]. Allostery involves communication between binding sites [4]–[6]. Following a binding event, the most populated conformation may differ from the one which prevails in the free state. This can be described by a shift in the free energy landscape, where the populations in the conformational ensembles are redistributed [1], [7]–[10]. The formation of the ternary complex of the KIX domain of the CREB (cyclic-AMP responsive element binding protein) binding protein (CBP or CREBBP), the transactivation domain of the MLL, and the proto-oncogene protein c-Myb is cooperative. Binding of MLL to KIX elicits an allosteric effect that enhances the binding of KIX to c-Myb [11]. The affinity of c-Myb binding to KIX-MLL binary complex is two-fold higher than of KIX alone [7], [12].
MLL and c-Myb are crucial elements for normal blood cell development. c-Myb is expressed widely in cells in the proliferative state, as in hematopoietic cells, whereas MLL ensures that the transcriptionally active state of a group of genes is maintained. CBPs play roles in cellular growth, differentiation, proliferation and apoptosis [13]–[15]. CBP binds to phosphorylated CREB, and is a modular coactivator of RNA polymerase II-mediated transcription [13], [16], [17]. It has several interacting domains which provide a scaffold for multiprotein assembly, including transcription factors, signaling molecules and hormone receptors [14], [18]–[20]. Thus, the control of the KIX binding states is crucial, and is related to many diseases, from cancer to neurodegenerative diseases [13], [18]–[23]. (See supporting information Text S1).
CBP has two distinct binding surfaces on the KIX domain that can be occupied simultaneously [12]. The activation domains of several proteins including MLL, Tax, Tat and c-Jun can bind these sites which are located roughly opposite to the c-Myb binding site [24], [25]. The KIX structure consists of a three helix bundle (α1, residues: Gln597–Ile611; α2, residues: Arg623–Tyr640; α3, residues: Arg646–Arg669) and a 310 helix (G1, residues: Trp591–His594) at the N-terminal. The C-terminal region is linked to a second 310 helix (G2, residues Pro617– Lys621) through a loop (L12) [11] (Figure 1). The binding of the MLL activation domain occurs on the hydrophobic groove formed by the L12 loop and the α2 helix region of KIX, specifically by the side chains of Ile611 and Phe612, aliphatic region of Arg624, Leu628, Tyr631 and Leu664. The side chains of Ile2849, Phe2852, Val2853, Leu2854 and Thr2857 (Ile849, Phe852, Val853, Leu854 and Thr857 with the numbering convention in the structure file [11]) of MLL make contacts with Phe612, Pro613, Arg624, Asn627, Leu628 and Tyr631 in the hydrophobic binding groove of KIX. Ile2844, Leu2845 and Pro2846 (Ile844, Leu845 and Pro846) of MLL fit in the hydrophobic binding groove of KIX consisting of residues Met639, Lys656, Ile660, Ile664 and Arg668 (Text S1).
The c-Myb activation domain-binding site resides on the hydrophobic groove which is formed by the first and third helices of KIX, specifically by the side chains of Leu599, His602, Leu603, the aliphatic portion of Lys606, Leu607, Ala610 from the α1 helix, Tyr650, Leu653, Ala654, Ile657, Tyr658 and the aliphatic region of Lys662 from the α3 helix [11]. This binding groove provides a docking surface for the nonpolar side of c-Myb consisting of the side chains of Ile295, Leu298, Leu301, Leu302, Met303, Thr305 and Leu309. This hydrophobic groove is responsible for more than half of the interactions between c-Myb and KIX. The electrostatic interactions of c-Myb Arg294 with Glu665 of KIX along with the interactions of c-Myb Glu306 and KIX Arg646 determine the orientation of c-Myb in the groove [26]. Stabilization of the C-terminal of the α3 helix of KIX, which has most of the allosterically important residues, was suggested to increase the electrostatic interactions by bringing Glu665 and Glu666 of KIX in close proximity to Arg294 and Lys291 of c-Myb, respectively. The two hydrophobic binding regions which c-Myb and MLL bind (Figure 1) have no direct contact [11]. However, binding is cooperative, which is why it is considered allosteric.
Significant changes are observed at the MLL binding interface of KIX in the ternary complex compared to the KIX∶c-Myb binary complex, with the disordered regions in KIX in the KIX∶c-Myb binary structure becoming structured in the ternary structure [11]. Two regions become more ordered: loop L12 (between helices α1 and G2 regions, Figure 1) is observed to be in contact in the ternary complex and the extended C-terminal end (the end of the α3 helix) is stabilized in the presence of MLL [11]. Recent findings suggest that formation of the binary KIX-MLL complex leads to a redistribution of the KIX conformational ensemble, allosterically activating the c-Myb binding site, and that the conformation of KIX in the KIX-MLL binary complex resembles that in the ternary complex [7].
In this work, we studied the c-Myb∶KIX∶MLL ternary structure (PDB code: 2AGH [11]) and models based on this structure (described in the Materials and Methods section), by explicit solvent molecular dynamics (MD) simulations. MD simulations can help in figuring out the allosteric events to understand functional mechanisms [2], [7], [27]–[31]. Comparative analysis of the dynamic behavior showed that there is allosteric communication between the two binding sites on KIX when MLL is bound. The shift in the conformational ensemble of KIX upon the binding of the MLL activation domain demonstrated that the binding pre-organizes the KIX∶c-Myb binding site, which explains the more favorable interaction.
From the RMSF (root mean square fluctuations) of KIX, KIX∶MLL, c-Myb∶KIX and c-Myb∶KIX∶MLL simulations, the most mobile region of KIX is at Phe612∶Lys621 (Figure 2). This region corresponds to the L12 and G2 parts of the KIX structure (Figure 1), which is located close to the MLL binding site [11]. Comparing the RMSF of the simulations of the four models, we observed that c-Myb binding reduces the mobility of the L12 and G2 region. Similar behavior is also observed in parallel simulations (Figure S1). Further, the movement of this region is accompanied by the motion of the Phe612 side chain, which is at the MLL binding site and has a role in signal transmission to the c-Myb binding site (Figure 3). Upon binding of MLL, the side chain of Phe612 fits itself in the core formed by residues Phe2851, Val2852 and Asn2856 (Phe851, Val852 and Asn856) forming hydrophobic contacts with the MLL amphipathic helix [11]. In Figure 3, the motion of the L12 and G2 region can be observed clearly. Without MLL and c-Myb (i.e. in the KIX-only simulation) this region bends towards the α1 helix; in the presence of c-Myb (the c-Myb∶KIX simulation) this region moves further apart restricting the conformational space of the loop region similar to the conformation in the KIX-only simulation (Figure 3). This conformation of the loop might have a role in MLL binding as the conformational space of MLL binding interface becomes more accessible. As the loop gets stabilized in this state, the α3 helix of KIX which is in contact with the tail of MLL [11], is observed to move away from the core of the protein, facilitating MLL binding. Zor and coworkers [26] have shown that the C-terminal helix of KIX is partially disordered. However, melting or unfolding has not been observed on the time scales of the MD simulations of this work. Here, we observe that the α3 helix of KIX moves away from the core of the protein with relatively higher amplitude fluctuations, however, still keeping its correlated fluctuations with the N-terminal region of KIX and c-Myb.
The network of correlated fluctuations of the KIX-only structure (Figure 4a) demonstrates that the fluctuations of the Ala630-Ser642 region from the α2 helix and the His592-Pro617 region from the G1 and α1 helix, which contain both the MLL and the c-Myb binding sites, are positively correlated. That can be an indication of the movement of the loop (G2 and L12) as discussed in the previous section. With MLL binding the correlation increases (Figure 4b), because the L12 and G2 region moves towards MLL and enhances the interaction network. In addition, the Lys606-Leu628 region, which includes L12, G2 and the beginning of the α2 helix, has stronger correlations with His651-Glu666 in the α3 helix in the KIX∶MLL bound form. Expectedly, those regions form a favorable groove for MLL binding. These two regions were previously suggested to be critical for forming a coupled network of interactions through which the propagation of the allosteric effect following MLL binding is transmitted to the c-Myb binding groove [7], [11], [32].
In the case of the c-Myb∶KIX simulation, the correlations of Lys606-Leu628 with His651-Glu666 fade again in comparison to the isolated KIX and KIX∶MLL simulations. Nonetheless, the correlations of the Lys621-Arg646 in the α3 helix and the Trp591-Asp616 region of the α1 helix along with the G2 and L12 regions can still be observed as compared to the KIX-only simulation (Figure 4c). Besides, from these correlations it can be inferred that c-Myb binding affects the α3 helix with which MLL makes contacts. When the side chains of the C-terminal are investigated, in the case of the KIX and KIX∶MLL binary complex simulations, the side chains of the α3 helix point inwards to the MLL binding groove. This could favor the formation of the network of interactions of the α3 helix with the rest of the structure. On the other hand, upon c-Myb binding (the c-Myb∶KIX simulation), the side chains of the α3 helix point outwards. Thus, it can be deduced that the presence of MLL affects the side chains of the α3 helix to strengthen the communication network with the rest of the structure which in turn would allosterically favor c-Myb binding.
When the cross correlations are calculated from the simulation of the ternary structure all correlations (Lys606-Leu628 with His651-Glu666, Lys621-Arg646 with Trp591-Lys621) are emphasized (Figure 4d). Expectedly, the correlations of two regions Lys606-Leu628 and His651-Glu666 of the α3 helix are more pronounced for the c-Myb∶KIX∶MLL case. This result is an indication that the two regions (Lys606-Leu628 and His651-Glu666) have roles in allosteric signal transmission. Similar results are obtained in parallel simulations (See Figure S3).
The most compelling conclusion based on the cross correlation analysis of the pseudo-dihedral angles along with the phi and psi angles (Figures S4, S5 and S6), is that the G2 and L12 region has strong correlations with the C-terminal α3 region, with both having a role in the formation of the network of interactions which pre-organize the microenvironment for the MLL binding. In parallel to our findings up to this point, these correlations fade in the cross correlations of the pseudo-dihedral angles as well as in the phi and psi angles calculated over the c-Myb∶KIX binary complex simulation. On the other hand, correlations of the same regions for the c-Myb∶KIX∶MLL complex remain as expected (Figures S4, S5 and S6). Thus, the formation of coupled correlations increases when MLL is present.
The KIX, KIX∶MLL, and c-myb∶KIX∶MLL simulations all show similar patterns in the cross-correlation maps, reflecting the outcome of the redistributions of the ensemble of conformations. To further investigate how the ensemble of KIX shifts and what are the prevailing states in the various forms (KIX, KIX∶MLL, c-Myb∶KIX, c-MYb∶KIX∶MLL), we perform clustering over the sampled conformations during the simulations. [33], [34] We follow a combinatorial approach: all the conformations which were created from the four models are provided as the input to the clustering algorithm. [34] Conformations within 2.75 Å RMSD were clustered, leading to three clusters of conformations as indicated by different colors in Figure 5. As can be clearly seen, the conformations from the KIX-only simulations exist in all three clusters (red, blue and green); yet the majority of the conformations distributed between two clusters (blue and green). With MLL binding, the KIX conformations are mainly observed in the blue cluster, which is the cluster of almost all of the KIX bound to both MLL and c-Myb conformations. This clearly indicates that following MLL binding, KIX already assumes the conformational ensemble which it presents when bound to both MLL and c-Myb. In contrast, the conformational ensemble of KIX bound to c-Myb, is distributed among the three clusters, where the conformations of KIX bound to both MLL and c-Myb are not as frequently visited as in the case of KIX bound to MLL (Figure 5). Quantitatively, 58% of the ensemble of isolated KIX conformations belongs to the ensemble of c-Myb∶KIX∶MLL simulation. This value increases to 99% when the ensemble of KIX∶MLL is considered, which implies that the binding of MLL to KIX results in a shift in the ensembles of KIX towards a more favorable state so that the c-Myb∶KIX∶MLL ternary structure can form. On the other hand, the c-Myb∶KIX, which is comprised of the c-Myb∶KIX binary structure, has only 47% resemblance to the ternary complex which is close to the case of the isolated KIX, suggesting that conformations suitable for ternary complex formation are also present in the c-Myb∶KIX case. The most visited cluster in the c-Myb∶KIX ensemble is shown in red. It consists of conformations with the loop region bending outwards. Furthermore, the percentage of the stable conformations (shown by blue dots) in the simulation of the KIX∶MLL binary complex is much higher than in the simulation of isolated KIX, clearly indicating that KIX∶MLL binds to c-Myb with a higher affinity than KIX itself. The results are consistent with those of parallel simulations (See Figure S7).
In light of these findings, it is clear that MLL binding results in a redistribution of the conformational ensembles of KIX such that it favors c-Myb binding.
According to the correlation maps and the observed motion of loop L12 and G2 in the Lys606-Leu628 region, we identified this loop region as essential for ternary complex formation with the sequence of events starting from MLL binding to KIX. In the presence of MLL, the loop (L12 and G2) is bent towards it. On the other hand, upon c-Myb binding to KIX, Met625-Leu628 which resides at the N-terminal of the α2 helix is observed to change its conformation in the opposite direction. The significant difference between KIX∶MLL and c-Myb∶KIX in conformations and in correlation maps suggests that these residues could be the key in controlling the bending of the L12 and G2 regions and thus the allosteric switch. To provide further evidence for the importance of residues in the Met625-Leu628 region for the coupling of the loop motions, in silico mutations are performed for selected residues. A change of one residue in that region could interfere with the motion of the loop and block the ternary complex formation.
In order for the Met625-Leu628 region to be more flexible and act like a hinge as in the case of KIX and c-Myb∶KIX simulations, the loop region should bend outward, rather than towards MLL. We selected Glu626 and Leu628 [35] in the KIX∶MLL structure as representative residues and mutated them to Ala. The selection of these particular residues is based on their correlation maps and the fact that the orientation of their side chains displayed considerable changes in the simulations of different models.
The total energies (the kinetic and potential energies of each frame) of E626A and L628A mutants and wild type of KIX∶MLL are presented in Figure S8, using an approach which is similar to that used in the previous in silico mutant MD simulation studies [30], [31]. As the figure shows, for both mutants, the total energies remain stable over time. The differences between the models simply arise from the change in the interaction network with the mutations. The RMSF and RMSD of the core region of KIX from the wild type, E626A and L628A mutants of the KIX∶MLL complex, are shown in Figures S9 and S10. None provide any evidence that the mutations destabilize the region. All simulations indicate energetically and conformationally stable ensembles for these two mutations. Several other mutations on the same loop were designed such as P613A, T614A and D622A all of which disrupted the structural and energetic stability of the structure. On the other hand, some mutations, such as P613A did not lead to any change in the cross-correlation maps or in the loop motion, retaining the allosteric relationship between the two binding sites of KIX.
Analysis of the Glu626Ala and Leu628Ala mutant trajectories of KIX∶MLL, revealed that the loop region bends outwards, as in the KIX and c-Myb∶KIX structure (Figure 6). This implies that the loop region interferes with the interacting region of the long MLL tail. This mutation clarifies that the outward movement of the loop hampers the interaction of KIX with MLL's tail, which in turn hampers the allosteric communication network. Since the correlations of Lys606-Leu628 with His651-Glu666 fade in the cross correlation maps of the mutant cases (Figure S11a–d) as compared to KIX∶MLL (Figure 4b), it appears that each mutation distorts the network of interactions of KIX in the KIX∶MLL structure; creates couplings similar to those in the c-Myb∶KIX (Figure 4c); and results in a loop conformation that is open, away from the core of the structure, as in c-Myb∶KIX. In this open conformation no distinct couplings are observed in the allosteric network. In our parallel simulations (Table S.1) we have obtained the similar results (Figures 4, S2 and S11). Thus, our mutations could disfavor the allosteric signaling and decrease the affinity of formation of the ternary complex, which is in agreement with the work of Arai and coworkers [35] in which they mutated Leu628Asp on KIX protein and observed very weak MLL binding.
The present MD simulations provide the dynamics on nanosecond time scale (See Table S1). The results are in agreement with parallel simulations, and as such can be considered as statistically significant in the time window of the simulations. However, this does not exclude any mutation-induced effect on the structure and dynamics that might emerge at longer time scales.
The allosteric effect upon formation of the MLL∶KIX complex favors c-Myb binding; on the other hand, formation of the c-Myb∶KIX complex does not enhance MLL binding [7], [12]. By simulating the KIX, KIX∶MLL, c-Myb∶KIX and c-Myb∶KIX∶MLL structures, we revealed the allosteric mechanism which explains these opposing effects that have crucial consequences in the activation of Pol II-mediated transcription. Our major conclusion is that the L12 and G2 loop region plays a key role in the conformation control, acting as a switch: in the KIX-only structure it swings toward the MLL-favored position, pre-organizing the MLL binding site. In contrast, when c-Myb binds to KIX, it swings in the opposite direction. When considering the opposite sequence of events, when MLL binds first, a clear correlation emerges between the fluctuations of the two binding sites, and the loop region. Clustering of the merged trajectories from the KIX, KIX∶MLL, c-Myb∶KIX and c-Myb∶KIX∶MLL simulations reveals that MLL binding redistributes the conformational ensemble of KIX and leads to a more stable ensemble and a higher percentage of population of the favored state, explaining the higher binding affinity of c-Myb to KIX∶MLL compared to KIX only [12]. Further, we observe that the conformation of L12 and G2 loop region is coupled to the α3 helix conformational changes, which further help MLL binding. Our correlation maps clearly demonstrate that the L12 and G2 loop (Lys606-Leu628) and the α3 helix (His651-Glu666) regions are involved in the transmission of the allosteric signal. These correlations with the MLL interaction sites, that is, the allosteric network, fade in the c-Myb∶KIX structure due to the absence of MLL. To further test our conclusions, we performed in silico mutations of Glu626 and Leu628 residues in the loop region in the KIX∶MLL structure. Our analysis revealed that these mutations led to the loop swinging away from MLL, similar to the conformation observed for the c-Myb∶KIX. That is, the mutations distorted the KIX structure in a way that would block the allosteric signal transmission of MLL.
Allosteric control has been suggested earlier based on conformational dynamics [36]. Here, our results suggest a mechanism of conformational control in the c-Myb∶KIX∶MLL assembly, and provide an insight which can help to understand how KIX can regulate Pol II-mediated transcription. The key element in these allosteric events is the loop region, which appears to be a pivotal component in the allosteric signal transmission from MLL via KIX to the c-Myb binding site. The detailed analysis of the conformational ensembles provides the mechanistic explanation of how the unidirectional allosteric coupling between the two binding sites of KIX is maintained by a non-binding site conformational switch of the loop.
To this end, the ensemble of conformations with residue fluctuations and their correlations provide a means to study inter-relationships between binding or other functional allosteric events. A dynamic control mechanism of function is essential; and here, through such residue fluctuations and correlations, we observed that loops can play a key role. While here we observed their role in KIX, we believe that a mechanism where loops govern allosteric communication is likely to be general. To this end, experimental studies may test the significance of loops and their role in allosteric control of signal transmission.
Four models were simulated to shed light on the underlying dynamics and allostery elicited by MLL binding to the KIX domain of CREB-binding protein in the c-Myb∶KIX∶MLL system. The initial structures were based on the NMR structure of c-Myb∶KIX∶MLL (PDB id 2AGH [11]). The systems simulated are: The isolated structure of the KIX domain of CREB-binding protein (wild type); the isolated structure of KIX domain and MLL binary complex (wild type, E626A mutant and L628A mutant); the isolated structure of c-Myb and KIX binary complex (wild type) and the original ternary complex of c-Myb, KIX domain and MLL. The models are created by removing either MLL or c-Myb or both from the NMR structure prior to solvation and minimization. Chlorine and Sodium ions are added to neutralize the overall charge of the system. The protein was solvated in a TIP3P type [37] octahedral water box. The solvated system went through an energy minimization procedure to remove steric overlaps and unfavorable contacts. In principle, obtaining different models by removing parts of a structure may lead to major conformational changes. The RMSD values of the core region of KIX taken from the equilibrated structures with respect to the NMR structure (PDB id 2AGH [11]) are summarized in Table 1. As can be seen, the modeled starting structures do not show considerable conformational changes. Comparison of the equilibrated KIX∶c-Myb structure to the c-Myb∶KIX NMR structure (PDB id: 1SB0) [26] displays a smaller RMSD (0.96 Å for the core region) than when compared to the co-crystal c-Myb∶KIX structure in the ternary NMR structure (1.63 Å for the core region) (PDB id: 2AGH) [11] (See Figure S12). Lastly, a summary of the MD simulations is given in Table 1.
MD simulations were carried out using the AMBER package with periodic boundary conditions to mimic the boundary effects and include the solvent effect with a relatively small number of particles [38]–[40]. The ff03 force field was used [41]. Isobaric periodic boundary conditions are used with isotropic position scaling. The initial velocities of atoms are generated at 10 K with a Maxwellian distribution. Next, the temperature was gradually raised to 300K and maintained. The pressure is kept at 1 bar by the Berendsen weak-coupling approach [42]. A time step of 2 fs was used in the Leapfrog algorithm. Coordinates are recorded every 1 ps. The simulation trajectories were then analyzed for the sampled conformations in all models.
|
10.1371/journal.pntd.0004789 | Development and Application of a Loop-Mediated Isothermal Amplification (LAMP) Approach for the Rapid Detection of Dirofilaria repens from Biological Samples | Dirofilariasis by Dirofilaria repens is an important mosquito vector borne parasitosis, and the dog represents the natural host and reservoir of the parasite. This filarial nematode can also induce disease in humans, and in the last decades an increasing number of cases have been being reported. The present study describes the first loop mediated isothermal amplification (LAMP) assay to detect D. repens DNA in blood and mosquitoes. Two versions of the technique have been developed and described: in the first, the amplification is followed point by point through a real time PCR instrument (ReT-LAMP); in the second, the amplification is visualized by checking UV fluorescence of the reaction mixture after addition of propidium iodide (PI-LAMP). The two variants use the same set of 4 primers targeting the D. repens cytochrome oxidase subunit I (COI) gene. To assess the specificity of the method, reactions were carried out by using DNA from the major zoonotic parasites of the family of Onchocercidae, and no amplification was observed. The lower limit of detection of the ReT-LAMP assay was 0.15 fg/μl (corresponding to about 50 copy of COI gene per μl). Results suggest that the described assay is specific, and its sensitivity is higher than the conventional PCR based on the same gene. It is also provide a rapid and cost-effective molecular detection of D. repens, mainly when PI-LAMP is applied, and it should be performed in areas where this emerging parasitosis is endemic.
| Dirofilaria repens is a filarial nematode which mainly infests the dog, but humans may be occasionally infested, too. The spread of the parasite is mediated by a number of mosquitoes species, which are well recognized as vectors of D. repens. The majority of reports of the disease come from the European Countries, especially those along the Mediterranean basin, but in the last decade several cases have been recorded also from Asian and African Countries, and this led the scientific community to consider such parasitosis an emerging disease. To date, diagnosis is based the morphologic analysis of microfilariae, isolated from the blood of infected hosts, but this may be time-consuming and the identification of parasite requires specialized parasitologists. The here described approach, based on the loop-mediated isothermal amplification (LAMP), allows the detection of D. repens genomic DNA directly from the biological samples, and it may be easily and rapidly performed, producing unequivocal results in less than a hour. We also presented two versions of the assays. The first, a real-time LAMP, is characterized by a very high sensitivity but it requires an expensive real time PCR instrument, while the second, performed with the addition of propidium iodide, does not need such equipment, therefore being very affordable. This makes it suitable to be carried out in field and whenever expensive equipment or specialized personnel lacks.
| Dirofilariases are parasitic diseases caused by nematodes of the genus Dirofilaria (Nematoda: Onchocercidae) and transmitted by hemathophagous arthropods. The subgenera Dirofilaria (Nochtiella) is represented by 22 species including Dirofilaria repens Railliet et Henry, 1911 [1]. The latter is a common mosquito–borne parasite of subcutaneous tissues of dogs and others carnivores in the Old World [2], including wolf and fox [3]. Cats and feline in general are less susceptible to the infestation, which consists of a low microfilaremia [4–5]. D. repens has zoonotic potential: it can infest humans, and it is considered one of the most important vector-borne parasitosis in humans in Europe [6].
Among these hosts, the dog is the most important and it also act as a reservoir for the parasite [7–8]. From dog, the parasite may be transmitted to several species of mosquitoes (Diptera, Culicidae) [9], which have been proved to act as a D. repens vector [10–11]. Specifically, they may transmit infective third-stage larvae from animals with microfilariemia to humans during the blood-feeding. Therefore, although the nematode is not highly pathogenic for the dog, where it usually causes subcutaneous tissue diseases, it is considered of primary veterinary importance due to its zoonotic behavior [10].
In humans the parasite may locate itself in the subcutaneous tissues, mainly in the upper half of the body, but all regions may be potentially involved. Secondarily, it can migrate from the initial site of infection to others sites, commonly leading to subconjunctival and periorbital infestation [12], which may lead to severe ocular complications. Pulmonary forms have been reported, although rarely [13]. Microfilaremia has never been observed in humans, with only one exception [14].
The disease is widely diffused and, currently, the number of reports from humans and animals is increasing. Most of them come from the Mediterranean Basin, an endemic area of dirofilariasis, caused by both D. repens and D. immitis [10]. However, in the last years, cases were recorded not only from the endemic areas of Europe, including Italy, France Spain, Russia and Turkey [10, 15, 13, 16–17] but also from several Asiatic and African countries, such as India [18], Vietnam [19], Iran [20], Tunisia [21], Egypt [22], and South Africa [23]. This is leading the scientific community to consider the disease as an emerging zoonosis. Several factors are thought to contribute to this expansion, such as the increase in frequency and numbers of travels and movements of animals, the wider distribution of vector-competent mosquito species due to the international trade and global warming, and the improvement of the diagnostic techniques which enable more accurate detection of the pathogen [10, 15, 24].
These circumstances stress the need to have diagnostic tools effective in terms of accuracy, sensitivity and user-friendliness. A recently developed amplification technique, named loop-mediated isothermal amplification (LAMP), owns most of those features, so that it has been finding large application as a diagnostic tool [25]. A brief description of the principle and mechanism, as firstly described by Notomi et al. [26], is reported in the supplementary material S1 Text. This study describes two versions of a novel loop-mediated isothermal amplification assay for a rapid diagnosis of D. repens in dogs and mosquitoes.
The available sequences (Accession numbers AJ271614, AM749230, AM749231, AM749232, AM749233, AM749234, DQ358814, JF461458 and KF692102) of D. repens cytochrome oxidase subunit I (COI) from GenBank were aligned by ClustalW, implemented in MEGA 6.0 [27]. The resulting consensus sequence was then used to design LAMP primers by the mean of the Primer Explorer Program (Available online at http://primerexplorer.jp/e/, latest accessed 8th April 2016). The primers are listed in Table 1. Because no complete sequences of D. repens COI are currently available, the relative positions of oligonucleotides are shown in Fig 1.
Two different variants of LAMP have been developed: a real-time LAMP (hereafter ReT-LAMP) and a propidium iodide LAMP (PI-LAMP). They differed for the visualization of results: in the first cases, the amplification is visualized as a curve in a real-time PCR instrument, while the second allows to visualize the amplification as UV fluorescence following the addition on propidium iodide.
The ReT-LAMP reactions were carried out in 25 μL containing 15 μL of Isothermal Master Mix with carboxyfluorescein (Optigene, Horsham, UK), primers DiRFIP and DiRBIP 0.4 μM each and primers DiRF3 and DiRB3 0.1 μM each. ROX was used as a passive reference dye. Five μL of total DNA were used as template. The mixture was incubated at 65°C for 40 min on a StepOne real time PCR system (Applied Biosystems, Milan, Italy).
The PI-LAMP reactions were carried out as previously described [28]. The reactions (final volume, 25 μL) were prepared by mixing the isothermal amplification buffer (20 mM Tris-HCl, 10 mM [NH4]2SO4, 50 mM KCl, 2 mM MgSO4, 0.1% Tween 20, pH 8.8), dNTPs 1.5 mM each, 1.6 μM each of the DIRFIP and DIRBIP primers, 0.3 μM each of the DIRF3 and DIRB3 primers, 2.5 μL of extracted DNA, and 8 U of Bst 2.0 Warm Start DNA Polymerase (New England Biolabs Inc., Ipswich, MA, USA). The reaction mixture was incubated in a heating block at 65°C for 45 minutes and then heated at 80°C for 5 minutes to terminate the reaction. Propidium iodide to a final concentration of 1 μg/μL was then added to the mixture. The tubes containing the mixture were exposed to UV, and digital images were acquired by the GelDoc System (BioRad Laboratories, Milan, Italy).
In order to confirm the results gathered by LAMP, and to compare the sensitivity of LAMP and traditional PCR, a COI-targeting PCR assays was performed according to the protocol by Casiraghi et al. [29], by using the primers COIintF and COIintR, which are expected to return a 689-pb amplicon from many nematoda species, including those considered in this study.
The products were analyzed in a 1.5% agarose gel and visualized after dying with ethidium bromide 0.5 μg/mL. Images were digitalized by the mean of a Gel Doc EZ system (BioRad Laboratories, Milan, Italy).
A preliminary specificity assay was performed in silico by comparing the primer sequences with the full or partial COI sequences listed in Table 2. The comparison was performed by BLAST [30]. The nucleotide sequences of DiRF1 and DiRF2c (which constituted DiRFIP), and of DiRB1 and DiRB2c (DiRBIP) were split and individually checked. Matches were considered significant when a portion at least corresponding to 95% of the sequence of the checked oligonucleotide was at least 90% identical to the target sequence, without mismatches in the last three bases of the 3' termini.
In vitro specificity assays were performed by using DNA from previously identified nematode species such as Dirofilaria (D.) immitis, Acanthocheilonema (Ac.) reconditum, Angiostrongylus (An.) vasorum, Brugia sp., Loa (L.) loa, Wuchereria (W.) bancrofti, Mansonella (M.) perstans and Cercopithifilaria sp.. Positive controls were carried out with genomic DNA from D. repens.
In order to check the reliability of the test with more complex starting samples, it was also performed by using DNA from canine blood and mosquitoes.
Specifically, 40 canine blood samples were used. They were collected from the cephalic vein of dogs hosted in a dog kennel and previously checked by the qPCR method described by Czajka et al. [31], designed to reveal the presence of Onchocercidae. Out of the forty blood samples, twelve (12/30) were positive.
Arthropod samples consisted of 46 pools of Culicidae mosquitoes belonging to the collection of IZSPB and previously collected in Southern Italy during a research project (R.C. IZSPB 005/2010) funded by Italian Minister of Health. The samples were also screened for the presence of filarial nematodes, and 6 of them were found positive [32]. An additional negative control, consisting of a pool of 10 Dirofilaria-free Aedes albopictus mosquitoes from IZSPB laboratory colony, was also included.
All samples (genomic DNA, blood and arthropod samples) were tested by performing both the ReT-LAMP and PI-LAMP protocols.
In order to definitively confirm the presence or absence of D. repens, the blood and mosquitoes samples were also tested by using the COI-targeting PCR [29]. The yielded amplicons were sequenced by the BigDye Terminator v3.1 (Applied Biosystems, Milan, Italy), and the nucleotide sequences were compared with those from GenBank by BLAST.
A COI-targeting PCR [29] was performed by using DNA from D. repens larvae as template. Larvae were part of the collection of IZSPB, and they were previously isolated from the blood of an infected dog. The yielded 689-bp amplicon, which included the LAMP-targeted region, was purified by the mean of the QIAquick PCR purification kit, cloned in pGEM T-easy vector (Promega, Milan, Italy) according to the manufacturer's instructions, and then transformed in Escherichia coli MACH1.
The recombinant plasmid was extracted from a positive clone by using the PureLink HiPure plasmid miniprep kit (Thermo Scientific, Milan, Italy). Following quantification by the mean of a UV spectrophotometer, ten-fold serial dilution of plasmid DNA were prepared, from an initial concentration of 30 ng/μL, corresponding to about 8,5x109 copies/μL, up to a concentration of 30 ag/μL (8–9 copies/μL).
Five μL of each dilution was used as template for ReT-LAMP, PI-LAMP and PCR assays.
The blood samples from stray dogs (without any known owner) were collected by professional veterinarians without causing injury or any kind of consequence to the animals. The blood collection was performed as a routine, according to the local rules, upon admittance to the dog kennel, to assess or exclude the presence of infectious diseases. No animal was sacrificed or euthanatized for the aims of this study.
When tested in silico, all but one tested genes harbored potential annealing regions for no more than three out of the six oligonucleotides which constituted the LAMP primer set. In no cases the sequences of DiRF2c and DiRB2c, which prime the initial amplification step, matched together within the COI of the same species (Table 2). The only exception was represented by the COI of W. bancrofti, which was found to harbor the complementary motif of four oligonucleotides, including DiRF2c and DiRB2c.
However, when it was tested in vitro, no amplification was showed. Equally, no amplification was registered when DNAs from samples of D. immits, Ac. reconditum, An. vasorum, Brugia sp., L. loa, M. perstans or Cercopithifilaria sp. were used as template (able 2). On the contrary, all the ReT:LAMP reactions with DNA of D. repens returned the expected amplification curve (Fig 2a and 2b), and all the PI-LAMP mixtures were fluorescent (Fig 3).
Similarly, when the LAMP assays were performed with DNA from blood and mosquitoes samples, all the samples that were previously found positive to Onchocercidae resulted positive to ReT-(Fig 2c) and PI-LAMP (Fig 3), while no amplification was observed from the other samples.
The positive samples, when tested by the mean of the COI-targeting PCR, returned the expected 689-bp amplicon. The nucleotide sequences of amplicons were 98–100% identical to those in GenBank from D. repens.
The minimum amount of template necessary to obtain an amplification profile by the ReT LAMP was 0.15 fg, corresponding to about 50 copies of target (Fig 4a), while an amplicon was yielded by PCR (Fig 4b) from a minimum starting amount of 15 fg (about 5,000 copies of target).
Conversely, the PI-LAMP showed a detection limit of 10 fg (about 3,000 copies of target (Fig 4c).
The broad diffusion of the dirofilariasis due to D. repens stresses the needing of suitable and affordable diagnostic tools, which might promptly detect the parasite in hosts or in vectors.
The presented results showed that LAMP may be a suitable approach for the laboratory confirmation of the D. repens infection. Currently, the diagnosis of dirofilariasis, due to D. repens or D. immitis, relies on the microscopic and morphological identification of the parasite, usually microfilariae isolated from the blood of infected hosts [33]. Other methods are based on serological screenings [34–35] or on the detection of DNA of the nematode. In particular, PCR-based strategies are currently described to amplify D. repens DNA from mammals [36–40] or mosquitoes samples [11, 31, 41]. Among the diagnostic molecular methods, the LAMP is a promising system, and nowadays, it is applied for the diagnosis of fungal, viral, bacterial and parasitic infections [28, 42–43], including zoonotic filarial nematodes, such as the canine heartworm D. immitis, [44], W. bancrofti [45], the human lymphatic filarial nematodes Brugia malayi and B. timori [46]. The here described assay represents the first LAMP protocol for the detection of D. repens DNA. It results highly sensitive, as it returns a detectable amplification product from 0.15 fg of template, corresponding to about 50 copies of target, when performed as ReT-LAMP. While lower than ReT-LAMP, the sensitivity of PI-LAMP remains high, closely consistent with the traditional PCR. The specificity remains high, as well. In fact, no false positive result was obtained, while all samples that resulted positive by bother methods were also ReT-and PI-LAMP positive, thus confirming the cogency of the method.
Both versions of LAMP can be performed in about 40 minutes, with a faster outcome than PCR, which takes about 2 hours. Additionally, the here described LAMP protocols are also faster than the previously described qPCR assays [31, 41], which takes about 100–150 minutes. Furthermore, the latter were designed and developed to target a wider range of organisms (both D. repens and D. immitis [41], or a large group of filarial nematodes [31]). Therefore, the species identification must rely on the analysis of the melting curve or the nucleotide sequence, and this make those methods more demanding in terms of time and work.
Furthermore, the amplification may be immediately visualized, without the need for an agarose gel electrophoresis. Finally, the results can be unequivocally interpreted, as the amplification curve (for ReT-LAMP) and fluorescence (for PI-LAMP) were clearly visible from positive samples, while no aspecific signal was observed from negative samples. In addition, the here described LAMP assays appear to be effective on different matrices: the reactions carried out from blood and mosquito specimens returned clear amplification signals when the parasite was present in the sample, while no kind of signal was found from those negative.
In the light if those consideration, the described method represents, to our knowledge, the first LAMP-based tool to detect D. repens directly from biological samples.
This makes the assay effective for the detection of D. repens in hosts with microfilaraemia, and it may reliably support the differential diagnosis with D. immitis, Ac. reconditum or Mansonella spp.. Therefore, a potential application of the method may be the screening of traveling live animals, which are often responsible for the introduction of the parasite in unaffected areas. Furthermore, the PI-LAMP, being a rapid and cost-effective assay, could be an useful and ancillary tool for screening a large number of culicid mosquitoes and to assess their positivity for D. repens during entomological surveys in endemic, or even in non-endemic areas. Finally, the PI-LAMP protocol can also be performed with very simple equipment, such as a heating device instead of the more expensive real time PCR instrument. This version has been found to be slightly less sensible if compared with the ReT-LAMP protocol, but it did not show any impairment in the specificity of the assay. This possibility could make the method suitable for application in field, especially in developing Countries, where expensive equipment and specialized personnel may often lack.
|
10.1371/journal.pgen.0030079 | A Mutation in the Myostatin Gene Increases Muscle Mass and Enhances Racing Performance in Heterozygote Dogs | Double muscling is a trait previously described in several mammalian species including cattle and sheep and is caused by mutations in the myostatin (MSTN) gene (previously referred to as GDF8). Here we describe a new mutation in MSTN found in the whippet dog breed that results in a double-muscled phenotype known as the “bully” whippet. Individuals with this phenotype carry two copies of a two-base-pair deletion in the third exon of MSTN leading to a premature stop codon at amino acid 313. Individuals carrying only one copy of the mutation are, on average, more muscular than wild-type individuals (p = 7.43 × 10−6; Kruskal-Wallis Test) and are significantly faster than individuals carrying the wild-type genotype in competitive racing events (Kendall's nonparametric measure, τ = 0.3619; p ≈ 0.00028). These results highlight the utility of performance-enhancing polymorphisms, marking the first time a mutation in MSTN has been quantitatively linked to increased athletic performance.
| An individual's genetic profile can play a role in defining their natural skills and talents. The canine species presents an excellent system in which to find such associative genes. The purebred dog has a long history of selective breeding, which has produced specific breeds of extraordinary strength, intelligence, and speed. We have discovered a mutation in the canine myostatin gene, a negative regulator of muscle mass, which affects muscle composition, and hence racing speed, in whippets. Dogs that possess a single copy of this mutation are more muscled than normal and are among the fastest dogs in competitive racing events. However, dogs with two copies of the same mutation are grossly overmuscled, superficially resembling double-muscled cattle known to possess similar mutations. This result is the first to quantitatively link a mutation in the myostatin gene to athletic performance. Further, it emphasizes what is sure to be a growing area of research for performance-enhancing polymorphisms in competitive athletics. Future implications include screening for myostatin mutations among elite athletes. However, as little is known about the health issues and potential risks associated with being a myostatin-mutation carrier, research in this arena should proceed with extreme caution.
| The wide variety of behaviors and morphological types exhibited among dog breeds and the overall low genetic diversity within each breed make the dog an excellent genetic system for mapping traits of interest [1,2]. Recently, owners of whippets, an established racing-dog breed, have reported a phenotype of heavy muscling occurring within the breed (http://www.k9community.co.uk/forums/index.php). The typical whippet is similar in conformation to the greyhound, a medium-sized sighthound, weighing about 9 kg and characterized by a slim build, long neck, small head, and pointed snout (Figure 1A) [3]. Heavily muscled dogs, termed “bully” whippets by breeders, have broad chests and unusually well-developed leg and neck musculature (Figure 1C). “Bully” whippets are easily distinguished from their normal littermates based on physical appearance alone (compare Figure 1A and 1C). Owners report that “bully” whippets do not have any health abnormalities other than muscle cramping in the shoulder and thigh. However, the dogs are often euthanized at an early age as they do not conform to the American Kennel Club breed standard. In addition, about 50% of “bully” whippets have a distinctive overbite.
The “bully” whippet phenotype is reminiscent of the double muscling phenotype seen in other species that is caused by mutations in the myostatin (MSTN) gene. Such variants have been observed in mice [4], cattle [5,6], sheep [7], and human, the latter described once in a German boy [8].
The myostatin protein has been shown to affect both the amount and composition of muscle fibers. For instance, the muscle mass of Mstn knockout mice is two to three times greater than that of wild-type mice [9]. Furthermore, the sequence of the MSTN gene is relatively conserved across species [9]. Therefore, we chose to interrogate the MSTN gene for possible mutations resulting in the “bully” whippet phenotype.
We sequenced the three exons and the majority of introns of the MSTN gene in an initial set of 22 whippets. A 2-bp deletion was discovered in the third exon of the MSTN gene (Figure 2). This deletion removes nucleotides 939 and 940 within exon three and leads to a premature stop codon at amino acid 313 instead of the normal cysteine, removing 63 aa from the predicted 375-aa protein. The lost cysteine is one of several highly conserved cysteines known to form disulfide dimers required for protein function [9].
Of the 22 whippets sequenced, all “bully” whippets tested (n = 4) were homozygous for the deletion (mh/mh) while all dogs that sired or whelped a “bully” whippet (n = 5) were heterozygous for the deletion (mh/+). None of the initial set of 13 normal-appearing whippets that lacked a family history of the “bully” phenotype carried the deletion; these dogs were designated wild type (+/+). An additional set of DNA samples from 146 whippets (both racers and nonracers) were collected at racing events and through the mail without regard to the dogs' family histories of the “bully” phenotype. These were sequenced across exon three to determine the frequency of the 2-bp mutation among the dogs sampled. Of these, two were homozygous for the deletion, 20 were heterozygous, and the remaining 124 did not carry the deletion.
The “bully” phenotype displays a simple autosomal recessive mode of inheritance, as all “bullies” resulted from the mating of carriers. The parents have a phenotype of intermediate musculature (Figure 1B). In order to quantify the allelic substitution and dominance effects of the deletion mutation we considered three measures of musculature: mass-to-height ratio, neck girth, and chest girth. For all three measures, heterozygous females (mh/+) were intermediate in musculature, mh/mh females had the highest measures, and female +/+ whippets had the lowest measures. Male mh/+ whippets were more muscular than wild-type males (Figure 3 and Table 1).
Mass (kg) and height at the withers (cm) information was available for 71 female and 55 male whippets; we analyzed the two sexes separately due to a lack of samples from male mh/mh whippets. Two analyses suggest strong statistical support for the idea that the mh deletion mutation affects the mass-to-height ratio. First, a nonparametric Kruskal-Wallis one factor Analysis of Variance for females using genotype as a factor (mh/mh, mh/+, and +/+) was highly significant (p ≈ 7.43 × 10−6) [10]. Likewise, a two-sample Wilcoxon rank sum test comparing mass-to-height ratio of mh/+ and +/+ males was also significant (p ≈ 0.00017), with whippets heterozygous for the mutation exhibiting, on average, 17% more mass per cm of height (i.e., 0.333 kg/cm for mh/+ as compared to 0.284 kg/cm for +/+ genotypes). Second, we used standard least-squares regression to estimate the allelic substitution (a = 0.0817; p < 2.44 × 10−15) and dominance effects (d = −0.0423; p = 0.00415) in female whippets and found both parameters were significantly different than zero. These results, along with box plots of the phenotype by genotype class (Figure 3), suggest that the mutation is partially recessive with heterozygotes (mh/+) having musculature closer to, but significantly different from, that of wild types (+/+). As reported in Table 1, we found that the mh mutation also affected neck and chest girth in a similar fashion in females with a highly significant effect of MSTN genotype on phenotype. Overall, we estimate that mh explains approximately 60% of the variation in both mass-to-height ratio and neck girth (i.e., r2 = 60%) and 31% of the variation in chest girth (Table 1). In male whippets, we also observed a highly significant difference in neck girth (p ≈ 0.0013, Wilcoxon rank sum test) and nearly significant difference in chest girth (p ≈ 0.11, Wilcoxon rank sum test) among wild type (+/+) and heterozygotes (mh/+).
We hypothesized that the increased muscle mass of the heterozygotes would allow for increased speed when compared to wild type (+/+) whippets. Analysis of 85 genotyped dogs for which we obtained racing grades revealed an association between a dog's genotype and racing grade using two separate (but not independent) approaches (Figure 4). Only one mh/mh dog competed in racing events and it was a grade-A racer, so we included this dog with the heterozygotes and considered the absence or presence of the deletion for all analyses (i.e., we sum across the mh/+ and mh/mh columns in Table 2. First, we find a significant positive correlation between racing grade (A, B, C, and D in order from fastest to slowest) and the frequency of dogs carrying the mh deletion in either homozygous or heterozygous state (p ≈ 0.00028, τ = 0.3619, Kendall's nonparametric measure). Secondly, standard contingency table analysis reveals strong evidence for heterogeneity in the frequency of dogs carrying the mh deletion among racing grade classes (Table 2; p = 0.00029, Fisher's exact test with degrees of freedom = 3). Since the 4 × 2 contingency table (i.e., combining columns two and three of Table 2) has three degrees of freedom, it is possible to partition the analysis into three tests of one degree of freedom each in order to identify outlier grades. There are several methods by which this can be accomplished, although they are not independent of each other. For example, 12 of 41 dogs in the fastest two racing grades, A and B, carried the deletion, while just one dog of 43 from the slowest two racing grades, C and D, was a heterozygote (p = 0.00073, Fisher's exact test), indicating a strong difference in frequency between (A, B) and (C, D). There is suggestive evidence for a difference between race grades A and B (p = 0.086, Fisher's exact test) but no evidence for a difference between C and D (p = 0.42, Fisher's exact test). A different approach for dividing the degrees of freedom is to compare A versus (B, C, D) (p = 0.00027, Fisher's exact test), C versus (B, D) (p = 0.099, Fisher's exact test), and C versus D (p = 0.42, Fisher's exact test). Thus, each of these two methods for partitioning the test suggests that the presence of the mh mutation strongly influences racing ability.
There was also a marginally significant difference in the mutation genotype frequency in whippets that participated in racing versus conformation events such as breed club regional specialties, where dogs are judged based on their conformation to the physical breed standard (height at the withers or shoulder, head shape, coat color, etc.). Twenty of 119 confirmed racing whippets were heterozygous for the deletion while just two of 42 whippets that competed in conformation events were heterozygous (p = 0.038, one-tailed Fisher's exact test).
To investigate how population substructure within whippets affects our candidate gene analysis, we genotyped 32 unlinked microsatellite markers [11] in 84 whippets with racing grades. Analysis of FST revealed that whippets exhibit a low, but measurable, degree of population differentiation with higher levels of interbreeding within racing grades. As a result, genetic distance correlates with racing grade so that there is a moderate differentiation between A and B grades (FST = 0.021), little differentiation between B and C (FST = 0.0044) or C and D (FST = 0.0067) racers, and the largest difference between the population of A and D racing dogs (FST = 0.041). These levels of population differentiation are quite typical for mammalian species and are not surprising, given that whippets are bred to race and positive assortative mating is expected (i.e., fast dogs are bred to fast dogs).
Analysis of the data using Structure [12] and InStruct [13] gave comparable results under a range of cluster numbers from K = 1 to K = 15. Neither program found a clustering that clearly correlated with racing grade. The neighbor-joining tree of the 84 dogs used in this analysis based on genetic similarity (i.e., kinship coefficient) tends to differentiate dogs within the A racing grade from those in other grades, but not in a fully exclusive manner (Figure S1). It also demonstrates that dogs carrying the 2-bp deletion are found on all branches of the tree and are not one another's closest relatives, although they do tend to cluster near one another on the tree.
Logistic regression analysis of 73 alleles at moderate frequencies in the sample confirmed that population substructure may be a potentially confounding effect for association mapping within breeds of dog. In particular, we found that 26 of the alleles (36%) across the 32 loci show a nominal p-value of 5% or lower, a huge inflation above the expected 5% (Figure S2A). As a result, many of the 32 loci showed association with racing grade using standard contingency table analysis. However, the association of the 2-bp deletion with racing grade is more significant than all but one of these, suggesting an empirical p-value of 1/72 = 0.014 (Figure S2B).
We sequenced exon three of MSTN in a set of approximately four dogs each from heavily muscled breeds including the bullmastiff, rottweiler, bulldog, Presa Canario, miniature bull terrier, American Staffordshire terrier, and Staffordshire bull terrier. In addition, we sequenced exon three of MSTN in a small set of dogs from breeds known to compete in racing: the greyhound, lurcher, and two mixed-breed dogs (whippet crosses). None of these dogs possessed the 2-bp deletion seen in the whippet.
To obtain haplotype information, we sequenced the MSTN gene and surrouonding area (Figure 5) in one to ten dogs each from the greyhound, Ibizan hound, Pharaoh hound, Afghan hound, saluki, Italian greyhound, mastiff, boxer, Akita, Basenji, Australian shepherd, beagle, German wirehaired pointer, and flat-coated retriever breeds. A golden jackal was also sequenced to determine the ancestral allele for each dog SNP and insertion/deletion (indel). None of the dogs sequenced from any of the above breeds nor the golden jackal possessed the 2-bp deletion. We discovered 28 SNPs and three indels. For the three indels and eight SNPs that were discovered within 50 kb of the MSTN mutation, we inferred haplotypes independently for each breed using PHASE. A total of 13 haplotypes were identified, but only six were observed in more than one breed. Two of these haplotypes are shared across several breeds. The MSTN mutation occurs on just one haplotype that is observed only in the whippet (Figure 5). A haplotype observed in 12 breeds differs from this one only at the mutation itself.
The whippet breed was developed in the late 1800s specifically for the sport of racing [3]. Despite its comparatively small stature it is a very fast dog capable of running up to 35 miles per hour [3]. We have discovered a 2-bp deletion in the whippet MSTN gene that in the homozygote state results in a double-muscling phenotype commonly referred to as the “bully” whippet. This deletion causes a premature truncation of the protein at amino acid 313, removing the latter 17% of the protein.
MSTN has been mapped to canine Chromosome 37 (CFA37) and consists of three exons spanning 5083 bp (http://genome.ucsc.edu). It is highly conserved across species [9] and in the human genome is located on Chromosome two. The gene is a member of the transforming growth factor β family and encodes the myostatin protein. Studies of Mstn knockout mice demonstrate that the gene is a negative regulator of skeletal muscle mass [9]. This is the result of a cascading pathway triggered by MSTN signaling that prevents myoblast cell progression from the G1 to S phase of the cell cycle. MSTN therefore controls the total number of muscle fibers by regulating overall myoblast proliferation [14]. In the absence of functional protein, greater numbers of muscle fibers are made [14].
Double muscling has been described in several breeds of cattle [5,6,15–17] and muscular hypertrophy (an increase in muscle-fiber size) has been described in sheep [7]. Muscular hypertrophy has also been described in the domestic cat [18]; however, a deficiency in dystrophin is the cause in this species rather than a mutation in MSTN. To date, six different mutations in the bovine MSTN gene have been reported to cause double muscling [5,6,15–17]. While all these mutations result in a loss of MSTN function, a subset of them and the one we describe here in the whippet likely change the three-dimensional shape of the protein by disrupting the “cysteine knot,” a structure important in the folding of all transforming growth factor β family member proteins [5,9]. The mutation in the whippet also removes nearly 20% of the protein.
We sequenced genomic DNA not only from whippets but also from multiple dogs from each of 14 additional breeds in order to determine the haplotype background on which this mutation arose (Figure 5). In each dog, 15 PCR amplicons that spanned the MSTN gene and amplicons spanning known dog SNPs within 50 kb of the MSTN mutation were sequenced. Using the resulting data we observed two haplotypes, termed two and seven, that occurred in a large number of breeds and that were identical except at position 3,676,629, which is located outside the gene, 15,801 bp downstream from the 2-bp deletion. The mh mutation occurs only on haplotype six, which is identical to haplotype seven except for the deletion itself. Not surprisingly, the golden jackal sequence has only the wild-type allele at the position of the mutation, indicating the mh allele represents the derived state. We conclude therefore that haplotype six likely derives from haplotype seven (Figure 5). Haplotype seven is the most common and widely dispersed haplotype spanning the gene and was found in 12 out of 15 breeds sequenced. Interestingly, we did not observe haplotype seven in the Afghan hound, Basenji, or boxer.
Our data do not exclude the possibility that the mutation occurs in breeds other than the whippet. However, we screened for the 2-bp deletion in several mastiff type breeds (rottweiler, bulldog, Presa Canario, miniature bull terrier, American Staffordshire terrier, Staffordshire bull terrier, and bullmastiff) and did not find it. These data argue that the changes in musculature exhibited by the whippet are unique and caused by the effects on MSTN associated with the deletion described in this study.
An excess of the mh/+ genotype was observed among the fastest racers, as defined by the highest racing grade achieved during a dog's career. This demonstrates that the heterozygote state carries a performance-enhancing polymorphism that provides a competitive edge. The optimal study of racing performance would use the racing points acquired by each whippet during their career as a quantitative measure of performance. However a dog's total career points are a function of the number of races run throughout their career and, as such, whippets of different ages are not easily compared. To compensate, the total number of points accrued over a lifetime of racing could be averaged over the number of races entered. However, as dogs age their performance declines. Some owners stop racing their dogs after their performance declines while others continue to race their dogs for months or even years longer. Using the average number of points accrued during a specific year of the dog's life, for instance age two or three, presents similar problems. Dogs reach their racing prime at different ages and the number of points will always reflect the number of races entered. While an average is satisfactory if many races are run in a given year, the average will be inaccurate if few races are run.
While cattle breeders have long selected for individuals that are homozygous for mutations in MSTN because of their increased musculature, which is optimal for beef production, this is the first example of breeders unknowingly selecting for individuals with a single polymorphism that increases athletic performance. Of interest, the trait appears to confer an undesirable appearance upon dogs competing in conformation. Only two mh/+ dogs were found among the dogs reported to compete in conformation events, and those dogs were reported to show poorly. This is consistent with the association seen between a dog's genotype and their relative muscle mass as defined by either a ratio of mass (kg) to height at the withers (cm) (p = 7.43 × 10−6; Kruskal-Wallis test) (Figure 3A) or the direct measure of an individual's neck girth (p = 3.47 × 10−5; Kruskal-Wallis test) (Figure 3B) or chest girth (p = 0.001462; Kruskal-Wallis test) (Figure 3C). We acknowledge that there are more accurate methods to measure muscle mass. However, many of these methods are either invasive, such as a muscle biopsy, or would need to be conducted post-mortem, neither of which was an option. These measurements were not designed to specifically eliminate contributions from body fat. However, obesity is rare in the whippet; indeed, the breed is characterized by an overall low body fat content. Thus, these measurements are the best achievable metrics of the phenotype.
Greyhounds and whippets share a common ancestral gene pool and as a result the breeds are difficult to separate in genetic clustering analyses [11]. This, together with the fact that both were bred to excel at racing, suggested that the mutation might also be found in racing greyhounds. However, none of the greyhounds tested carried the mutation. There are three possible explanations for this result. First, an insufficient number of samples have been tested if the mutant allele is relatively rare in the greyhound population. Second, the mutation may only be present in a subset of greyhound lines, none of which were among those tested. Finally, the mutation may not be carried in the greyhound population at all, indicating that it is a relatively new mutation in the purebred dog population. This may be because the mutation offers no advantage to greyhound racers. Indeed, it may even be disadvantageous. Studies of muscle composition in Mstn knockout mice demonstrate a higher proportion of both fast type II and glycolytic fibers, versus slow type I and oxidative fibers when compared to wild-type mice [19]. While this change in muscle composition may offer an advantage to whippets, which typically race a short sprint of 200–300 m, it may be disadvantageous to greyhounds, whose races extend to 900 m and where endurance is more important. In addition, Belgian Blue cattle that are homozygous for a MSTN mutation display a decrease in the size of several organs, including the lungs [20]. If heterozygous dogs have even a slightly reduced lung capacity, it is possible that a MSTN mutation would actually be disadvantageous for racing longer distances as greyhounds do. Finally, it remains to be determined whether additional health problems are associated with being a carrier of this mutation.
We examined the microsatellite data set for evidence of population substructure and found that there is not random gene flow across the racing classes. All groups display positive FST values with the greatest found between the grade A racers and all others. This is not unexpected. The very presence of the “bully” phenotype is evidence that breeders choose to mate dogs with increased musculature to one another. Reducing the mating population of a breed to a small proportion of the whole population has consequences, particularly for genetic mapping of complex traits. This is evidenced by our analysis of the same marker set for association with racing grade. While we find low p-values at many of the alleles, only one of 73 had a p-value smaller than the 2-bp deletion, confirming that the association between racing grade and the MSTN mutation is not simply a spurious result of population structure. Overall, these results suggest that the population structure within breeds is likely to have an important confounding effect on association mapping in the domestic dog.
Our findings have implications for competitive and professional sports. Here, we show that a disruption in the function of the MSTN gene increases an individual's overall athletic performance in a robust and measurable way. To date, the muscular hypertrophy phenotype has been described in a single human child [8]. This child possessed two copies of a G-to-A transition in the noncoding region of the human MSTN gene. This mutation results in the mis-splicing of precursor mRNA, which most likely truncates the myostatin protein. The child's mother, a former professional athlete, was heterozygous for this mutation and also appeared muscular, although not to the same degree as her child. Perhaps additional mutations in MSTN have yet to be discovered in other species that competitively race, such as the horse or humans. As discussed by others [21,22], human athletes could undergo so-called gene doping via disruption of MSTN. The potential to increase an athlete's performance by disrupting MSTN either by natural or perhaps artificial means could change the face of competitive human and canine athletics. Given the poorly understood consequences for overall health and well-being, caution should be exercised when acting upon these results.
An initial set of 22 whole-blood samples were collected from whippets that participated in racing, conformation, or were simply privately owned pets. Of these 22 dogs, four were reported by owners to be “bullies,” five dogs had either sired or whelped a “bully,” and the owners of the remaining 13 stated there was no known family history of “bullies” in their dog's pedigree. After initial analysis of these 22 samples an additional 46 whole-blood samples and 100 buccal swabs were collected from a mixture of racing, conformation, and pet whippets. No restrictions were placed on age, gender, or relatedness of the dogs sampled. The dog's sex was recorded for 165 (98%) of the dogs (74 males and 91 females). Samples were collected both by mail and at sanctioned Whippet Racing Association (WRA), National Oval Track Racing Association (NOTRA), and Continental Whippet Alliance (CWA) racing events. Blood samples from each dog were collected as whole blood in ACD tubes. Buccal swabs were collected using standard protocols with Cytosoft cytology brushes (Medical Packaging Corporation, http://www.medicalpackaging.com). DNA was extracted from the brushes using a QIAamp Blood Mini kit (Qiagen, http://www.qiagen.com) following the manufacturer's protocol. DNA was extracted from the blood samples using a standard phenol/chloroform extraction method [23].
DNA samples were also collected from 33 greyhounds, two mixed-breed dogs (whippet crosses), seven lurchers, five each of bullmastiffs, rottweilers, and bulldogs, four each of Presa Canarios, miniature bull terriers, mastiffs, Staffordshire bull terriers, Ibizan hounds, and salukis, three each of American Staffordshire terriers, Italian greyhounds, boxers, and Pharaoh hounds, two each of Akitas, Afghan hounds, Australian shepherds, beagles, flat-coated retrievers, and German wirehaired pointers, a single Basenji, and a single golden jackal. Samples were either received as DNA from collaborators or DNA was extracted by the aforementioned methods after collection at dog shows, events, or provided directly by owners and breeders. Informed consent was obtained for all newly collected samples and all protocols were approved by the Animal Care and Use Committee of the Intramural Program of the National Human Genome Research Institute at the National Institutes of Health.
Owners of whippets were asked to provide detailed information about their dogs including American Kennel Club or other registration number, pedigree information, and the events in which their dogs participated. Front and side photos of individual dogs were obtained for comparison. The dog's mass and height at the withers were recorded for 126 dogs (55 males and 71 females). A set of body measurements including neck girth and chest girth were also collected from 137 of the whippets (61 male and 76 female) either by owners or laboratory members as described (N. B. Sutter, D. S. Mosher, and E. A. Ostrander; personal communication).
There are four organizations that govern whippet racing in the United States; straight racing is sponsored by the WRA, CWA, and the North American Whippet Racing Association (NAWRA), while oval racing is sponsored by NOTRA. The standard track length of a straight race is 182.88 m, while the standard oval track is 350 m. A race meet is composed of four heats. Based on their placement in all four heats, dogs are given a total meet score. A dog's racing grade is a simplified assessment of its performance over the last three meets. As a result, a dog's racing grade will vary throughout its career as grade is reassessed following each meet. We used the WRA racing grades: Grade A is 15.0–29.0 points, grade B is 10.0–14.9 points, grade C is 4.0–9.9 points, and grade D is 0–3.9 points (http://www.whippetracing.org/Rules/2006/2006Chapter5.htm). We categorized dogs based on the highest grade ever achieved. Grades for 85 racing whippets were obtained from the WRA website (http://www.whippetracing.org).
We employed least-squares regression to estimate the allelic substitution (a) and dominance (d) effects of the mutation using a Falconer parameterization for genotypic means [24]. Briefly, we assumed the phenotypes within genotypic classes are normally distributed with mean for wild type (+/+), heterozygotes (mh/+), and homozygotes (mh/mh) of μ+/+ = μ − a, μmh/+ = μ + d, and μmh/mh = μ + a, and with common variance σ2 within classes. This amounts to using two “dummy” variables to encode the design matrix for the regression with values (−1, 0, and 1) and (0, 1, and 0) for the genotype classes (+/+, +/mh, and mh/mh). All measurements were normalized (i.e., mean subtracted and observations divided by the standard deviation) and Q-Q plots inspected visually to assess the appropriateness of normality assumption within genotypic classes. All statistical analyses were carried out using R 2.4.1 (www.r-project.org).
We genotyped 32 unlinked microsatellite markers [11] in 84 whippets with racing grades. All loci were found to be variable with a total of 135 alleles segregating across all markers. We used GDA [25] to calculate Wright's fixation index (FST) and estimated kinship coefficients using Ritland's method [26] and Rousset's genetic distance [27] using Spagedi 1.2f [28]. Population clustering was assayed using the Bayesian clustering algorithm Structure [12], which assumes Hardy-Weinberg equilibrium within clusters and InStruct [13], a Structure-like algorithm that estimates a generalized inbreeding coefficient for each cluster. A neighbor-joining tree was constructed by using Rousset's genetic distance [27] as input into Phylip 3.66 (http://evolution.genetics.washington.edu/phylip.html) and racing grade state was traced across the tree using MacClade 4.0 (http://macclade.org/) (Figure S1).
We analyzed the control loci for association between genotype and racing grade using standard logistic regression analysis for differentiating A racers from B, C, or D racers. For each control locus, we fit a saturated model with nl parameters, where nl is the number of microsatellite alleles found at locus l (each analysis had (nl − 1) allele effects and one intercept parameter). Figure S2 shows the distribution of p-values for the 73 allele effects with p-values below 0.98. In total there were (135 − 32) = 103 independently estimated allele effects, but for 30 of these there was little power to detect an effect because so few dogs carried the allele.
The entire canine MSTN gene was sequenced except for a 1,039-bp GC-rich region in intron one. Twelve pairs of overlapping primers covering the remaining regions of the gene and three primer pairs for SNP genotyping were designed using Primer 3 software [29] (http://frodo.wi.mit.edu/cgi-bin/primer3/primer3_www.cgi). Exon/intron boun-daries were based on the complete canine MSTN mRNA sequence. The resulting amplicons averaged 700 bp in length (653 to 799 bp). PCR was performed in a total volume of 10 μl containing 10 ng of dog DNA, 1× reaction buffer (Applied Biosystems, http://www.appliedbiosystems.com), 0.1 mM dNTP (Promega, http://www.promega.com), 1.5 mM MgCl2, 0.5 U of AmpliTaq Gold polymerase (Applied Biosystems) and 0.03 μM each specific primer. Touchdown PCR was carried out as follows: 7 min at 95 °C followed by 20 cycles of 30 s at 95 °C, 30 s at annealing temperature (beginning at 61 °C and decreasing 0.5 °C per cycle) and 30 s at 72 °C then 20 cycles of 30 s at 95 °C, 30 s at 51 °C and 30 s at 72 °C and a final extension phase for 3 min at 72 °C.
The resulting PCR products were sequenced using Big Dye version 3.1 on an ABI 3730xl capillary sequencer (Applied Biosystems). Sequence reads were aligned and analyzed using Phred, Phrap, and Consed [30,31,32]. Polyphred [33] was used to assist in the identification of all SNPs and indel polymorphisms.
The Genbank (http://www.ncbi.nlm.nih.gov/gquery/gquery.fcgi) accession number for the complete canine MSTN mRNA sequence is AY367768. The Genbank accession number for the MSTN protein is NP_01002959. |
10.1371/journal.pbio.2003782 | Identifying off-target effects of etomoxir reveals that carnitine palmitoyltransferase I is essential for cancer cell proliferation independent of β-oxidation | It has been suggested that some cancer cells rely upon fatty acid oxidation (FAO) for energy. Here we show that when FAO was reduced approximately 90% by pharmacological inhibition of carnitine palmitoyltransferase I (CPT1) with low concentrations of etomoxir, the proliferation rate of various cancer cells was unaffected. Efforts to pharmacologically inhibit FAO more than 90% revealed that high concentrations of etomoxir (200 μM) have an off-target effect of inhibiting complex I of the electron transport chain. Surprisingly, however, when FAO was reduced further by genetic knockdown of CPT1, the proliferation rate of these same cells decreased nearly 2-fold and could not be restored by acetate or octanoic acid supplementation. Moreover, CPT1 knockdowns had altered mitochondrial morphology and impaired mitochondrial coupling, whereas cells in which CPT1 had been approximately 90% inhibited by etomoxir did not. Lipidomic profiling of mitochondria isolated from CPT1 knockdowns showed depleted concentrations of complex structural and signaling lipids. Additionally, expression of a catalytically dead CPT1 in CPT1 knockdowns did not restore mitochondrial coupling. Taken together, these results suggest that transport of at least some long-chain fatty acids into the mitochondria by CPT1 may be required for anabolic processes that support healthy mitochondrial function and cancer cell proliferation independent of FAO.
| Oxidation of long-chain fatty acids inside of the mitochondrial matrix provides an essential source of energy for some cells. Since long-chain fatty acids cannot freely pass into the mitochondrial matrix, they rely on a protein called carnitine palmitoyltransferase I (CPT1) for transport. Prior research has found that many tumors exhibit increased expression of CPT1 and/or sensitivity to CPT1 inhibition by a drug called etomoxir. These findings have led to thinking that cancer cells rely on fatty acid oxidation for energy. Here we present data that indicate otherwise, showing that inactivation of fatty acid oxidation has no effect on the proliferation of at least some cancer cell lines. Instead, these cells alter their utilization of other nutrients (such as glutamine) to compensate for the loss of fatty acid oxidation. We describe 2 discoveries that provide new insight into the role of fatty acid oxidation in cancer and help rationalize previous results. First, etomoxir has the off-target effect of inhibiting complex I of the electron transport chain. Second, CPT1 has other cellular functions that are independent of fatty acid oxidation. We suggest that one such function may be importing long-chain fatty acids into the mitochondria for anabolic fates, rather than catabolic oxidation.
| During the last decade, carnitine palmitoyltransferase I (CPT1) has been identified as a potential therapeutic target for a growing list of cancers that include breast cancer, prostate cancer, glioblastoma, colon cancer, gastric cancer, myeloma, and others [1–6]. In these cancers, CPT1 expression is increased, and/or CPT1 inhibition is reported to have antitumor effects. CPT1 is an enzyme associated with the outer mitochondrial membrane that transfers a long chain acyl group from coenzyme A to carnitine [7, 8]. Importantly, this transformation is required to transport long-chain fatty acids into the mitochondrial matrix. Long-chain fatty acids reaching the mitochondrial matrix are generally assumed to be oxidatively degraded, thereby implicating fatty acid oxidation (FAO) as a potentially important pathway in cancer metabolism [9].
FAO is thought to support cancer metabolism primarily in 2 ways. First, given their highly reduced state, fatty acids may provide an important source of ATP to fuel tumor growth [10]. For every pair of carbons in a fatty acid that is completely oxidized, up to 14 ATP can be produced—assuming NADH and FADH2 yield 2.5 and 1.5 ATP, respectively [11]. Ten of these 14 ATP are produced by oxidizing acetyl-CoA in the tricarboxylic acid (TCA) cycle. Oxidation of exogenous fatty acids might be particularly relevant to tumors that grow in adipocyte-rich environments, such as breast cancer [12]. Here, fatty acids transported from neighboring adipocytes may constitute an important energy reservoir [13]. A second potential benefit of cancer cells oxidizing fatty acids is the production of NADPH [14]. Although FAO does not make NADPH directly, the acetyl-CoA it produces in the mitochondria can be shuttled to the cytosol as citrate once acetyl-CoA condenses with oxaloacetate. Each molecule of citrate exported to the cytosol can then produce 1 molecule of NADPH through either isocitrate dehydrogenase 1 or malic enzyme 1. It has been suggested that some cancer cells rely on this source of NADPH to neutralize oxidative stress [9]. Indeed, inhibition of CPT1 in human glioblastoma cells causes a reduction in NADPH levels and an increase in reactive oxygen species [15].
A major challenge of considering FAO as an essential pathway in cancer metabolism is that cancer cells are also thought to rely heavily on fatty acid synthesis [16]. While one can rationalize the coexistence of FAO and fatty acid synthesis on the basis of subcellular compartmentalization, conventional thinking would indicate that it is unproductive to run both pathways simultaneously [9]. Additionally, recent data from our laboratory suggest that such a futile cycling process occurs to only a minimal extent in at least some proliferating cells [17].
As noted, the focus on FAO in cancer cells has mostly been driven by experimental findings related to CPT1 [6]. The assumption has been that increased CPT1 expression and sensitivity to CPT1 inhibition represents a demand for FAO. In this work, we consider an alternative possibility that CPT1 has important metabolic roles independent of FAO. We present evidence that long-chain fatty acids transported into the mitochondria via CPT1 have important anabolic fates that are essential for proliferation. We also provide data suggesting that etomoxir, a drug commonly used to inhibit CPT1 in cancer studies, has off-target effects that may complicate the interpretation of some experiments.
We focus much of our attention on the breast cancer cell line BT549, because the essential role of CPT1 in these cells has already been thoroughly demonstrated [18]. We show that inhibiting FAO by as much as 90% had no effect on BT549 cell proliferation. At this level of pharmacological CPT1 inhibition, minimal labeling from 13C-enriched fatty acids could be detected in citrate. These results suggest that BT549 cells do not require FAO as a major source of ATP or NADPH. When CPT1 is knocked down, however, we found that BT549 cell proliferation was significantly reduced. Under these conditions, the function of the mitochondria was impaired, and changes in the levels of complex lipids within mitochondria were detected. The cells could not be rescued by acetate or octanoic acid supplementation. These data support a role for CPT1 in the proliferation of some cancer cells that is independent of FAO.
All cells were cultured in high-glucose DMEM (Life Technologies) containing 10% FBS (Life Technologies) and 1% penicillin/streptomycin (Life Technologies) at 37 °C with 5% CO2. All culture media for growing cells were supplemented with 100 μM palmitate-BSA and 100 μM oleate-BSA to approach the physiological concentrations of free fatty acids. When counting cells manually, BT549 cell media were refreshed to control or experimental media 24 hours after the cells were seeded (at t = 0) to assess growth. At selected time points, cells were collected and counted in trypan blue with an automatic cell counter (Nexcelom). Doubling time was calculated by linear regression against the logarithm of cell density in exponential phase. For assessing proliferation, cells were grown under various experimental conditions for 48 hours, and proliferation was determined by using an MTT assay (ATCC) according to the manufacturer’s instructions. Absorbance was measured at 570 nm by using the Cytation 5 microplate reader (BioTek) with a reference wavelength set at 670 nm. We note that comparable changes in cell proliferation were measured using the MTT assay and manual cell counting when BT549 cells were treated with 200 μM etomoxir for 48 hours (S1 Fig), indicating that the 2 techniques to assess cell proliferation provided consistent results in our experiments. Etomoxir was purchased from Cayman Chemical (purity ≥ 98%). Etomoxir was dissolved in water to create a concentrated stock solution. The vehicle control was water alone.
CPT1A silencing was achieved by using a validated pool of small interfering RNA (siRNA) duplexes directed against human CPT1A (Trifekta Kit, IDT) and Lipofectamine RNAiMAX Transfection Reagent (Invitrogen) according to the manufacturer’s instructions (see S1 Text for the dicer-substrate short interfering RNA [DsiRNA] sequence) [19]. The knockdown (KD) efficiency was determined by measuring CPT1A mRNA levels with a premade primer (IDT) and quantitative RT-PCR (Applied Biosystems). The expression levels were normalized to an HPRT endogenous control. Cells given scrambled siRNA were used as a negative control. For overexpression of human CPT1A, the cDNA was cloned in the pcDNA3.1+ vector (GenScript) under a constitutive CMV promoter. The codon was optimized to be resistant to the siRNA added. The catalytically dead CPT1A had an identical sequence (see S1 Text) to the wild-type siRNA-resistant CPT1A, with the exception of G709E and G710E mutations to abolish catalytic activity (GenScript). For transduction, CPT1A was first knocked down with Lipofectamine RNAiMAX for 24 hours. Next, cells were transduced with plasmids using Lipofectamine 3000 (Invitrogen) for 4 hours. Media were then refreshed, and cells were assayed 48 hours post plasmid transduction (72 hours post siRNA knockdown). The control vector was pcDNA3.1+N-eGFP (GenScript), which expresses GFP instead of CPT1A.
BT549 cells were treated with either a scrambled siRNA control or siRNA targeting CPT1A for 12 hours. Next, nutrients were added to each culture plate and incubated for 48 hours before assessing cell proliferation with an MTT assay. Each compound (sodium acetate, octanoic acid, uridine, and sodium pyruvate) was added separately and evaluated in an independent experiment relative to vehicle controls. For sodium acetate, the vehicle control was sodium chloride.
Cells were lysed with RIPA buffer (Thermo Fisher Scientific) in the presence of a protease inhibitor cocktail (Thermo Fisher Scientific) and sonicated for 30 seconds. Lysates were separated by SDS-PAGE under reducing conditions, transferred to a PVDF membrane, and analyzed by immunoblotting. Rabbit anti-CPT1A (No. 12252) (Cell Signaling Technology) was used as a primary antibody. Immunoblotting for β-tubulin by mouse anti-β-tubulin antibody (Santa Cruz Biotechnology) and COX IV by rabbit anti-COX IV antibody (Cell Signaling) was used as a loading control for whole-cell lysates and mitochondrial lysates, respectively. Anti-rabbit and anti-mouse secondary antibodies were from Cell Signaling Technology and Thermo Fisher Scientific, respectively. Signal was detected using the ECL system with X-ray film development (Thermo Fisher Scientific and GE Healthcare Life Sciences) or a LI-COR C-Digit blot scanner (LI-COR) according to the manufacturer’s instructions.
Cells were preincubated with the vehicle control or 200 μM etomoxir for 48 hours. On the day of the assay, cells were trypsinized, washed 2 times with cold PBS buffer, and extracted according to the manufacturer’s instructions. The NADH/NAD+ ratio was measured and calculated using an NAD/NADH Quantification Colorimetric Kit (BioVision).
To assess the activity of FAO, cells were treated with vehicle control, etomoxir, scrambled siRNA, or CPT1A siRNA for 48–72 hours. Next, the medium was refreshed with new medium containing 100 μM uniformly 13C labeled (U-13C) palmitate-BSA and 100 μM natural abundance oleate-BSA. After labeling for 24 hours, cells were harvested, extracted, and analyzed as previously described [17]. For U-13C glucose, U-13C glutamine, and U-13C palmitate tracing experiments, cells were transferred to media containing 13C label and either vehicle control or 200 μM etomoxir for 12 hours, 6 hours, and 24 hours, respectively. The polar portion of the extract was separated by using a Luna aminopropyl column (3 μm, 150 mm × 1.0 mm I.D., Phenomenex) coupled to an Agilent 1260 capillary HPLC system. Mass spectrometry detection was carried out on an Agilent 6540 Q-TOF coupled with an ESI source operated in negative mode. Isotopic labeling was assessed comprehensively by using the X13CMS software [20]. The identity of each metabolite was confirmed by matching retention times and MS/MS fragmentation data to standard compounds. The isotopologue distribution patterns presented were obtained from manual evaluation of the data and calculated by normalizing the sum of all isotopologues to 100%. Data presented were corrected for natural abundance and isotope impurity.
After incubating cells in fresh media for 24 hours, the spent media were collected and analyzed. Known concentrations of U-13C internal standards (glucose, lactate, glutamine, glutamate, and palmitate; Cambridge Isotopes) were spiked into media samples before extraction. Extractions were performed in glass to avoid plastic contamination as previously reported [21]. Samples were measured by LC/MS analysis, with the method described above. For each compound, the absolute concentrations were determined by calculating the ratio between the fully unlabeled peak from samples and the fully labeled peak from standards. The consumption rates were normalized by cell growth over the experimental time period.
Mitochondria were isolated from BT549 cells as previously described [22]. In brief, cells were harvested, pelleted, and resuspended in cold mitochondrial isolation medium (MIM) (300 mM sucrose, 10 mM sodium 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid [HEPES], 0.2 mM ethylenediaminetetraacetic acid [EDTA], and 1 mg/mL bovine serum albumin [BSA], pH 7.4). Cells were then homogenized with a glass-Teflon potter. After homogenization, samples were centrifuged at 700 g at 4 °C for 7 minutes. The supernatant containing mitochondria was centrifuged again at 10,000 g for 10 minutes. Mitochondrial pellets were washed with cold BSA-free MIM, and the protein amount was determined by using a Bradford protein assay (Bio-Rad).
The oxygen consumption rate (OCR) of whole cells and isolated mitochondria was determined by using the Seahorse XFp Extracellular Flux Analyzer (Seahorse Bioscience). Cells were first incubated with vehicle control, 10 μM etomoxir, or 200 μM etomoxir for 1 hour prior to measuring respiration (we note that etomoxir was present in the assay medium as well). For CPT1A knockdowns, cells were treated with scrambled siRNA control or CPT1A siRNA for 48 hours. Cells were trypsinized and plated on a miniplate with the same seeding density 24 hours prior to the Seahorse assay. The assay medium consisted of 25 mM glucose, 4 mM glutamine, 100 μM palmitate-BSA, and 100 μM oleate-BSA in Seahorse base medium. The OCR was monitored upon serial injections of oligomycin (oligo, 2 μM), FCCP (1 μM, optimized), and a rotenone/antimycin A mixture (rot/AA, 1 μM). To measure the respiration of isolated mitochondria, freshly isolated mitochondria from BT549 cells were resuspended in cold mitochondrial assay solution (MAS). For the composition of MAS, see [22]. Samples were loaded on a miniplate with 5 μg of protein per well. Mitochondria were attached to the plate by centrifuging at 2,000 g (4 °C) for 20 minutes. After centrifugation, prewarmed MAS-containing substrates (10 mM pyruvate, 2 mM malate, 4 mM adenosine diphosphate (ADP), vehicle control, or etomoxir) were added to each well without disturbing the mitochondrial layer and then inserted into the XFp analyzer. OCR was monitored upon serial injections of rotenone (rot, 2 μM), succinate (suc,10 mM), and antimycin A (AA, 4 μM). Whole-cell OCR was normalized to the final cell number as determined by manual cell counting. Data presented were corrected for nonmitochondrial respiration.
Cells were incubated with 100 nM MitoTracker Red CMXRos (Thermo Fisher Scientific) or 4 μM JC-1 (Cayman Chemical) dissolved in complete media at 37 °C for 30 minutes. Cells were washed twice with PBS and then subjected to live imaging, or cells were fixed with 4% paraformaldehyde in PBS. Fixed cells were permeabilized with 0.1% Triton X-100 (Sigma Aldrich). Next, cells were washed twice with PBS, and nuclei were stained with DAPI. Cells were then mounted with ProLong Gold (Thermo Fisher Scientific). For live imaging, nuclei were stained with Hoechst 33342 (Thermo Fisher Scientific). Cells were imaged using a Zeiss LSM 880 confocal microscope equipped with Airyscan. Images were acquired with a Zeiss 20x, 40x, 63x/1.4 NA objective using the ZEN Black acquisition software. Samples were excited with 405 (for DAPI and Hoechst 33342), 514 (for JC-1 monomers), and 543 (for Mitotracker Red and JC-1 aggregates) laser lines. Images were processed and prepared using the ZEN Black software.
Samples were fixed in 2% paraformaldehyde/2.5% glutaraldehyde (Polysciences) in 100 mM sodium cacodylate buffer, pH 7.2, for 1 hour at room temperature. Samples were washed in sodium cacodylate buffer and postfixed in 1% osmium tetroxide (Polysciences) for 1 hour. Next, samples were rinsed extensively in dH2O prior to en bloc staining with 1% aqueous uranyl acetate (Ted Pella) for 1 hour. Following several rinses in dH2O, samples were dehydrated in a graded series of ethanol and embedded in Eponate 12 resin (Ted Pella). Sections of 95 nm were cut with a Leica Ultracut UCT ultramicrotome (Leica Microsystems), stained with uranyl acetate and lead citrate, and viewed on a JEOL 1200 EX transmission electron microscope (JEOL USA) equipped with an AMT 8 megapixel digital camera and AMT Image Capture Engine V602 software (Advanced Microscopy Techniques).
Isolated mitochondria (with known concentrations of internal standards) were extracted with chloroform/methanol/water (1:1:1) and vortexed for 1 minute. After centrifuging at 3,000 g for 15 minutes, the chloroform layer was dried under nitrogen gas and reconstituted with methanol/chloroform (95:5) according to the protein amount. Samples were separated using a Kinetex evo C18 column (2.6 um, 150 mm × 2.0 mm I.D., Phenomenex) coupled to an Agilent 1290 UPLC system. Mass spectrometry detection was carried out on an Agilent 6540 Q-TOF or a Thermo Scientific Q Exactive Plus coupled with an ESI source operated in both negative mode and positive mode. The lipid identities were confirmed by accurate mass as well as by matching retention times and MS/MS fragmentation patterns to standards. Absolute quantitation was achieved by normalizing to internal standards for (PC(14:1/14:1), PE(16:1/16:1), CL(14:0/14:0/14:0/14:0), PG(15:0/15:0), PS(14:0/14:0), PA(12:0/12:0), LPE(14:0), LPC(17:0), SM(d18:1/12:0), and Cer(d18:1/17:0)).
The first question we sought to address is whether FAO is dispensable in rapidly proliferating cancer cells, such as BT549. We pharmacologically targeted FAO by using the drug etomoxir (ethyl 2-[6-(4-chlorophenoxy)hexyl]oxirane-2-carboxylate), which has been regarded as a specific inhibitor of CPT1 [23, 24]. It is common in cancer studies to use etomoxir at hundreds of micromolar concentrations [5, 15, 18, 25, 26]. Here, we started by considering etomoxir at doses an order of magnitude lower. When BT549 cells were treated with 10 μM etomoxir, we measured over an 80% decrease in acylcarnitine species (the products of CPT1 activity, Fig 1A). Since changes in acylcarnitine levels may not reflect the same change in FAO, we directly assessed FAO by feeding cells uniformly labeled 13C-palmitate (U-13C palmitate) and measuring the labeling of FAO products. During FAO, U-13C palmitate is degraded to 13C2-acetyl-CoA. This acetyl-CoA then condenses with oxaloacetate in the TCA cycle to produce 13C2-citrate (the M+2 isotopologue). Upon treatment with 10 μM etomoxir, 13C2-citrate labeling from U-13C palmitate decreased by approximately 90% compared to vehicle controls (Fig 1B). These data demonstrate that 10 μM of etomoxir effectively blocks most of FAO.
Surprisingly, 10 μM etomoxir did not affect the proliferation rate of BT549 cells relative to vehicle controls (Fig 1C). Increasing the concentration of etomoxir by a factor of 10 to 100 μM led to further decreases in acylcarnitine levels and citrate labeling from U-13C palmitate, but we still did not observe a statistically significant change in the proliferation rate of BT549 cells (S2A Fig). Comparable results were observed in HeLa cells. When HeLa cells were treated with 100 μM etomoxir, no FAO activity was detected, yet we observed no alteration in proliferation (S2B and S2C Fig). An analysis of 6 additional cell lines produced similar results for B16, 3T3, MCF7, and HS578t cells (S2A Fig). Only 2 cell lines tested (H460 and T47D) showed a statistically significant decrease in proliferation with 100 μM etomoxir treatment. These data suggest that FAO is not an essential source of ATP or NADPH in some cancer cells, such as BT549.
Given that studies evaluating the role of CPT1 in cancer have commonly used concentrations of etomoxir at the hundreds of micromolar or even 1 mM level [15], we next assessed whether higher concentrations of etomoxir affected cell growth. Although 10 μM etomoxir was sufficient to inhibit most of FAO, residual FAO could be further reduced with increasing concentrations of etomoxir. This was reflected by additional small decreases in acylcarnitine pools (Fig 1A) and additional small decreases in the labeling of citrate from U-13C palmitate (Fig 1B). Despite the relatively small differences in FAO between 10 and 200 μM etomoxir-treated BT549 cells, we found that 200 μM of etomoxir resulted in a statistically significant reduction in cellular proliferation rate, while 10 μM did not (Fig 1C). These data are consistent with previous reports of the effects of 200 μM etomoxir on BT549 cells [18]. Interestingly, even though no FAO could be measured at 200 μM (Fig 1B), higher concentrations of etomoxir continued to result in further reductions in cell proliferation for BT549 cells (S3A Fig). Similar results were obtained from other cell lines tested (S3B Fig). Taken together, these observations suggest that high concentrations of etomoxir influence proliferation rate independent of FAO.
Since impairing approximately 90% of FAO did not change the rate of BT549 cell proliferation, we hypothesized that these cells might compensate for losses in ATP or NADPH production by increasing the oxidation of metabolic substrates other than fatty acids (e.g., glucose or glutamine). We therefore analyzed cell culture media to evaluate nutrient-uptake and waste-excretion rates of cells treated with etomoxir. Interestingly, when 90% of FAO was inhibited with 10 μM etomoxir, we observed no change in the rate of glucose uptake or lactate excretion (Fig 2A). Instead, with 10 μM etomoxir, we observed a 30% decrease in glutamate excretion. We note that cells treated with 10 μM etomoxir did not alter their glutamine uptake. These data suggest that when FAO is mostly blocked, BT549 cells can possibly compensate for the loss of energy/reducing equivalents by up-regulating glutaminolysis, by which glutamine carbons are fed into the TCA cycle instead of being excreted as glutamate (see S11 Fig, introduced below). Additionally, we observed a 30% decrease in the uptake rate of fatty acids (palmitate and oleate) in etomoxir-treated cells compared to vehicle controls, presumably because drug-treated cells cannot degrade these fatty acids by FAO.
Consistent with our proliferation results, different concentrations of etomoxir resulted in strikingly distinct nutrient utilization profiles that did not correlate with the small differences we observed in FAO. While 10 μM etomoxir did not change glucose uptake or lactate excretion, we observed an increase in glycolysis (indicated by a 3.5% increase in glucose consumption and a 4.8% increase in lactate excretion) when cells were given 200 μM etomoxir (Fig 2A). We also found that in contrast to the decrease in palmitate and oleate uptake we observed in cells treated with 10 μM etomoxir, cells treated with 200 μM etomoxir took up approximately 30% more palmitate and oleate even though these fatty acids could not be oxidized. Most notably, instead of decreasing by 30% as we observed with 10 μM etomoxir treatment, glutamate excretion increased by nearly 2-fold with 200 μM etomoxir (Fig 2A). Considering that glutamine uptake was unaltered, this result suggests that less glutamine carbon is available for oxidation in the TCA cycle at high concentrations of etomoxir (see S11 Fig, introduced below).
Although etomoxir is often assumed to be a specific inhibitor of CPT1, our observations above prompted us to consider other possible off-target activities, particularly at high drug concentrations as are often used in cancer studies [27]. We first examined the OCRs of BT549 cells treated with etomoxir. Cells were assayed in nutrient-rich media containing 25 mM glucose, 4 mM glutamine, 100 μM palmitate, and 100 μM oleate in the presence of vehicle control or etomoxir. Cells were treated with etomoxir for 60 minutes prior to making the oxygen consumption measurements. As we expected based on our nutrient-utilization data and proliferation results, the mitochondrial respiration profiles of cells treated with 10 μM etomoxir were not significantly different from vehicle controls (Fig 2B). With 200 μM etomoxir, however, mitochondrial respiration was significantly impaired. We measured a 65% decrease in basal respiration and a 65% decrease in maximal respiratory capacity after treating cells with 200 μM etomoxir. Moreover, we detected only minimal oxygen-driven ATP production, and the calculated mitochondrial coupling efficiency was therefore determined to be nearly zero (Fig 2C).
Given the impaired mitochondrial respiration that we observed with 200 μM etomoxir treatment in whole cells, we hypothesized that high concentrations of etomoxir might directly inhibit the activity of the electron transport chain. To test this possibility, we isolated intact mitochondria from BT549 cells and measured changes in oxygen consumption upon etomoxir treatment. By using isolated mitochondria instead of whole cells, we could control the availability of substrates for respiration. For this experiment, it is critical to point out that isolated mitochondria were assayed in buffer free of fatty acids, acyl-CoA species, acylcarnitines, and carnitine. Under such conditions, no FAO is occurring, and hence, CPT1 inhibition will not affect respiration. Any change in oxygen consumption upon etomoxir administration can therefore be attributed to off-target effects.
We evaluated mitochondrial respiration in 4 time segments over which various respiratory substrates and inhibitors were added (Fig 2D). The purpose of this experimental design was to distinguish respiration driven by complex I (state I respiration) from respiration driven by complex II (state II respiration). At time zero, mitochondria were provided pyruvate, malate, and ADP. These substrates enable turnover of the TCA cycle and production of NADH. Oxidation of NADH by respiratory complex I drives oxygen consumption. In time segment 2, we added rotenone to the mitochondria. Rotenone inhibits complex I and therefore blocks oxygen consumption under these conditions by preventing the electron transport chain from accepting its only source of electrons. In time segment 3, we provided mitochondria an alternative source of electrons in the substrate succinate. Oxidation of succinate feeds electrons into respiratory complex II of the electron transport chain, which is independent of the rotenone-inhibited complex I and therefore reintroduces oxygen consumption. Finally, in time segment 4, mitochondria were treated with antimycin A. Antimycin A inhibits respiratory complex III, thereby preventing the electron transport chain from oxidizing any of the substrates present. Under these conditions, there is no mitochondrial oxygen consumption. Data from vehicle controls (Fig 2D, black) were as expected.
Next, we independently considered isolated BT549 mitochondria treated with 200 μM etomoxir for 15 minutes. We performed the respiration measurements detailed above over the same 4 time segments. Notably, relative to the vehicle controls, there was a 35% decrease in state I respiration upon etomoxir treatment (Fig 2D, red). However, there was no statistically significant change in state II respiration. In contrast, 10 μM etomoxir resulted in similar OCRs for both state I and state II respiration (S4A Fig). These data suggest that high concentrations (200 μM) of etomoxir inhibit respiratory complex I but do not affect downstream proteins in the electron transport chain. The results also indicate that low concentrations (10 μM) of etomoxir do not have off-target effects on the electron transport chain (S4A Fig). Similar to etomoxir, we note that the complex I inhibitor rotenone also slowed down BT549 cell proliferation when given in culture media (S4B Fig).
We surmised that this off-target effect of 200 μM etomoxir on respiratory complex I might prevent regeneration of NAD+ from NADH and hence inhibit the turnover of the TCA cycle, thereby contributing to increased glycolysis and decreased glutaminolysis. Indeed, the intracellular NADH/NAD+ ratio was increased in cells treated with 200 μM etomoxir (S5 Fig). To further test our prediction, we fed BT549 cells U-13C glucose or U-13C glutamine and measured labeling in TCA cycle metabolites. Compared to vehicle controls, labeling of glycolytic intermediates from U-13C glucose was slightly increased, while labeling of TCA cycle metabolites from U-13C glucose was slightly decreased in cells treated with 200 μM etomoxir (Fig 2E, S6 and S7 Figs). These data are consistent with results shown in Fig 2A, indicating that cells treated with 200 μM etomoxir direct more glucose carbon into lactate instead of aerobic respiration. Although BT549 cells treated with 200 μM etomoxir showed only a modest increase in glycolysis, we note that much larger increases in glycolysis were observed for other cell lines treated with 200 μM etomoxir (S8 Fig). In BT549 cells treated with 200 μM etomoxir, we also observed a decrease in the overall labeling of citrate and other TCA cycle intermediates from U-13C glutamine relative to vehicle controls (Fig 2F, S9 Fig). Additionally, the pools of TCA cycle intermediates were decreased, with the exception of α-ketoglutarate, which is the entry point of glutamine into the TCA cycle (S10 Fig). These results are consistent with decreased glutaminolysis, indicated by similar glutamine uptake but increased glutamate excretion (Fig 2A). The relative TCA cycle activity can also be inferred by the ratio of the M+2 isotopologue to the M+4 isotopologue (i.e., M+2/M+4) of malate (S11A Fig). The M+2/M+4 ratio was higher when BT549 cells were treated with 10 μM etomoxir compared to vehicle control, while the M+2/M+4 ratio was lower when BT549 cells were treated with 200 μM etomoxir (S11B Fig). Interestingly, in cells treated with 200 μM etomoxir, we detected increased labeling of the M+5 isotopologue in citrate from U-13C glutamine. This result is consistent with a relative increase in the reductive metabolism of glutamine, which is a metabolic signature of cells under hypoxic stress [28].
Having established that etomoxir has off-target effects, we chose to use genetic methods to inactivate CPT1. There are 3 subtypes of CPT1 that are encoded by different genes and show tissue-specific distribution [29]. CPT1B is expressed in muscle, heart, and adipose tissue and CPT1C in neurons, whereas CPT1A is more widely expressed and has been previously implicated as a therapeutic target in breast cancer cells [2, 30, 31]. Using siRNA, we knocked down CPT1A mRNA levels by >90% relative to scrambled siRNA controls (S12A Fig). All of our assays to phenotype CPT1A knockdown (CPT1AKD) cells were performed at least 48 hours post transfection and completed within 96 hours, over which time CPT1A mRNA levels and protein levels remained greatly reduced (S12B Fig).
As evidence that knockdown of CPT1A blocked transport of fatty acids into the mitochondria, we observed major reductions in the levels of acylcarnitine species (S13 Fig), and we detected no 13C-labeled citrate after 24 hours of U-13C palmitate labeling (Fig 3A). These data indicated that CPT1A knockdown inactivated most of FAO. Notably, CPT1AKD cells had a significantly impaired proliferation rate (Fig 3B), with a 50% increase in doubling time (42.5 hours) compared to control wild-type cells with scrambled siRNA (27.8 hours). Given that the end product of β-oxidation is acetyl-CoA and that acetyl-CoA is readily produced from acetate, acetate supplementation has been shown to rescue cellular functions dependent upon FAO [19]. In our cells, however, impaired proliferation due to CPT1A knockdown could not be rescued by acetate supplementation (Fig 3C), suggesting again that CPT1A affects the growth of BT549 cells independent of FAO. Interestingly, supplementation of acetate slightly impaired BT549 cell growth. This could be partially explained by the osmotic effects of sodium, acetate’s counter ion (S14 Fig). We also attempted to rescue the proliferation of knockdown cells by supplementing them with octanoic acid, which can passively diffuse through the inner mitochondrial membrane independent of CPT1 and therefore compensate for impaired FAO [32]. Similar to acetate, supplementing cells with various concentrations of octanoic acid did not restore their proliferation (S15 Fig), further supporting that CPT1A knockdown influences cell phenotype independent of FAO.
To rule out the possibility that decreased cell proliferation in CPT1A knockdown cells was a result of off-target effects of siRNA, we performed 2 analyses. First, we tested 2 different siRNA sequences and observed comparable protein depletion and growth inhibition in both (S16A and S16B Fig). Given that growth inhibition is a common off-target effect of siRNA, however, we performed a second experiment in which we attempted to rescue CPT1A knockdown cells by overexpressing siRNA-resistant CPT1A protein (CPT1Aresistant) (S16C Fig). CPT1Aresistant protein led to a significant increase in FAO and cellular proliferation rate relative to vector controls (S16D and S16E Fig). Together, these data indicate that decreased proliferation in siRNA-treated cells is due to CPT1A loss of function rather than off-target effects.
We also observed changes in nutrient utilization upon CPT1A knockdown (Fig 4A). CPT1AKD cells had a nearly 2-fold increase in glucose uptake and lactate production relative to scrambled siRNA controls, indicating a substantial increase in glycolytic flux. Additionally, relative to wild-type cells with scrambled siRNA, CPT1AKD cells had a 2-fold increase in palmitate uptake and a 6.5-fold increase in oleate uptake. Yet, in contrast to cells treated with 200 μM etomoxir, CPT1AKD cells increased their uptake of glutamine by 45% and began uptaking glutamate instead of excreting it (Fig 4A). The increased utilization of glutamine and glutamate carbon suggests increased glutaminolysis and thus increased TCA cycle activity in CPT1AKD cells, whereas data from the etomoxir experiments indicate that 200 μM treated cells have a truncated TCA cycle due to complex I inhibition.
Increases in glycolysis and glutaminolysis are indicative of a change in mitochondrial activity [33]. Thus, we next examined oxygen consumption in whole cells after CPT1A knockdown. Unlike cells treated with 200 μM etomoxir, CPT1AKD cells had similar responses to respiratory inhibitors as wild-type cells with scrambled siRNA (Fig 4B). Compared to control cells, however, CPT1AKD cells had a 40% increase in proton leak and a 60% decrease in ATP production. Taken together, CPT1AKD cells had a 70% decrease in mitochondrial coupling efficiency, which compromised their ability to efficiently use respiratory substrates for ATP production. Possibly to compensate for this loss in energy, CPT1AKD cells show increased basal and maximal respiration (Fig 4C). These data are consistent with the observed increase in glucose, glutamine, and glutamate uptake (Fig 4A). We also note that 200 μM etomoxir similarly inhibited respiration in CPT1AKD cells, which is consistent with etomoxir having off-target effects on the respiratory chain independent of CPT1A protein (S17 Fig).
To further examine mitochondrial dysfunction in CPT1AKD cells, we applied fluorescence imaging and electron microscopy (EM). We first stained mitochondria with MitoTracker red, a positively charged fluorescent probe that accumulates as a function of membrane potential. We observed a significant increase in fluorescence intensity from MitoTracker red in CPT1AKD cells relative to controls, suggesting an alteration in mitochondrial membrane potential (Fig 5). Since interpreting this change with respect to increased or decreased mitochondrial membrane potential is complicated by the quenching effects of MitoTracker red at the concentration used, we also compared CPT1AKD and control cells with JC-1 staining [34, 35]. JC-1 accumulates in the mitochondrial matrix as a function of the mitochondrial membrane potential. In the cytosol, JC-1 exists in its monomer form and fluoresces green. Upon its accumulation in the mitochondria, JC-1 forms aggregates that fluoresce red. Accordingly, depolarized mitochondria are characterized by a decrease in the red/green fluorescence intensity ratio [36]. In CPT1AKD cells, we found a decreased ratio of red J-aggregates to green J-monomers relative to control cells (S18 Fig). As expected on the basis of our respiration measurements, these data are consistent with a depolarized mitochondrial membrane due to uncoupling in the CPT1AKD cells. Interestingly, upon CPT1A knockdown, we also observed multinucleated cells, which is a signature of cell-cycle arrest [37]. With electron microscopy (EM) imaging, we determined that more than 50% of the mitochondria in CPT1AKD cells had abnormal vesicular morphology compared to the well-defined cristae structure of control cells. Indeed, vesicular cristae shape has been associated with respiratory complex assembly and respiratory efficiency [38–40]. We did not observe abnormal mitochondrial morphology in etomoxir-treated cells (S19 Fig), possibly due to a less complete inactivation of CPT1 compared to knockdowns. We note that although FAO is mostly inhibited in both BT549 cells treated with 200 μM etomoxir (Fig 1B) and in CPT1AKD cells (Fig 3A), the isotopologue distribution patterns of citrate after U-13C palmitate labeling cannot be used to compare the level of CPT1A inhibition. This is because 200 μM etomoxir has the off-target effect of inhibiting complex I, which impairs the regeneration of NAD+ and thereby influences the oxidative degradation of U-13C palmitate.
Pyruvate and uridine enable some cells lacking a functional mitochondrial electron transport chain to proliferate [41, 42]. Thus, we sought to test whether pyruvate and uridine could rescue growth in CPT1A knockdowns with dysfunctional mitochondria. When BT549 cells with knocked down CPT1A were given pyruvate and uridine, their proliferation rate remained significantly less than that of controls (S20 Fig). These results are consistent with CPT1A knockdown cells having a functional electron transport chain that can regenerate oxidized cofactors and suggest that their dysfunctional mitochondria impair cell growth by a different mechanism.
Our data suggest that knocking down CPT1A affects cell proliferation through a mechanism that is independent of FAO. As one such potential mechanism, we considered the possibility that CPT1A plays an important structural function essential to the integrity of the mitochondrial membrane. To assess this hypothesis, we expressed CPT1A having G709E and G710E mutations in BT549 cells. The replacement of glycine residues 709 and 710, which are part of the catalytic site, with glutamate abolishes CPT1A activity (S21A Fig) [43, 44]. We refer to this catalytically dead CPT1A as CPT1Amutant. We also note that CPT1Amutant was resistant to any siRNA added to knock down wild-type CPT1A. This allowed us to knock down wild-type CPT1A in BT549 cells, without affecting CPT1Amutant expression. We found that expression of CPT1Amutant protein did not rescue cells in which wild-type CPT1A had been knocked down. Specifically, expression of CPT1Amutant did not restore proliferation or mitochondrial membrane potential in wild-type CPT1A knockdowns (S21B–S21D Fig). These data do not support a structural role for CPT1A that is independent of FAO.
As another mechanism for how CPT1A may influence cell proliferation independent of FAO, we considered the possibility that CPT1A mediates transport of long-chain fatty acids into the mitochondria for anabolic purposes. That is, instead of oxidizing long-chain fatty acids transported into the mitochondria by CPT1A for energy, we hypothesized that the carnitine shuttle provides an indispensable source of fatty acids to synthesize complex lipids during cellular proliferation [45]. To test our hypothesis, we isolated mitochondria from CPT1AKD and wild-type cells and applied lipidomic profiling to quantitate differences in mitochondrial lipids. Consistent with our prediction, many complex lipid species had decreased levels in CPT1AKD cells relative to wild-type cells (Fig 6A). In our untargeted profiling experiment, 87% of the dysregulated lipids were decreased (see S1 Table). We then quantified the change in concentrations of these altered lipid features, which included complex structural lipids such as phospholipids, sphingolipids, and cardiolipins (Fig 6B). We also observed a nearly 2-fold decrease in complex signaling lipids such as lactosylceramide and glucosyl/galactosylceramides. Smaller decreases were found in other signaling lipids such as lysophospholipids and diacylglycerols. Whether they are a direct consequence of limited long-chain fatty acid availability or a downstream consequence of altered mitochondrial metabolism, these data suggest that CPT1A plays a role in regulating the levels of mitochondrial lipids.
In recent years, multiple cancers have been found to have increased expression of CPT1 and/or sensitivity to CPT1 inhibition [6, 9]. In the conventional textbook picture of mammalian metabolism, CPT1 commits long-chain fatty acids to catabolic oxidation [46]. Thus, increased expression of CPT1 and/or sensitivity to CPT1 inhibition has been assumed to represent a demand for FAO and the ATP or NADPH provided. Our work here reveals 2 complications with this interpretation: (1) pharmacological inhibition of CPT1 with high concentrations of etomoxir, as is often used in cancer studies, leads to off-target effects, and (2) CPT1 influences the proliferation of several cancer cell lines independent of FAO.
Treatment of BT549 breast cancer cells as well as several other cancer cell lines with 200 μM etomoxir significantly slowed cell proliferation, which is consistent with previous studies [18]. However, decreased cell proliferation at 200 μM etomoxir is not a result of inhibiting the primary target of etomoxir (i.e., CPT1). Rather, 200 μM etomoxir inhibits complex I of the electron transport chain (an off-target effect) and leads to decreased cell proliferation independent of FAO. We note that 10 μM etomoxir efficiently blocked 90% of FAO and did not exhibit off-target effects on respiration; however, 10 μM etomoxir did not reduce BT549 cell proliferation. When most of FAO was inhibited with 10 μM etomoxir, BT549 cells adjust their uptake and utilization of other nutrients to compensate for the loss of FAO. These data indicate that FAO provides a dispensable source of ATP and reducing equivalents under standard cell-culture conditions.
FAO generates acetyl-CoA, FADH2, NADH, ATP, and potentially cytosolic NADPH. Importantly, all of these products can be derived from other nutrient sources without using CPT1. Glucose, for example, can provide cytosolic NADPH via the pentose phosphate pathway and acetyl-CoA from glycolysis and the pyruvate dehydrogenase complex. FADH2, NADH, and ATP can be obtained from the oxidation of glucose carbon through the TCA cycle. Similarly, reducing equivalents and ATP can be readily derived from glutamine [47]. Thus, while the products of FAO are highly valuable to a cell and may serve as a major energy source, they are not unique to the FAO pathway. Our results suggest that some cells, such as BT549, can therefore compensate for the loss of FAO by adjusting nutrient uptake and utilization.
Inhibiting approximately 90% of FAO by pharmacological inhibition of CPT1 did not affect the proliferation rate of BT549 cells, but genetic knockdown of CPT1A did. Moreover, genetic knockdown of CPT1A altered mitochondrial morphology and caused mitochondrial uncoupling, while pharmacological inhibition of CPT1 did not. These data together with the observations that acetate and octanoic acid did not rescue CPT1A knockdowns indicate that CPT1A has a function affecting cell proliferation that is independent of its role in FAO. We first considered a structural function of CPT1A as a scaffolding protein. However, expression of a catalytically dead CPT1A in BT549 cells in which wild-type CPT1A had been knocked down did not restore mitochondrial membrane potential.
As another possible function of CPT1 that is independent of FAO, we considered the need to use CPT1 for purposes other than catabolic oxidation of lipids. Without CPT1, cells cannot transport long-chain fatty acids into mitochondria, and therefore, downstream mitochondrial pathways using these substrates are impaired (Fig 7). Sources of long-chain fatty acids (or long-chain fatty acyl-CoAs) inside the mitochondria that do not rely on the CPT1 transport system are limited [48–50]. Complex lipids synthesized in the endoplasmic reticulum can be transported to the mitochondria and deacylated to make long-chain fatty acids [51, 52], or long-chain fatty acids can be generated in the mitochondrial matrix by type II mitochondrial fatty acid synthesis, a pathway that resembles fatty acid synthesis in bacteria [53]. Although the fates of long-chain fatty acids generated by these processes remain poorly understood, disrupting mitochondrial fatty acid synthesis slows cell growth, influences mitochondrial phospholipid composition, and alters mitochondrial morphology [54–58], phenotypes which are highly consistent with those that we observed here with CPT1A knockdown. One possible explanation for these findings is that long-chain fatty acids generated in mitochondria are involved in phospholipid side-chain remodeling [54]. The de novo synthesis of cardiolipin in the mitochondria, for example, is followed by cycles of deacylation and reacylation. This remodeling process is essential to mitochondrial structure and function and, at least in part, uses acyl-CoA substrates in the mitochondrial matrix [59, 60]. Another possible demand for long-chain fatty acids in the mitochondria is protein acylation, which may be used for protein anchoring, cell signaling, or protein trafficking. Although acylation of mitochondrial proteins remains largely unexplored, many mitochondrial proteins have been shown to be modified with long acyl chains in the mitochondrial matrix [61, 62]. It is important to note that any anabolic demand for long-chain fatty acids transported by CPT1A in BT549 cells is likely to be low, since pharmacologically inhibiting most of CPT1 activity with low concentrations of etomoxir does not result in decreased cell proliferation or mitochondrial dysfunction. Interestingly, the demand for mitochondrial fatty acid synthesis is also low, but its disruption similarly results in decreased cell proliferation and mitochondrial dysfunction [54]. Our results therefore suggest that, like mitochondrial fatty acid synthesis, the CPT1 system may provide an indispensable source of long-chain fatty acids in the mitochondria to support processes that do not demand much carbon (such as phospholipid remodeling and protein acylation) but are essential to healthy mitochondrial function and cancer cell proliferation. We also point out that the results obtained for the cancer cells studied here are unlikely to be generalizable to all cancer cells; however, they demonstrate that additional evidence independent of CPT1 is necessary to implicate FAO as an antitumor target.
|
10.1371/journal.pcbi.1004255 | Understanding Voltage Gating of Providencia stuartii Porins at Atomic Level | Bacterial porins are water-filled β-barrel channels that allow translocation of solutes across the outer membrane. They feature a constriction zone, contributed by the plunging of extracellular loop 3 (L3) into the channel lumen. Porins are generally in the open state, but undergo gating in response to external voltages. To date the underlying mechanism is unclear. Here we report results from molecular dynamics simulations on the two porins of Providenica stuartii, Omp-Pst1 and Omp-Pst2, which display distinct voltage sensitivities. Voltage gating was observed in Omp-Pst2, where the binding of cations in-between L3 and the barrel wall results in exposing a conserved aromatic residue in the channel lumen, thereby halting ion permeation. Comparison of Omp-Pst1 and Omp-Pst2 structures and trajectories suggests that their sensitivity to voltage is encoded in the hydrogen-bonding network anchoring L3 onto the barrel wall, as we observed that it is the strength of this network that governs the probability of cations binding behind L3. That Omp-Pst2 gating is observed only when ions flow against the electrostatic potential gradient of the channel furthermore suggests a possible role for this porin in the regulation of charge distribution across the outer membrane and bacterial homeostasis.
| Porins are the main conduits for hydrophilic nutrients and ions uptake into the periplasm of Gram-negative bacteria. Their translocation permeability is determined by the amino-acid distribution on their extracellular loop L3. Bacterial porin channels have long been known to undergo step-wise gating, under the application of a transmembrane potential. Yet the exact molecular mechanism by which gating is achieved and the exact relevance of this evolved characteristic remain elusive. In the present study, we report on electrophysiology experiments and molecular dynamics simulations on the two general-diffusion porins of Providencia stuartii, Omp-Pst1 and Omp-Pst2. Our results show that gating in Omp-Pst2 occurs as the result of L3 displacement, which follows from the binding of cations in acidic niches between L3 and the barrel wall and effects in exposing the side chain of a highly conserved aromatic residue at the tip of L3 in the channel lumen. That Omp-Pst2 displays asymmetric voltage sensitivity and that the likelihood of gating is increased when cations transit from the extracellular to the intracellular side suggests voltage-gating underlies a regulatory role in bacterial homeostasis. Rational antibiotic-design strategies based on the maximization of antibiotic penetration and accumulation at their target sites, should take this role into account.
| The outer membrane of Gram-negative bacteria is sprinkled of homotrimeric channels comprised of three 16-stranded β-barrels. They are the principal conduits for the passive penetration of hydrophilic molecules into the periplasm, and are often referred to as the general diffusion porins. Classical examples include Escherichia coli OmpF [1], OmpC [2] and PhoE [1]. The general diffusion porins display a conserved β-barrel architecture with eight periplasmic turns and eight extracellular loops (L1–L8). Also conserved is the presence of a constriction zone (CZ), at mid height of the channel. The CZ is contributed by the plunging of extracellular loop L3 into the channel lumen, where it adopts a helix-turn-loop fold and interacts with the barrel wall through hydrogen bonding and Van der Waals (VDW) interactions. The amino-acid distributions in L3 and on the barrel wall opposite to it (“anti-L3 region”) determine the sieving properties of the porins, i.e. their ion specificity and size exclusion limit [3, 4]. The L3 and anti-L3 region generally display opposed charge distribution, with L3 being negatively charged and anti-L3 positively charged.
The general diffusion porins can switch from their open state to gated states when a transmembrane potential is applied [5–10], a phenomenon termed as voltage gating (VG). VG is characterized by step-wise, long-lasting closed states that persist until the transmembrane potential is suppressed. The critical voltage required for voltage gating (Vc) varies among porins, but is generally in the order of hundreds of mV. Vc can be influenced by a variety of environmental cues, including pH and salt concentration [11, 12], membrane constitution [8, 13], polarity of the transmembrane potential [9], or the presence of effectors such as oligosaccharides [13] and polyamines [14]. Accumulated evidences have suggested that voltage sensing in general diffusion porins occurs at the CZ [15–21]. It was shown that replacement of L3 charged residues by uncharged ones invariably results in alteration of voltage sensitivity, channel conductance and/or ion selectivity [15, 21, 22]. In particular, replacement of negatively charged residues from L3 leads to an increase of Vc in cation-selective E. coli OmpF, while that of the positively charged residues in anti-L3 regions causes the decrease of Vc in anion-selective E. coli PhoE [15]. In addition, mutagenesis studies on OmpF showed that destabilization of L3 by deletion of residues at its tip leads to increased voltage sensitivity (lower Vc) and reduced conductance [18, 19], whereas mutations of residues involved in L3 stabilization result in reduced voltage sensitivity (higher Vc) [19, 20].
More than a decade ago, Tieleman et al reported the first molecular dynamics (MD) simulation of E. coli OmpF embedded in explicit lipids. Their results revealed instability in L3 due to a breakdown of the hydrogen-bonding network (HBN) anchoring L3 to the barrel wall [23]. Suspecting that the fluctuation in the CZ may have been due to the protonation setting, Im et al and Varma et al respectively implemented MD simulations using different ionization states to the charged residues on L3 [24–26]. The invariable observation was that L3 is prone to large fluctuations, suggesting that this loop could intervene in translocation across porin and possibly also in voltage-gating. Owing to constant improvements in MD simulation algorithms [27, 28] and the successful implementation of artificial transmembrane potentials [29], it has become possible to simulate ion mobilizing within the channel and thus to study the channel transport properties at a molecular level. For example, Pezeshki et al showed that mutation of one charged residue within CZ leads to visible effects on ion permeation and selectivity in OmpF [4]; Faraudo et al found that removal of negative charges in CZ influences the distribution of cations along OmpF channel [30]. However, owing both to limited computational resources and to the low voltage sensitivity (high Vc) of the hitherto simulated porin (OmpF), the molecular mechanism by which the general diffusion porin gating occurs has not yet been reported.
In the following, we report on extensive comparison of sub-microsecond scale MD simulations that provide insights into the molecular basis of voltage gating in general diffusion porins. Simulations were conducted on Omp-Pst1 and Omp-Pst2, two general diffusion porins from Providencia stuartii [31]. The sequence identities with OmpF of Omp-Pst1 and Omp-Pst2 are 50% and 46.1% respectively, while the RMS deviation of their Cα atoms (as measured from their respective X-ray structures) are 0.94 Å and 0.89 Å respectively. Omp-Pst1 and Omp-Pst2 show a high level of structural similarity, but owing to a completely different pattern of charge distribution along their channel wall, the two porins display opposite ion selectivities. Furthermore, whereas Omp-Pst1 gates at voltages above 199 mV, Omp-Pst2 undergoes the typical three-step gating at voltages as low as ~20 mV, making it the most voltage sensitive bacterial porin studied to date. Taking advantage of this striking contrast, we constructed parallel simulations between Omp-Pst1 and Omp-Pst2 at positive, negative and none transmembrane potentials (VTM; extracellular to intracellular). In Omp-Pst2, we observed gating at VTM < 0 V, but not at VTM > 0 V, consistent with asymmetrical gating observed experimentally. At VTM < 0 V, the gating stems from stable binding of cations in acidic niches behind L3, which in turn disrupts the HBN anchoring L3 to the barrel wall and thereby allows a local conformational change in conserved W111 at the tip of L3. The repositioning of W111 aromatic side chain in the middle of the CZ effectively halts ionic permeation across the Omp-Pst2 channel. At VTM > 0 V, the HBN between L3 and the barrel wall strengthens. The stabilized L3 impedes cation binding, and thus thwarts gating. In Omp-Pst1, gating was not observed, regardless of the VTM applied. Additional hydrogen bonds in HBN contribute to a more resilient L3. Altogether, our results suggest that the voltage sensitivity of Omp-Pst1 and Omp-Pst2 is encoded in the HBN that anchors L3 onto the channel wall, and that conformational changes in the side chain of conserved W111 at tip of L3 leads to channel closing.
Reconstitution of a single Omp-Pst1 / Omp-Pst2 channel in a planar lipid bilayer showed a single trimer conductance of 2.7 ± 0.1 / 3.7 ± 0.2 nS respectively at 1M KCl, pH 7. In the ion selectivity measurements performed as described in [32], Omp-Pst2 shows strong cation selectivity whereas Omp-Pst1 shows indistinctive cation selectivity (Table 1). For Omp-Pst1, the critical voltage (Vc) for observing the typical three-step gating was ≥199 mV in all measurements (n = 8). For Omp-Pst2, pore-to-pore variation was observed, and Vc measurements varied between 20 and 90 mV, with a median at 50 mV (n = 20).
The crystallographic structures of Omp-Pst1 and Omp-Pst2 trimers were used as starting model for the simulations. Briefly, each porin was inserted in a lipid bilayer, solvated at pH 7, and the ionic concentration was adjusted to 1 M KCl. After equilibration, the two systems were subjected to either a negative (VTM direction pointing from extracellular to intracellular) or a positive transmembrane potential (VTM direction pointing from intracellular to extracellular) and each simulation was ran for > 500 ns. Additional simulations without a transmembrane potential were also carried out for both porins and each lasted for 100 ns. The β-barrel remains rigid in all voltage conditions, showing RMS fluctuation lower than 1 Å (S1 Fig). Larger fluctuation is observed in extracellular loops, especially in L1, L4, L5 and L6. L1 lies on the interface of the trimer and interacts with the positively charged residues at anti-L3 regions from another monomer. L4, L5 and L6 form secondary structure elements that associate to cover the extracellular entrance of the channels. Applying transmembrane potentials does increase the amplitude of loop fluctuation. However, according to the principal component analysis, their movements are constrained in a radius of ~5 Å, as imposed by a strong network of polar interactions between adjacent loops (S2 Fig).
In simulations with VTM ≠ 0, anions and cations separate into two pathways at the CZ: cations (K+) trail along negatively charged L3 (residues Y98 to D123 in Omp-Pst1 and Y95 to D120 in Omp-Pst2, respectively), while anions (Cl-) duct on the positively charged anti-L3 region (Fig 1A and 1B). In Omp-Pst1, K+ mainly shuffle between the acidic side chains of L3 residues D109 and D117, while Cl- interact with the basic side chains of anti-L3 residues K16, R20, R41, R59, K65, R78, K163 and K170. In Omp-Pst2, these residues are D106, D114 and D117, on the one hand, and R20, R38, R56, R75, K160 and K168, on the other.
Steady ionic currents were developed in the voltage-applied systems except for Omp-Pst2 at VTM < 0 mV. As we show below, gating occurred in this simulation, thus impacting ion current across the channels (Figs 2A and S3). In the other simulations, ion currents across the channel reached excellent agreement with the experimental measurements after fluctuating for ~200 ns. The current curve suggests that long equilibration time (~100–200 ns) is required when simulating porins at high ionic concentration (1M in our case). That this requirement was not fulfilled in earlier simulations on porins at high ionic concentrations may explain the discrepancy between their simulated and experimentally measured currents [4, 33]. From the stable K+ and Cl- permeation after 200 ns or before gating in the case of Omp-Pst2 at VTM < 0 mV, we derived the simulated conductance via least-square linear regression of the I/V curve (VTM = [–1, 0, 1] V). The raw conductance for Omp-Pst1 and Omp-Pst2 are 3.64 nS and 4.01 nS respectively, while the corrected values based on ion diffusion coefficients (See Methods) are 3.95 nS and 4.47 nS respectively. The deviation from experimental observables might result from poor data samples in regression fitting.
Simulations indicate a strong cation-selectivity for Omp-Pst2 and a mild anion-selectivity for Omp-Pst1 (Fig 2C), in line with experimental electrophysiology data and also with the analysis of their structures. The net charge of Omp-Pst1 is indeed +1 e at the constriction zone (within 5 Å radius of channel’s narrowest point) and +5 e along the pore, while that of Omp-Pst2 is -4 e at the constriction zone and -3 e along the pore. In the case of Omp-Pst2, we observed that translocation of cations from the intracellular to the extracellular side is more efficient than the other way around (and conversely for Cl- ions), as grounded by its two times more pronounced cation-selectivity at VTM > 0 than at VTM < 0 (Fig 2C). This asymmetry in ion selectivity correlates with the asymmetry in charge distribution (and thus with the gradient of electrostatic potential) along Omp-Pst2 channel, which features more acidic residues on the extracellular side than on the intracellular side. In Omp-Pst1, where the distribution of charged residues is even along the pore (Fig 1C), ion selectivity and translocation rate are less affected by the direction of the transmembrane potential (Fig 2C). Our simulations thus suggest that the ion selectivity of porins is not only determined by the charge distribution at their constriction zone [30, 34, 35], but also by the profile of charge distribution along the channel.
After 100 ns, a decline in ion fluxes was observed in the simulation of Omp-Pst2 at VTM < 0. Examination of ion fluxes on a monomer basis reveals that the decline mainly stems from one monomer in the trimer (monomer B) undergoing gating. At VTM < 0, K+ ions are forced to translocate from the extracellular to the intracellular side, and inversely for Cl- ions. This direction of transit is unfavourable for cations, as suggested by the electrostatic potential developed along Omp-Pst2 channel and established by the two times slower uptake of cations in comparison to the other direction. As this unfavourable flux gets heavier across the channel of monomer B, it effects in disorganizing the HBN within L3 (S4 Fig). A main-chain flip consequently occurs after 84 ns in 112-GA-113, resulting in the opening a highly acidic niche (niche I) between L3 and the barrel wall (Fig 3A). The acidic nature of the niche mainly results from the side chain oxygens of E258, a barrel-wall residue hitherto shielded from the bulk, but also from the side-chain oxygens of N102, T105, T115 and N276. A K+ ion rapidly (within 0.5 ns) lodges into niche I (Fig 3F), dragging along the side chain of D106. This results in the latter adopting a barrel-facing conformation, thereby reinforcing of the acidic nature of the niche. The change in orientation of D106 propagates to D114, whose side chain draws towards the channel lumen, leading to yet another main chain rearrangement in L3 (~114 ns) (Fig 3C). Originating in 112-GA-113, these changes rapidly transduce to 111-WGAD-114 (~131 ns) and result in W111 side chain wandering away from the barrel wall (Fig 3D and 3E). W111 movement effects in uncovering D312, a highly conserved barrel wall residue whose side chain oxygens hydrogen bond to the tip of L3 (main chain nitrogens of L110 and W111) in the crystallographic structure of Omp-Pst2. Exposure of D312 side chain generates a second acidic niche (niche II) in which another K+ ion binds ~10 ns latter (Fig 3F). Occupation of this niche effects in unleashing the tip of L3, and most notably, the side chain of W111 (Fig 3B). Originally constrained by Y20, K314 and V334, W111 first positions itself in the middle of L3 (~187 ns), but finally ends up in the channel lumen (~260 ns). There, it imposes steric and hydrophobic hindrance to ions translocation, notably through the formation of a hydrophobic belt by Y95, Y99 and A103 side chains (Fig 3G). In the last 100 ns, translocation of K+ ions is diminished by ~3/4 in monomer B, while that of Cl- ions by ~1/4 (S5 Fig).
In the two other monomers of the Omp-Pst2 trimer simulated at VTM < 0, a comparable sequence of events is observed, albeit incomplete and spanning a longer time scale. In monomer C, a potassium ion binds in niche I after ~110 ns, following a main chain flip in 112-GA-113 (~80 ns). Similar to that in monomer B, binding of this first K+ ion induces a conformational change in 111-WGAD-114 (~240 ns) that results in the detachment of W111 from the barrel wall (243 ns) and the concomitant opening of niche II. However, binding of a K+ ion in this niche occurs at more than 200 ns latter (~ 453 ns). Thus, D312 side chain remains in its native conformation and migration of W111 side chain toward the channel lumen does not complete (S6 Fig). Accordingly, only a slight reduction of K+ flux is observed, while Cl- flux remains unaffected. In monomer A, binding of a K+ ion in niche I occurs after 327 ns, and the resulting conformational change in 112-GA-113 after 350 ns. By the end of the simulation time, however, no K+ ion lodges into niche II. W111 consequently stays in place, and ionic fluxes remain steady (S7 Fig). It is interesting to recall that in electrophysiology experiments, gating of the three monomers in a trimer also occurs sequentially. Thus monomers A, C and B could represent different Omp-Pst2 intermediates in the process of gating.
No voltage gating was observed in the simulation of Omp-Pst2 at VTM > 0. This observation is consistent with experimentally observed asymmetrical gating behaviour [10, 36, 37]. Trajectory analysis reveals that it is the stronger stabilization of L3 at VTM > 0 that impeaches gating. In short, a main chain rearrangement in 112-GA-113 leads to exposure of E258 and concomitant binding of K+ ions in niche I after ~340, 20 and 95 ns in monomers A, B and C respectively. In the two latter, main chain rearrangements in W111 (after 210 and 340 ns in monomers B and C, respectively) precede binding of K+ ions in niche II (after 245 and 351 ns in monomers B and C, respectively). Nevertheless, K+ binding to niche II is less stable at VTM > 0 and gating is hence not observed (S8–S10 Figs).
Two factors are responsible for the reduced gating sensitivity of Omp-Pst2 at VTM > 0. First and most importantly, the HBN anchoring L3 to the barrel wall is more robust at VTM > 0 than at VTM < 0 (Fig 4A), owing to a more favourable orientation of acidic side chains. At VTM > 0, the acidic residues on Omp-Pst2 L3 are indeed facing toward the barrel wall, facilitating interactions with non-L3 residues. Major stabilizing interactions include: i/ a salt-bridge between E119 and R126; ii/ alternating hydrogen bonds between D117 and either Y99 or N102; and iii/ an hydrogen bond between D114 and Y294 before the main chain rearrangements in 112-AG-113 occurs (Fig 4B). In strong contrast, L3 acidic residues adopt outward conformations in the simulation at VTM < 0, which prevents them from forming hydrogen bonds or salt-bridges with neighbouring residues. Secondly, that the transit of K+ ions is energy favoured at VTM > 0 and that these ions thus transit faster across the CZ furthermore diminishes their likelihood of forming stable interaction with D312. Their intermittent binding to niche II does not trigger the repositioning of W111 side chain.
From a structural point of view, Omp-Pst1 shares the critical features of two acidic niches under L3, with E266 and D321 (equivalent to Omp-Pst2 E258 and D312) being the main contributors to niche I and II, respectively. Likewise, an aromatic residue is found at its L3 tip, viz. W114 (equivalent to W111 in Omp-Pst2). Yet, gating was not observed in Omp-Pst1 on the time-scale of our simulations, regardless of the directionality of the transmembrane potential. Accordingly, the HBN holding L3 attached to the barrel wall, W114-D321, remained unaffected throughout all simulations (Fig 5A). These observations are in faithful agreement with experimental data, which established a higher Vc for Omp-Pst1 than Omp-Pst2 (Table 1 and S11 Fig).
At VTM < 0, binding of a K+ ion in niche I (~200 ns) occurs in monomer A, where it results in a conformational jump of L3 residues 115-AG-116 (equivalent to 112-GA-113 in Omp-Pst2) toward the extracellular side. In monomers B and C, however, neither binding of a K+ ion in niche I nor a conformational change in L3 is observed. Niche II meanwhile remains unoccupied in all three monomers (S12–S14 Figs). At VTM > 0, binding of a K+ ion in niche I is observed in monomers A and B (after ~196 and ~254 ns, respectively), which precedes a conformational change in 115-AG-116 (after ~300 and ~500 ns, respectively). But again, niche II remains unoccupied in all three monomers (S15–S17 Figs). Thus at both VTM, and in all three Omp-Pst1 monomers, D321 remains covered by the side chains of L113 and W114. The tip of L3 is therefore kept fastened to the barrel wall, and W114 maintains its native, open-state conformation.
The discrepancy in the voltage sensitivity can be rationalized by a comparative analysis of Omp-Pst1 and Omp-Pst2 structures. In Omp-Pst2, the tip of L3 is attached to the barrel wall solely through hydrogen bonding of L110 and W110 main chain nitrogens to D312, whereas the presence of Q270, S280 and Y309 in Omp-Pst1 contributes three additional hydrogen bonds that manifestly increase the energy barrier of exposing the acidic niches (Fig 5B). The lower sensitivity to voltage of Omp-Pst1 (S18 Fig) is thus encoded in the HBN holding L3 attached to the barrel wall. Of note, we also observed asymmetry in HBN resilience in simulations of Omp-Pst1. At VTM > 0, and not at VTM < 0, the hydrogen bond between Y309 and V110 is disrupted at the end of the simulation. Therefore, Omp-Pst1 could gate faster at this VTM (Fig 5C). Opposed asymmetries in voltage sensitivity were reported earlier for E. coli OmpF and PhoE [10].
This study reports on four 0.5 μs time-scale simulations conducted on the two porins from P. stuartii, viz. Omp-Pst1 and Omp-Pst2, at different transmembrane potentials. Under application of a negative transmembrane potential, we observed partial gating in one monomer of Omp-Pst2 (monomer B). Trajectory analysis revealed that a sequential changes on L3, including loop rearrangement and exposure of the hitherto hidden acidic niches, leads to protrusion of W11 side chain in the channel lumen. The presence of this large indol ring in the middle of the CZ hinders ion transit and results in the conductance of the monomer being decreased by ~50%. Based on curve fitting (I(t) = a × e-b×t) to the calculated currents, we estimate that the time required for full closure is about tens of microseconds. Direct comparison from experiments under similar transmembrane potentials (VTM = ±1 V) is not available, as lipid membranes do not withstand application of high voltages in experiments. However, we note that it was experimentally demonstrated that the time required for a bacterial porin to gate fully decreases exponentially as the applied voltage increases [6, 38]. Thus the estimated closing constant of Omp-Pst2 could be plausible. In the rest of our voltage-applied systems, stronger attachment of L3 loop onto the barrel wall was observed and the porins of study failed to undergo gating. We thus propose that the strength of this hydrogen bond network determines voltage sensitivity in the two porins from P. stuartii.
Based on our simulations, the most critical residues for porins (Omp-Pst2/1) gating are the acidic niches residues E258/E266 and D312/D321, on the one hand, and W111/W114 on the other hand. Sequence alignment with E. coli porins of known structures shows good conservation of these residues. First, all E. coli porins feature an aspartic acid at the position equivalent to D312 in Omp-Pst2 (niche II; viz. D312, D315 and D302 in OmpF, OmpC and PhoE, respectively). While E258 is not strictly conserved in terms of sequence position, negatively charged residues emanating from the barrel wall are found at other locations behind the L3 loop in E. coli porins, where they could contribute to an equivalent of niche I (E296 in OmpF; E260 and D299 in OmpC; E248 and D287 in PhoE) (S19B and S19C Fig). Also, an aromatic residue is found at the L3 tip of all E. coli porins (F118, F110 and F111 in OmpF, OmpC and PhoE, respectively) (S19A Fig). That residues critical for the voltage-gating of Omp-Pst2 are highly conserved in E. coli porins suggests that our gating model could apply to them as well. Furthermore, the lower sensitivity to transmembrane potential of OmpF, OmpC and PhoE (~155, ~185 and ~135 mV, respectively) correlates with the higher stabilization of their L3 loop onto the barrel wall. E. coli porins indeed all feature an additional glutamic acid at the L3 tip (E117, E109 and E110 in OmpF, OmpC and PhoE, respectively), which reinforces the attachment of their L3 tip to the barrel wall by one (Y22 in PhoE) or two hydrogen bonds (Y22/Y22 and Y310/Y313 in OmpF/OmpC). Thus, our proposal that it is the strength of the HBN attaching the tip of L3 onto the barrel wall that determines the likelihood of gating—and thus the sensitivity to voltage—is supported by examination of E. coli porins structures and by their Vc.
Further support to our proposed model for gating comes from mutagenesis data gathered on OmpF, which, similar to Omp-Pst2, is cation-selective. Replacement of acidic residues L3 (E117C) and niches residues (D312N, E296L, E296A and E296Q) by neutral ones, or stabilization of L3 by disulphide-bridge tethering onto the barrel wall (between E117C and A333C) [19, 22] both leads to increased Vc in OmpF. Reversely, replacement of basic residues in the anti L3 region (K16A, K16D, R42C, R82C, R132A, and R132D) [17, 18] leads to decreased Vc, as expected from a heavier trafficking of cations at the CZ—and thus an easier binding behind L3. That disulphide-tethering of L3 onto the barrel-wall does not suppress gating but merely increases the required voltage [20] is in accordance with our observation that gating does not require major conformation changes in L3, but a mere protrusion of an aromatic side chain in the channel lumen (here, W111 from Omp-Pst2). It needs to be acknowledged that fitting into our gating model mutagenesis data from anion-selective PhoE is less straightforward. While the critical residues are conserved (S19D Fig), mutagenesis data are pointing to the opposite direction, as discussed earlier [10, 15]. Mutation of residues involved in the attachment of the L3 tip to the barrel wall (E110C) or in the constitution of the acidic niche under L3 (E302C) indeed induce a decrease in Vc, while mutations of anti-L3 residues (R37C, R75C, R37C/R75C, K18C) all provoke an increase in Vc [15]. We believe that similar calculations need to be conducted on PhoE to provide an atomic level understanding of its possibly peculiar gating behaviour.
A tormenting question is whether voltage gating is a mere experimental artefact or underlies a functionally relevant regulatory mechanism. In early days, it was proposed that voltage gating could be a means by which porins that are mistakenly inserted into the inner membrane keep in the closed state [39, 40], because the Donnan potential of the outer membrane is ≤ -30 mV [39], while the inner membrane displays a transmembrane potential of about 160–200 mV (i.e. a value close to the Vc of most porins). Yet, the observation that Omp-Pst2 displays a strong propensity to gate (low Vc) suggests that this porin could, in the physiological context, rest in the gated state. That the gating propensity is asymmetric, and that this asymmetry correlates with the transportation of cations in an unfavourable direction—i.e. against the gradient of transmembrane potential—furthermore raises the question as to whether voltage gating conceals a role, for this porin, in the regulation of the cationic content of the periplasm. In humans, the primary habitat of P. stuartii, is the urinary tract, where the ammonium (NH4+) concentration is high. As P. stuartii features a urease activity that degrades urea into ammonia and carbonate, the higher propensity of Omp-Pst2 to transport cations from the intracellular to the extracellular side could participate in cleansing the bacterial periplasm of abounding cations. That Omp-Pst2 display a strong propensity to gate when the cationic flux occurs from the extracellular to the intracellular side furthermore suggests an active participation in the regulation of cationic fluxes across the outer-membrane. As new contributions of porins to bacterial development are discovered [41, 42], the functional meaning of voltage gating requires to be re-visited. In this context, molecular dynamics simulations hold the promise of allowing functional insights at the atomic level of resolution, as shown by the present work.
The crystal structure of Omp-Pst1 and Omp-Pst2 were solved at 3.3 Å and 2.2 Å respectively (PDB code 4D64 and 4D65 respectively). The crystal packing revealed an organization as dimers of trimers, viz. two symmetric, face-to-face trimmers. Chain A, B, and C of the crystal structures were taken out and used as starting models for the simulations. To create the lipid bilayer whereto embed our porins during the simulations, we used dimyristoylphosphatidylcholine (DMPC), as this lipid is of adequate length to model the thickness of a bacterial outer-membrane and has been used in a number of simulations studies on other bacterial porins [25, 43]. The bacterial outer-membrane is asymmetric, featuring lipopolysaccharides (LPS) on its extracellular leaflet. In our simulations, we did not use LPS, given that both experimental data collected on Omp-Pst1 and Omp-Pst2 reconstituted in LPS-containing bilayers and studies on other porins [44] established that the translocation properties of porins are not influenced by the presence of LPS. The trimeric structure of Omp-Pst1 and Omp-Pst2 were thus inserted into a pre-equilibrated lipid bilayer consisting of 512 DMPC by using the Perl script InflateGRO [45]. After the insertion, 58 DMPC were deleted, on the basis of steric clashes with the protein. In the final configuration, the area per lipid (APL) reached 60 Å2. Lipid-embedded Omp-Pst1 and Omp-Pst2 were then wrapped with 51,943 and 51,976 water molecules, respectively, in a cubic simulation box of 13.7 × 13.7 × 13.7 nm3. Potassium and chloride ions were added to reach a salt concentration of 1 M, and adjusted to provide neutral simulation systems. The standard protonation state at neutral pH was used for charged residues.
Molecular dynamics simulations were performed using GROMACS 4.5 package [46], and all-atom CHARMM force field for proteins [47] and lipids [48]. The TIP3P model [49] was used for water molecules. In all simulations, periodic boundary conditions were used in x, y and z directions. Electrostatic interactions were computed using the particle mesh Ewald (PME) method [50]. A Fourier spacing of 0.11 nm was used to avoid spurious drifts in the center of mass of the system [28, 30]. The LINCS method [51] was used to restrain bond lengths, allowing integration steps of 2 fs and updating of the neighbor list every 5 fs (cut-off distance of 1.2 nm). Lennard-Jones and Coulomb cut-off distances were set to 1.4 and 1.2 nm, respectively.
To prepare the simulation systems, we used the following procedure. First, the initial configurations of the lipid-embedded Omp-Pst1 and Omp-Pst2 were optimized by four steps of energy minimization, during which positional restrains were imposed on i/ all none-hydrogen atoms, ii/ main-chain atoms, iii/ Cα atoms and iv/ no atoms. Thereby, maximum forces lesser than 100 kJ.mol-1.nm-1 was attained. The two systems were then thermalized to 310 K in six steps of NPT ensemble, each lasting 500 ps. In the NPT ensembles, the pressure was kept constant at 1 bar independently on the x-y plane (containing the lipid bilayer) and the z-axis direction (normal to the lipid bilayer) by semi-isotropic coupling to a Parrinello-Rahman barostat with τP = 1.0 ps and a compressibility of 4.6x10-5 bar [52], while the temperature was maintained at the target temperatures (50 K, 100 K, 150 K, 200 K, 250 K, and 310 K) by weakly (τT = 0.1 ps) coupling lipids, protein and solvent separately to a V-rescale thermostat [53]. Each system was then subjected to another 1 ns NVT ensemble at 310 K. After equilibration, the two systems were coupled to a homogenous electrostatic field E aligned with the z-axis, allowing for the simulations with an artificial transmembrane potential (E = VTM/LZ) of either +1 or -1 V (extracellular to intracellular). As for controls, we also subjected the Omp-Pst1 and Omp-Pst2 simulation systems to NVT ensembles in the absence of an electrostatic field (VTM = 0 V). Simulations at VTM = 0 V were carried out for 100 ns while those at VTM ≠ 0 for 500 ns each. A recapitulation of the simulations is given in S1 Table.
Net ion permeation events and residency times within the channels were calculated using g_flux tool [54]. Ions that bound in niche I and II and the occupancy of the two niches were counted using the GROMCAS tool of g_select. Ions were required to move from one side of the lipid to the other to count as a complete permeation event. The selectivity and conductance of Omp-Pst1 and Omp-Pst2 were determined by calculating the current-voltage relationship under different applied electric fields. In each simulation, the current contributed either by K+ or Cl- ions was determined through linear regression of the net ion crossing at every time interval of Δt = 50 ns. At each voltage, the total current was the sum of K+ and Cl- currents, and conductances were calculated as the slope of the current-voltage curves [55]. Voltage-specific ion selectivity was calculated by using the ratio of K+ and Cl- current at each voltage. Simulated ion diffusion coefficients were calculated from our 100 ns simulations at VTM = 0. Diffusion coefficients obtained for K+ and Cl- were 1.83 cm2/s and 1.86 cm2/s in the simulation of Omp-Pst1, and 1.72 cm2/s and 1.95 cm2/s in the simulation of Omp-Pst2. The ratio of the experimentally determined diffusion coefficients (1.96 cm2/s for K+ and 2.02 cm2/s for Cl- [56]) to the calculated ones was used to get the corrected conductance. In practice, the K+ and Cl- currents were scaled by the ratio before calculation of the slope of the I-V curves. Ion densities within the channels (S4 Fig) were calculated using the density calculation module in MDAnalysis tool [57]. Hydrogen-bonding network analysis was performed using VMD (Figs 4A and 5A). Sequence alignments were performed using CLUSTALW. PhoE residue numbering was adjusted to fit the nomenclature used in previous papers [10, 15, 18].
Planar lipid bilayers were formed according to the monolayer technique of Montal and Mueller [58]. The bilayer was formed across a hole that was about 50 μm in diameter in a 25 μm thick polytetrafluoroethylene (PTFE) film. A lipid bilayer was prepared by spreading 1 μL of a 5 mg/mL solution of 1,2-diphytanoyl-sn-glycero-3-phosphocholine in a solvent mixture of n-pentane in the aperture. Ag/AgCl electrodes were used to detect the ionic currents. The electrode on the cis side of the cell was grounded, whereas the other one on the trans side was connected to the headstage of an Axopatch 200B amplifier. Purified detergent-solubilized porins (1 ng/mL) were added to the cis side of the chamber in presence of 1M KCl, 20mM PO4 pH 4 and inserted into the bilayer membrane by applying a 150–200 mV voltage. The recordings were made after diluting the same chamber with 1M KCl, 10mM HEPES pH 7.
|
10.1371/journal.ppat.1002334 | RNA Polymerase II Stalling Promotes Nucleosome Occlusion and pTEFb Recruitment to Drive Immortalization by Epstein-Barr Virus | Epstein-Barr virus (EBV) immortalizes resting B-cells and is a key etiologic agent in the development of numerous cancers. The essential EBV-encoded protein EBNA 2 activates the viral C promoter (Cp) producing a message of ∼120 kb that is differentially spliced to encode all EBNAs required for immortalization. We have previously shown that EBNA 2-activated transcription is dependent on the activity of the RNA polymerase II (pol II) C-terminal domain (CTD) kinase pTEFb (CDK9/cyclin T1). We now demonstrate that Cp, in contrast to two shorter EBNA 2-activated viral genes (LMP 1 and 2A), displays high levels of promoter-proximally stalled pol II despite being constitutively active. Consistent with pol II stalling, we detect considerable pausing complex (NELF/DSIF) association with Cp. Significantly, we observe substantial Cp-specific pTEFb recruitment that stimulates high-level pol II CTD serine 2 phosphorylation at distal regions (up to +75 kb), promoting elongation. We reveal that Cp-specific pol II accumulation is directed by DNA sequences unfavourable for nucleosome assembly that increase TBP access and pol II recruitment. Stalled pol II then maintains Cp nucleosome depletion. Our data indicate that pTEFb is recruited to Cp by the bromodomain protein Brd4, with polymerase stalling facilitating stable association of pTEFb. The Brd4 inhibitor JQ1 and the pTEFb inhibitors DRB and Flavopiridol significantly reduce Cp, but not LMP1 transcript production indicating that Brd4 and pTEFb are required for Cp transcription. Taken together our data indicate that pol II stalling at Cp promotes transcription of essential immortalizing genes during EBV infection by (i) preventing promoter-proximal nucleosome assembly and ii) necessitating the recruitment of pTEFb thereby maintaining serine 2 CTD phosphorylation at distal regions.
| Epstein-Barr virus (EBV) is associated with the development of a number of human cancers including Burkitt's lymphoma, Hodgkin's lymphoma, nasopharyngeal carcinoma and post-transplant lymphoproliferative disease. The virus infects B cells rendering them immortal through the production of a small number of viral proteins in the latently infected cell. Many of the viral proteins required for B-cell immortalization are produced from a very long protein-coding RNA message that initiates at the main viral latency promoter C, and our results provide important new information on how this message is produced. Specifically we show that the production of this long RNA is driven by the recruitment of the elongation factor (pTEFb) to paused transcription complexes at the C promoter. We show that pTEFb is recruited by the chromatin-associated protein, Brd4. Treatment of cells with a recently developed Brd4 inhibitor and inhibitors of the pTEFb elongation factor inhibits production of transcripts derived from the long EBV message thereby highlighting Brd4 and pTEFb inhibitors as potential anti-EBV agents.
| Epstein-Barr virus (EBV) is causally associated with the development of numerous tumours including Burkitt's lymphoma, Hodgkin's lymphoma, nasopharyngeal carcinoma and post-transplant lymphoproliferative disease and immortalizes resting B cells in vitro generating latently infected lymphoblastoid cell-lines (LCLs) [1]. LCLs express 9 viral latent proteins: EBV Nuclear Antigens (EBNAs 1, 2, 3A, 3B, 3C and LP) and three membrane proteins (LMP 1, 2A and 2B). Following initial infection, EBNA-LP and EBNA 2 are expressed from the viral W promoter (Wp). EBNA 2 then drives promoter switching through activation of the upstream viral C promoter (Cp) to produce a long message (up to 120 kb) that is differentially spliced to produce transcripts encoding all nuclear antigens required for immortalization [2]. EBNA 2 also activates two promoters that direct transcription of the EBV oncogene latent membrane protein 1 (LMP1) and the viral LMP 2A and 2B genes [3]–[4]. EBNA 2 is directed to promoters via association with the cellular DNA binding proteins RBP-Jκ and PU.1 [5]–[8]. Transcriptional activation by EBNA 2 involves the promotion of transcription initiation through associations with histone acetyltransferases [9], chromatin remodelling complexes [10]–[11], and the basal transcriptional machinery [12]–[14] and leads to Histone H3 and H4 acetylation at target gene promoters in vivo [15]. The association of EBNA 2 with target promoters is increased by asymmetric arginine dimethylation in the arginine-glycine repeat region of the protein [16] and is inhibited by phosphorylation on serine 243 during mitosis and viral lytic cycle [17]–[19].
The carboxy-terminal domain (CTD) of RNA Polymerase II (pol II) plays a central role in regulating efficient transcription initiation, elongation and RNA processing. It contains 52 heptapeptide repeats (Y1S2P3T4S5P6S7) and is phosphorylated largely on serines 2 and 5 during transcription [20]. Following pol II recruitment, promoter-proximal serine 5 CTD phosphorylation is mediated mainly by the TFIIH kinase, CDK7. Serine 2 CTD phosphorylation catalysed by CDK9/Cyclin T1 (positive transcription elongation factor b; p-TEFb) subsequently peaks at the 3′ end of genes. Using the specific inhibitors 5,6-dichloro-1-β-D-ribofuranosylbenzimidazole (DRB) and Flavopiridol, pTEFb has been shown to be required for productive elongation [21]–[22] by functioning as a CTD kinase and a regulator of the pol II-associated complexes DRB sensitivity-inducing factor (DSIF) and Negative Elongation Factor (NELF). DSIF and NELF induce promoter-proximal pausing that is relieved following the phosphorylation of DSIF, NELF and the pol II CTD by pTEFb [23]–[26]. Although NELF is localized to promoters and promoter-proximal regions in vivo, DSIF, in its phosphorylated form, continues to associate with pol II throughout genes and functions as a positive elongation factor capable of reducing polymerase stalling at pause sites, preventing transcript release and stimulating elongation [25], [27]. Interestingly, although initial studies in Drosophila documented the negative effects of NELF-induced promoter-proximal pausing on transcription of a subset of genes, including the Hsp70 locus [28]–[29], recent studies have demonstrated that the presence of NELF is required for the efficient transcription of the majority of Drosophila genes, forming a barrier to nucleosome assembly around the promoter [30].
Although pTEFb activity may be required for the efficient transcription of many cellular genes [22], not all gene transcription is pTEFb-dependent [31]–[32] and it is clear that many cellular and viral transactivators recruit and/or activate pTEFb to facilitate high-level gene-specific transcription elongation [33]. The bromodomain protein Brd4 can also recruit pTEFb to promoters via acetylated histones [34]–[35].
We have shown that EBNA 2 transcriptional activation requires pTEFb activity and promotes serine 5 CTD phosphorylation [36]. In this study we investigated how long-range EBV transcription required for viral immortalization is driven from the EBNA 2-responsive C promoter. Our results provide the first demonstration that significant levels of pol II accumulate specifically at Cp in association with the pausing factors DSIF and NELF and that pTEFb is recruited to the promoter at high level. Our data indicate that promoter-proximal pol II accumulation at Cp is directed by specific DNA sequences and maintains nucleosome depletion. Pausing facilitates pTEFb recruitment via Brd4 to drive high-level serine 2 CTD phosphorylation to promote production of the EBNA-encoding transcripts required for EBV immortalization.
We have previously demonstrated that EBNA 2 increases serine 5 CTD phosphorylation at viral latency promoter (Cp)-proximal and downstream regions and requires pTEFb for activation of both Cp and LMP1p [36]. To examine the role of EBNA 2 in facilitating transcription of the very long primary transcript (∼120 kb) encoding all EBV nuclear antigens (EBNAs) from Cp, we probed EBNA 2-driven changes in CTD phosphorylation at distal genome regions. We examined Cp transcription in a pair of EBV-positive Burkitt's lymphoma (BL) clonal cell-lines that either maintain the original EBNA 1-only (Latency I) BL tumour phenotype (Mutu I) or have drifted in culture to express the full panel of latent antigens including EBNA 2 (Mutu III) [37]. We found that EBNA 2 binding to Cp peaked around the RBP-Jκ site in Mutu III cells and was undetectable in Mutu I cells as expected (Figure 1B). EBNA 2-activated transcription in Mutu III cells resulted in increased serine 2 CTD phosphorylation which was evident from +295 and started to increase significantly in the W repeat region of the genome (typically 7.6 repeats located +666 to +24020 downstream from Cp), remaining high up to approximately 60 kb downstream (Figure 1C). In line with our previous observations in cells expressing conditionally active EBNA 2 [36], we found that EBNA 2-activated Cp transcription resulted in large increases in pol II recruitment and serine 5 CTD phosphorylation at promoter-proximal regions consistent with the promotion of transcription initiation (Figure 1D and E). In this study we found that increased serine 5 CTD phosphorylation was maintained at distal regions (Figure 1D). ChIP assays using an antibody that precipitates total pol II detected increases in the association of pol II with distal regions in Mutu III cells compared to Mutu I cells (Figure 1E), consistent with the promotion of transcriptional elongation by EBNA 2 to drive synthesis of the full panel of EBNAs expressed in Mutu III cells. Importantly, the observed changes in distal pol II CTD phosphorylation could not be accounted for by increased pol II presence alone, since increases in phospho-epitope levels exceeded the increases in total pol II (Figure 1E). We confirmed that distal serine 2 CTD phosphorylation required functional EBNA 2 using cells expressing a conditionally-active estrogen receptor-EBNA 2 fusion protein [38]; high level serine 2 CTD phosphorylation was detectable up to 75 kb downstream from Cp only in the presence of beta-estradiol (Figure S1).
Interestingly, our experiments revealed a large peak of pol II accumulation at Cp consistent with significant pol II stalling despite the fact that Cp was constitutively active (Figure 1E). In contrast, high levels of pol II were not detectable at the alternative promoter Q (Qp, +38800 downstream from Cp) that drives EBNA 1 transcription in Mutu I cells (Figure 1A & E), despite the fact that Q was fully active (Figure S2), indicating a lack of high-level pol II recruitment or stalling at Qp.
We have previously demonstrated that EBNA 2-activated transcription requires pTEFb activity [36]. Since CDK9 predominantly phosphorylates the pol II CTD on serine 2 during elongation through association with the travelling pol II complex [39], we examined pTEFb recruitment at Cp. ChIP assays using anti-CDK9 and anti-cyclin T1 antibodies demonstrated that high levels of both subunits of pTEFb were associated with Cp in Mutu III cells (Figure 2). Consistent with a role for pTEFb in distal serine 2 CTD phosphorylation, pTEFb was detectable in the W repeats and at 31 kb downstream (Figure 2) but fell to levels below the limits of detection of our ChIP assays thereafter. Previous studies have shown that pTEFb levels can drop significantly and be barely detectable using standard ChIP methods even 2 kb downstream from promoters, despite clear evidence of pTEFb function (i.e. Serine 2 phosphorylation) at these regions [40]–[41]. To further confirm that pTEFb was the kinase responsible for pol II CTD phosphorylation beyond 31 kb, Mutu III cells were treated with the pTEFb inhibitor, DRB. Our results demonstrated that DRB ablated serine 2 phosphorylation on the pol II CTD and severely reduced polymerase retention at distal regions (Figure 2). We also observed a reduction in pol II phosphorylation on serine 5 at distal regions, supporting previous observations of a role for pTEFb in catalysing serine 5 phosphorylation during elongation (Figure S3) [39]. ChIP for Tata box binding protein (TBP) confirmed that DRB treatment did not have general non-specific effects on Cp pre-initiation complex assembly (Figure S3). DRB treatment of cells expressing conditionally active EBNA 2 also confirmed the requirement for pTEFb for distal serine 2 CTD phosphorylation and pol II retention up to 75 kb downstream during EBNA2-dependent transcription (Figure S3).
To further investigate pol II stalling at Cp, we examined the association of the pausing complexes NELF and DSIF with the promoter. We detected high-levels of the NELF-A subunit of the NELF complex and the Spt5 subunit of the DSIF Spt4-Spt5 heterodimer at Cp in Mutu III cells (Figure 3) consistent with DSIF and NELF-induced polymerase stalling. Unlike NELF, which was absent from distal regions of the template, Spt5 remained detectable at distal regions consistent with a role for DSIF in promoting transcriptional elongation [25], [27] (Figure 3). EBNA 2-dependent pausing complex recruitment to Cp was confirmed in cells expressing conditionally active EBNA 2 (Figure S4). Our results suggest that recruitment of pTEFb to Cp is likely to be required to overcome stalling induced by DSIF and NELF and promote elongation to distal regions through serine 2 phosphorylation of the pol II CTD.
To determine whether polymerase stalling, high level pTEFb recruitment and large increases in serine 2 CTD phosphorylation were evident at other shorter EBNA 2-responsive transcription units, we performed ChIP assays using primers specific for the EBNA 2-activated LMP genes (Figure 4). Transcription of the LMP2A gene is regulated by EBNA 2-via two RBP-Jκ sites; EBNA 2-dependent LMP1 transcription is driven by a bidirectional promoter located in the reverse orientation in the EBV genome via EBNA 2 binding to both RBP-Jκ and PU.1 (Figure 4A). This bidirectional promoter also drives transcription of the LMP2B gene. The LMP2A and LMP1 transcription units therefore overlap and ChIP assays with primer sets 3–8 detect transcription complexes associated with either or both genes (Figure 4A). ChIP assays detected the same or higher levels of EBNA 2 binding to the LMP1 and LMP2A promoters in Mutu III cells to those detected at Cp (Figure 4B). Interestingly however, pTEFb recruitment to LMP promoters was barely detectable and no pol II stalling was evident at either LMP promoter (Figure 4). To rule out the possibility that we had failed to detect a pol II peak at LMP2Ap due to the location of our primer sets (−268 to −185 and +150 to +231), we designed an additional primer set spanning the transcription start site (−50 to +34). This primer set did not detect any higher pol II signal than the flanking primer sets (Figure S5). NELF and DSIF recruitment to the LMP locus was also minimal (Figures 4 and S6). Consistently, pol II CTD serine 2 phosphorylation did not reach the high levels observed at distal Cp regions and serine 5 phosphorylation on the pol II CTD was also much reduced (Figure S6). Similar results were obtained when Cp and the LMP locus were compared in an EBV-infected LCL (Figure S7).
To exclude the possibility that low-level pol II and transcription factor association with the LMP gene locus simply reflected low-levels of LMP transcription in the cell-lines under study, we used real-time PCR to determine the levels of Cp-initiated EBNA 2 and EBNA 1 transcripts and compared these to levels of LMP1 transcripts in Mutu III cells and two EBV-infected LCLs (Figure S8). We found that LMP1 transcript levels were equivalent to the levels of Cp-initiated EBNA 2 and EBNA 1 transcripts produced in the same cell-line, although there were variations in the level of transcripts produced between cell-lines probably as a result of differences in EBV genome copy number (Figure S8).
Taken together, our data indicate that pol II accumulation and high-level pTEFb recruitment is not a general characteristic of EBNA2- activated promoters, but is specific to Cp. Moreover, the level of promoter-associated pol II does not simply reflect the level of gene transcription from Cp and LMP1p.
Pol II stalling has recently been implicated in the promotion of gene activity through the maintenance of a promoter-proximal nucleosome-free region [30]. We therefore investigated whether the region around Cp was depleted of nucleosomes in the presence of stalled polymerase in Mutu III cells and an LCL where Cp is active, compared to Mutu I cells where Cp is inactive. Nucleosome levels were measured in ChIP assays using antibodies against the core histone, histone H3 [41]. Strikingly, we detected an 84% decrease in nucleosome occupancy at Cp in Mutu III cells compared to Mutu I cells using primer sets that spanned the region −208 to −96 bp upstream of the transcription start site and a 78% and 73% decrease with primer sets spanning regions +48 to +167 and −430 to −337, respectively (Figure 5). Nucleosomes were similarly depleted from these regions in an EBV-infected LCL (Figure 5). In contrast, levels of nucleosome depletion at similar regions around LMP2Ap and LMP1p were much lower, consistent with the absence of stalled pol II at these promoters (Figure 5). It is therefore clear that in the absence of Cp activity in Mutu I cells, nucleosomes assemble over promoter regions, but in the presence of stalled polymerase in Mutu III cells, Cp is maintained in a nucleosome-depleted state. In contrast, the low levels of pol II initiating at the LMP promoters are unable to maintain a highly nucleosome-depleted region and transient remodelling is likely to facilitate initiation.
Gene-specificity of polymerase stalling may be directed by the ability of promoters to recruit high levels of the general transcription factor TFIID and thus high levels of polymerase molecules [42]–[43]. Promoters that contain DNA sequences less favourable for nucleosome assembly may therefore be predicted to recruit TFIID and transcription complexes more efficiently and accumulate stalled pol II in association with DSIF and NELF. To test whether this could explain the specificity of polymerase stalling at Cp, we examined the propensity of the DNA sequences around Cp to assemble into nucleosomes using a nucleosome occupancy prediction program http://genie.weizmann.ac.il/software/nucleo_prediction.html [44]. This revealed a dramatic difference in the probability of nucleosome occupancy at Cp compared to the LMP promoters (Figure 6). The region of Cp encompassing the TATA signal appears much less likely to be occupied by nucleosomes compared to the equivalent regions of LMP1p and LMP2Ap (TATA boxes are located at −31 to −26, −32 to −27 and −28 to −23 at the C, LMP1 and LMP2A promoters respectively). Consistent with these predictions, ChIP assays using an anti-TBP antibody detected dramatically lower levels of TBP binding at the LMP promoters compared to Cp (Figure 6). We detected high-level TBP association around the Cp TATA box (−107 to −2) and upstream (−208 to −96) presumably as a result of cross-linked interactions between TBP (TFIID) and the transcription complex following initial TBP binding to the more accessible TATA signal. Our data are therefore in agreement with a model in which initial recruitment of high levels of pol II to Cp, presumably in association with the pol II binding factors NELF and DSIF, is driven by increased accessibility of the promoter to TBP. It is clear however, that in the absence of active Cp transcription in Mutu I cells, nucleosomes are able to assemble at Cp (Figure 5) and that the reduced probability of nucleosome occupancy may provide an initial advantage to pre-initiation complex assembly, but does not completely preclude nucleosome assembly. Moreover, the presence of stalled polymerase maintains nucleosome depletion further upstream and downstream from the Cp regions predicted to be less likely to be occupied by nucleosomes since primer sets spanning −430 to −337 and +295 to +406 detect reduced histone H3 levels (Figure 5).
Although pTEFb can be recruited to promoters via association with activators, we have been unable to demonstrate binding of EBNA 2 to pTEFb (data not shown). To investigate the mechanism of recruitment of pTEFb to Cp further, we examined the association of the pTEFb binding protein, Brd4, with the promoter. Brd4 is recruited via binding of its bromodomains to acetylated lysines in Histones H3 and H4 [45]. We detected large increases in Histone H3 and H4 lysine acetylation at Cp in Mutu III cells and recruitment of the Histone acetyl transferase p300, previously shown to interact with EBNA 2 [9] (Figure 7). Accordingly, Brd4 was recruited to Cp in Mutu III cells over a region spanning the peaks of histone acetylation (Figure 7). We next examined whether high-level Brd4 recruitment was Cp-specific. Our data demonstrated that Brd4 was also recruited to the LMP1 and LMP2A gene promoters consistent with the peaks of Histone H3 and H4 lysine acetylation and recruitment of p300 (Figure S9). Brd4 recruitment per se could therefore not account for Cp-specific pTEFb recruitment (Figures 2 and 4). Interestingly however, experiments carried out in the presence of cellular stress revealed that pTEFb recruitment to Cp correlates with Brd4 binding. In the presence of cellular stress, such as that induced by exposure to Actinomycin D, DRB or UV, pTEFb is released from an inactive pool, where it is complexed with 7SK snRNA and the HEXIM1 protein, as part of a stress response aimed at increasing transcription factor availability. Released pTEFb then associates with Brd4 and the levels of the pTEFb/Brd4 complex are increased [34], [46]–[47]. DRB treatment has been previously shown to result in a 2-fold increase in the level of pTEFb/Brd4 complexes [34]. Consistent with these observations, we found that treatment of Mutu III cells with DRB resulted in a two-fold increase in both Brd4 and pTEFb recruitment to Cp indicating that Brd4 is responsible for recruiting pTEFb to Cp. Histone H4 acetylation was increased by DRB treatment at Cp, perhaps as a result of the protection from deacetylation provided by the preferential binding of Brd4 to acetylated Histone H4 residues. Previous studies have described inducible Brd4 recruitment via acetylated histone H4 but not acetylated histone H3 residues [48] and our data indicate that the pattern of Brd4 binding more closely resembles the profile of histone H4 rather than histone H3 acetylation (Figures 7 and S9). In sharp contrast, DRB treatment led to loss of Brd4 from the LMP1 promoter and decreases in Histone H3 and H4 acetylation (Figure 8G-J) (pTEFb is not detectably recruited to LMP1; Figure 4 and S5). Since the key difference between the Cp and LMP1 promoters is the presence of high levels of stalled pol II at Cp, these results suggest that pTEFb is efficiently recruited to Cp via Brd4 as a result of stable interactions between the pTEFb/Brd4 complex and the large numbers of stalled polymerases present at the promoter. Thus at LMP1p, in the absence of an accumulation of pol II molecules, pTEFb complexes brought in by Brd4 have little polymerase with which to stably associate and Brd4/pTEFb complex binding is not stabilized.
Interestingly, the essential EBV replication and transcription factor EBNA 1 has been shown to recruit Brd4 to a region of the latent origin of replication (OriP), that functions as an EBNA 1-dependent Cp enhancer [49]–[50]. We therefore investigated the possibility that EBNA 1 may recruit pTEFb to OriP via Brd4 and contribute to the level of pTEFb at Cp through DNA looping effects. Since EBNA 1 is expressed in Mutu I and Mutu III cells, Brd4 would be expected to be recruited to OriP by EBNA 1 in both cell types. ChIP analysis in Mutu I and Mutu III cells using primers sets close to the family of repeats (FR) element in Ori P where Brd4 was previously detected [49] revealed some Brd4 binding to Ori P in Mutu I and Mutu III cells, equivalent to that detected in the GAPDH gene (Figure S10). The level of Brd4 detected was however much lower than that present at Cp and did not appear to result in significant recruitment of pTEFb to this region of the genome (Figure S10). Our data therefore indicate that it is unlikely that pTEFb recruitment via Brd4 at OriP contributes to the level of pTEFb at Cp.
We next sought to obtain direct evidence that Brd4 binding is required for Cp but not LMP1 transcription by treating Mutu III cells with the novel small molecule Brd4 bromodomain inhibitor, JQ1, previously shown to block Brd4 association with acetylated histones [51]. Strikingly, treatment with 50 nM JQ1 for 48 hrs reduced levels of Cp-initiated EBNA 2 and EBNA 1 transcripts by 74% and 65% respectively, but had no effect on LMP1 transcript levels (Figure 9A). ChIP analysis confirmed that JQ1 dramatically inhibited Brd4 association with Cp promoter regions (Figure 9). The loss of Brd4 resulted in a significant decrease in pTEFb association with Cp, consistent with Brd4-dependent recruitment of pTEFb to Cp (Figure 9). In summary, our data indicate that the binding of Brd4 to Cp is required for Cp transcription since it facilitates the stable association of pTEFb with the stalled polymerases present at Cp.
Since inhibition of Brd4 binding was sufficient to selectively inhibit Cp transcription presumably through reduced pTEFb recruitment, we investigated the effects of pTEFb inhibitors on Cp and LMP transcription in Mutu III cells. We have previously demonstrated that EBNA 2 activation of both Cp and LMP1 reporter constructs was inhibited by treatment with DRB or overexpression of a dominant negative form of the pTEFb kinase, CDK9 [36]. However, our current study indicates that LMP promoters in vivo show little detectable pTEFb recruitment (Figure 4). Consistent with the selective high-level recruitment of pTEFb to Cp in vivo, we found that the pTEFb inhibitors DRB and Flavopiridol were both able to inhibit Cp transcription at concentrations at which LMP1 transcription was unaffected (Figure 10). The discrepancy between our previous results and these observations is likely explained by the fact that the promoter context in transiently transfected reporter constructs differs significantly from the appropriately assembled chromatin structures found at promoters actively engaged in transcript production in latently infected cells. Our data indicate that a reduced propensity for nucleosome assembly around Cp allows high level recruitment of TFIID and establishes polymerase pausing at the constitutively active C promoter in infected cells. These Cp-specific features may not have been established in transient assays. Thus pTEFb may be important for EBNA 2-dependent Cp and LMP promoter activity in reporter assay systems, but differences in pTEFb requirements are evident in the context of latently infected cells.
EBV relies on the transcription of a long polycistronic mRNA to encode the nuclear antigens (EBNAs) essential for immortalization. Following initial production of EBNA-LP and EBNA-2 from a cellular factor-driven promoter (Wp) after primary infection [52]–[54], EBNA 2 activates an upstream promoter (Cp) leading to long-range transcription and the full panel of EBNA expression [2]. The molecular mechanisms behind this strategy have not been fully elucidated. Our data show that the necessity for this promoter switch goes beyond the simple advantage of utilizing a virally-controlled promoter, and may reflect a requirement to promote efficient transcriptional elongation ensuring production of the long (approximately 120 kb) primary message. Our results indicate that the specific recruitment of high levels of pTEFb to Cp in the presence of EBNA 2 is required to promote distal transcriptional elongation through serine 2 CTD phosphorylation and to overcome promoter-proximal pol II stalling induced by the high levels of the NELF/DSIF complex present at Cp. Our data suggest that Cp-specific pol II stalling may play dual positive roles in promoting transcription (i) by triggering the recruitment of pTEFb and promoting distal elongation and (ii) by maintaining a nucleosome-free region at the promoter that promotes initiation.
Our results document the presence of stalled RNA polymerase at an actively transcribing viral gene locus, unlike the situation observed at heat-shock genes, where genes temporarily in the ‘OFF’ state maintain promoter-proximally paused pol II to enable a rapid transcriptional ‘ON’ response to heat-shock that results in a re-distribution of polymerase along the gene [55]. Paused polymerase does not remain detectable when Cp is ‘OFF’ in the Mutu I cells used in this study because Cp is silenced in EBV positive Latency I cells through CpG DNA methylation, thus inhibiting transcription factor binding and pol II recruitment [56]–[57]. Recent fine-mapping confirmed Cp methylation in Mutu I cells and demonstrated a peak of 5-methyl cytosine close to Cp that increased 7-fold in Mutu I cells compared to an LCL generated from Mutu virus, where Cp is active [58].
Significantly, we demonstrate that regulation of Cp is distinct from the regulation of the latent membrane protein promoters, where only low levels of pol II, pTEFb, NELF and DSIF are detectable and serine 2 CTD phosphorylation does not substantially increase at distal regions. The EBNA 2-dependent LMP1, LMP2A and 2B genes encode transcripts of 2.8 kb, 11.7 and 8.4 kb in length (in the B95-8 EBV genome sequence NC_007605.1) so these shorter transcription units may therefore be less dependent on elongation factor function. It is worth noting that the LMP 2 transcription units can increase in size due the presence of varying numbers of ∼500 bp terminal repeat (TR) elements that are present within these genes. On entry into host cells, the EBV genome is initially in its linear form and the TR region is the site of recombination-directed genome circularization. Although the B95-8 genome sequence we have used for transcript annotation contains 4 TRs spanning 2.1 kb, TR regions of up to ∼12 kb have been reported, indicating that LMP2A transcripts can be up to 25 kb in length [59]. Nonetheless, the ∼120 kb Cp transcription unit requires pol II to elongate over considerably longer distances and our results indicate that it possesses distinct regulatory features that promote long-range transcription.
The specificity of pol II stalling at Cp appears to be driven by the presence of DNA sequences upstream of Cp that are less favourable for nucleosome assembly. These sequences encompass the TATA box and therefore allow increased access to TBP resulting in high-level recruitment of pol II in association with the pausing factors NELF and DSIF. Once pol II stalling is established at Cp, a more extensive region around the promoter is then maintained in a nucleosome-depleted state. We have previously demonstrated that pTEFb activity is required for the activation of both Cp and LMP1 promoter-reporter constructs by EBNA 2 [36]. In the present study, the pTEFb components cyclin T1 and CDK9 were virtually undetectable at either the LMP1/2B or LMP2A promoters. In the context of latently infected EBV immortalized cells we now show that endogenous Cp transcription can be selectively inhibited by the pTEFb inhibitors DRB and Flavopiridol at concentrations that do not affect LMP1 transcript production. It is likely that chromatin structure in our previous transient reporter constructs differs significantly from the chromatin context present in proliferating infected cells in vivo and thus pausing factor and elongation factor requirements at Cp may not have been faithfully recapitulated. The results presented in this manuscript show that nucleosome occupancy is likely to be the critical determinant that sets up Cp-specific polymerase pausing and pTEFb recruitment on endogenous EBV templates. It would be interesting to test whether Cp-specific regulatory features can be conferred by Cp promoter sequences alone by generating recombinant viruses in which LMP promoter regions are replaced with Cp promoter regions.
Since the studies described here have exclusively examined the nature of constitutive Cp transcription in established cell-lines, it will be interesting to investigate the kinetics of the establishment of polymerase stalling during primary B-cell immortalization, when transcription switches from Wp to Cp approximately 6 days post-infection [60], and the effects promoter switching has on CTD phosphorylation and the elongation properties of pol II. Interestingly, when we extended our nucleosome prediction analysis to include Wp, we found that the region around the Wp TATA box has a high probability of being occupied by nucleosomes, similar to the results obtained for the LMP promoters (Figure S11). Thus, based on our findings, Wp would be less likely to accumulate stalled polymerases as a result of increased TFIID access and pol II recruitment. The switch from Wp to Cp usage may therefore be advantageous for the virus, enabling high level pTEFb recruitment and increased efficiency of elongation. It is interesting however, that Cp-deleted viruses capable of transforming B-cells have been described and more recently rare Burkitt's lymphoma cells were identified that exclusively use Wp to drive EBNA transcription [61]–[62]. These observations suggest that Wp may be able to achieve sufficient long-range transcription required for EBNA production during immortalization in vitro or in certain cell backgrounds. Importantly, the W promoter is present in multiple copies in the EBV genome (e.g. 7.6 copies in the prototype Type 1 EBV strain, B95-8) and transcription initiation from a number of W promoters may be required to generate sufficient levels of downstream transcripts. The fact that Cp deletion is a rare event however, supports the notion that Cp plays a crucial role during EBV immortalization and infected cell growth in vivo.
Since we detected the presence of Spt5 at distal regions, it is also possible that the recruitment of DSIF may play a positive role in Cp transcription, as documented for HIV transcription. pTEFb-mediated phosphorylation of the Spt5 subunit of DSIF at promoter-proximal regions converts DSIF into a positive-acting elongation factor that travels with polymerase to promote processivity and inhibit further pausing [25], [27]. Spt5 has been shown to promote transcriptional activation by Gal4-VP16 and is recruited to the HIV-1 LTR to co-operate in the stimulation of transcriptional elongation by HIV-1 Tat [25], [63]. Further experiments involving RNA interference would be useful in dissecting the roles of DSIF in the regulation of Cp transcription.
EBV strains can be classified into two virus types (1 and 2, formerly A and B) largely based on sequence differences between the EBNA 2 genes, which share only 55% homology. Despite the prevalence of type 2 viruses in Africa and their association with BL, type 2 viruses transform resting B-cells much less efficiently than type 1 strains and differences between the EBNA 2 genes appear to be the major determinant of this property [64]. Recent work has identified cellular genes that are differentially regulated by type 1 and type 2 EBNA 2 indicating that reduced gene activation by type 2 EBNA 2 may contribute to the reduced transforming potential of type 2 viruses [65]. Since type 2 EBNA 2 also initially activates the LMP1 promoter to a reduced extent [65], it is conceivable that differences in Cp control by type 1 and type 2 EBNA 2 may also be evident. However, the results described here indicate that high-level pTEFb recruitment to Cp is driven by polymerase stalling initiated by DNA sequences that promote reduced nucleosome occupancy around Cp, increasing access to TFIID, rather than through specific properties of EBNA 2 (that may vary between strains). To investigate whether type 2 Cp sequences also possessed this property, we performed nucleosome occupancy predictions using the sequence of the type 2 viral strain AG876. We found that nucleosome occupancy at the type 2 C, LMP and W promoters was predicted to be virtually indistinguishable (data not shown) from that of the type I B95-8 strain previously examined (Figure 6A and S11). This is perhaps not surprising given the sequence similarity between the two strains throughout most of the genome. Thus type 2 Cp sequences show the same reduced likelihood of nucleosome occupancy as type 1 viruses, compared to the respective LMP and W promoter sequences, and would be just as likely to accumulate stalled polymerases and recruit high-levels of pTEFb. It is therefore unlikely that pTEFb recruitment would contribute to the reduced transforming potential of type 2 viruses, but further work is necessary to address this point unequivocally.
Since we have been unable to detect an interaction between EBNA 2 and pTEFb to date, we investigated alternative mechanisms for the recruitment of pTEFb to Cp. The double bromodomain protein, Brd4, has been shown to bind to the active form of pTEFb and recruit it to promoters to stimulate elongation [34]–[35]. Our data indicate that pol II stalling facilitates the association of Brd4-recruited pTEFb with the C promoter by providing high levels of pol II with which pTEFb can associate. In support of this hypothesis the interaction of pTEFb with Brd4 has been shown to be weak in nature and is disrupted by low salt concentrations [34]. Thus despite similar levels of Brd4 binding to Cp and the LMP gene loci, the lack of stalled pol II at LMP does not facilitate stable pTEFb binding to the transcription complex. Brd4 has been shown to play important roles in regulating viral transcription and in tethering viral genomes to chromatin [66]. Brd4 enhances HIV-1 transcription and promotes transcriptional activation of G1/S cyclin genes by murine gammaherpesvirus 68 (MHV-68) through direct interaction with MHV-68 orf73 [35], [67]. Brd4 also plays an important role in the repression of human papillomavirus transcription by the viral E2 protein and tethers bovine and human papillomavirus genomes to mitotic chromosomes [68]–[69]. Brd4 was also recently shown to bind to the EBV latent antigen EBNA-1 and to play a role in EBNA-1 activation of transcription; knock-down or overexpression of Brd4 inhibited EBNA-1 activated transcription in reporter assays [49]. It therefore appears that Brd4 may play multiple roles in the EBV life-cycle. Our data demonstrating specific inhibition of Cp-driven EBV transcription by the novel Brd4 inhibitor JQ1 highlights the potential for drug-like derivatives of this compound as anti-EBV agents. In addition, our further evidence for the role of pTEFb in promoting EBV transcription and the inhibition of Cp transcription by pTEFb inhibitors adds weight to the possible use of the pTEFb targeting anti-cancer drug, Flavopiridol, in the treatment of EBV-associated tumours.
In summary, we demonstrate that polymerase stalling may play a role in facilitating immortalization by the tumour virus EBV. High-level recruitment of pol II and associated pausing factors to the viral C promoter maintains nucleosome depletion and necessitates pTEFb recruitment to overcome pausing. This provides high levels of pTEFb to promote the distal serine 2 CTD phosphorylation required for production of the long viral transcript encoding key EBV immortalizing genes.
ER/EB 2.5 cells [38] were maintained as described previously [36]. Mutu I (clone 179), Mutu III (clone 48), IB4 (provided by Martin Rowe), PER 142 B95-8 LCL and PER 253 B95-8 LCL (provided by Heather Long) were cultured as described [70]. For Brd4/pTEFb inhibition experiments, Mutu cells were resuspended at 5×105 cells/ml and incubated in the presence of DMSO, JQ1/SGCBD01 [51] (kindly provided by Stefan Knapp, Structural Genomics Consortium, University of Oxford), DRB (Sigma) or Flavopiridol (Sigma) for 24 or 48 hours.
ER/EB 2.5 cells were washed and resuspended at 5×105 cells/ml in medium without β-estradiol. After 3 days 1 µM β-estradiol (Sigma) was added for 5 hours and chromatin prepared as described previously [36]. ER/EB 2.5 cells were treated with 100 µM DRB (or DMSO as a control) for 2 hrs as required prior to addition of β-estradiol.
Mutu cells were diluted to 5×105 cells/ml 24 hrs prior to chromatin preparation and resuspended at 1×107 cells/mL in fresh media before crosslinking. Cells were treated with 500 µM DRB for 2hrs prior to chromatin preparation.
ChIP methods were optimised for each target using a number of alternative strategies.
For ER/EB 2.5 cells ChIP assays were carried out as described previously [36] by overnight incubation at 4°C with 6 µg of polyclonal antibodies (anti-Pol II; N-20, anti-Spt5; H-300, Santa Cruz Biotechnology, Inc) followed by precipitation with protein A sepharose beads pre-blocked with salmon sperm DNA. EBNA 2 immunoprecipitations were carried out using 8 µg of monoclonal antibody (PE2) and an additional incubation with secondary antibodies [36]. DNA was purified using the QIAquick Gel extraction Kit (Qiagen) and eluted in 110 µl sterile millipore water. Phospho serine 2 immunoprecipitations in ER/EB 2.5 cells were carried out using a double-round ChIP protocol immunoprecipitating first pol II and then the phospho-specific form. Immune complexes from pol II precipitations were eluted and diluted by addition of 850 µl IP dilution buffer. Second round immunoprecipitations were carried out using protein sepharose A/G beads (1∶1 mix of protein A and G sepharose) preabsorbed first with rabbit anti-mouse IgM immunoglobulins (20 µg) in 500 µl IP dilution buffer overnight, and then with 25 µg anti-phospho ser 2 (H5) for 3–5 hours at 4°C. Prior to collection of immune complexes, 100 µl of a 50% slurry of antibody pre-coated beads were blocked using 350 µg salmon sperm DNA for 1 hr at 4°C. Immune complexes were collected by rotation at 4°C overnight.
ChIP assays for EBNA 2 using Mutu cell chromatin were carried out as for ER/EB 2.5 cells using 8 µg (PE2) antibodies. Pol II, Spt5, acetylated Histone H3 and acetylated Histone H4 immunoprecipitations were carried out as described previously [36] by overnight incubation of chromatin lysates with 5 µg of anti-pol II, anti-Spt5 (H-300), anti-CDK9 (H-169), anti-Cyclin T1 (H-245), anti-Brd4 (H-250) (Santa Cruz Biotechnology, Inc), anti-acetyl-Histone H3 or H4 (Millipore) antibodies. ChIP assays for core Histone H3 were carried out using 2 µg anti-Histone H3 antibody (Abcam, ab1791). For NELF-A, immunoprecipitations were carried out using a polyclonal antibody (Santa Cruz Biotechnology, Inc) and precoating protein A/G sepharose beads with 5 µg (anti-NELF-A; A20) antibody overnight. Immune complexes were collected overnight following blocking of pre-coated beads with salmon sperm DNA as above. Phospho serine 2 and 5 immunoprecipitations using Mutu cell chromatin were carried out in a single round ChIP by precoating protein A/G sepharose beads with 10 µg rabbit anti-mouse IgM overnight, prior to the addition of 25 µg H5 or 5 µg H14 antibodies and then salmon sperm DNA as above. All controls were treated identically but without addition of antibodies.
Cells were diluted to 5×105/ml, harvested after 24 hrs and total RNA extracted using TriReagent (Sigma). RNA samples were purified using the RNeasy kit (Qiagen) and cDNA was then synthesised using the ImProm II reverse transcription system using random oligonucleotides (Promega). For Brd 4 and pTEFb inhibitor experiments, cDNA was prepared from 105 cells using Power SYBR Green Cells-to-CT Kit (Applied Biosystems) according to the manufacturer's instructions.
Quantitative PCR (Q-PCR) was performed as described previously [36] using an Applied Biosystems 7500 real-time PCR machine (95°C for 10 mins, 40 cycles at 95°C for 15 sec and 60°C for 1 min and dissociation curve analysis). For ChIP analysis, an input control standard curve was generated for each primer set (Table S1). Generally, cDNA samples were analysed using the absolute quantitation method with standard curves generated from Mutu I or Mutu III cDNA. Transcript levels were determined using Qp or Cp-specific primers [71], cDNA-specific EBNA 2, EBNA 1 (YUK) or LMP1-specific primers [72] and either the 18S rRNA Quantitect primer assay (Qiagen) or actin primers as normalization controls (Table S1). For Brd 4 inhibition experiments, Q-PCR was carried out using Power SYBR Green Cells-to-CT Kit (Applied Biosystems) and cDNA-specific EBNA 2, EBNA 1 (YUK) or LMP1-specific primers [72] with actin as the endogenous control and analysed by Relative Quantification (ddCt).
SDS-PAGE analysis and immunoblotting was carried out as described previously [36], [70]. Blots were probed with human M.S. serum at 1/200 to detect EBNA 1 (gift from Martin Rowe), PE2 at 1/300 to detect EBNA 2 and anti-actin at 1/5000 (A-2066, Sigma). HRP-conjugated anti-mouse (Dako) or anti-rabbit antibodies (Cell Signalling Technology) were used to detect EBNA 2 and actin respectively, and HRP-conjugated protein A (1/1000, Amersham Biosciences) was used to detect EBNA 1 primary antibodies.
The type 1 EBV genome used for primer design, transcription start sites and nucleosome predictions is the annotated sequence from the B95-8 strain (NC_007605.1). The type 2 EBV genome used was from the AG876 strain (NC_009334.1).
|
10.1371/journal.ppat.1005508 | PI3Kγ Is Critical for Dendritic Cell-Mediated CD8+ T Cell Priming and Viral Clearance during Influenza Virus Infection | Phosphoinositide-3-kinases have been shown to be involved in influenza virus pathogenesis. They are targeted directly by virus proteins and are essential for efficient viral replication in infected lung epithelial cells. However, to date the role of PI3K signaling in influenza infection in vivo has not been thoroughly addressed. Here we show that one of the PI3K subunits, p110γ, is in fact critically required for mediating the host’s antiviral response. PI3Kγ deficient animals exhibit a delayed viral clearance and increased morbidity during respiratory infection with influenza virus. We demonstrate that p110γ is required for the generation and maintenance of potent antiviral CD8+ T cell responses through the developmental regulation of pulmonary cross-presenting CD103+ dendritic cells under homeostatic and inflammatory conditions. The defect in lung dendritic cells leads to deficient CD8+ T cell priming, which is associated with higher viral titers and more severe disease course during the infection. We thus identify PI3Kγ as a novel key host protective factor in influenza virus infection and shed light on an unappreciated layer of complexity concerning the role of PI3K signaling in this context.
| Acute respiratory viral infections like influenza virus can cause life-threatening disease in infected individuals. Phosphoinositide-3-kinases have been suggested to be important factors used by the virus to infect and replicate in host cells, and thereby cause viral pneumonia. However, to date the role of these signaling molecules has not been thoroughly addressed in the context of an infection in whole animals, rather than just cell culture systems. Here we show that one of the PI3K subunits, PI3Kγ, is in fact critically required for the clearance of the infection. This is because PI3Kγ regulates the immune response against the virus through the generation and maintenance of antiviral CD8+ T cell responses. We show that in the absence of PI3Kγ a specialized dendritic cell subset in the lung is deficient and this leads to a strongly impaired immune response against influenza virus. We thus identify PI3Kγ as a novel host molecule that is important for the immune defense against influenza virus infection
| Phosphoinositide 3-kinases (PI3K) are classified into three main groups (class I, class II and class III) according to sequence homology of the catalytic subunit and their substrate specificity [1]. Class I PI3K are further divided into class IA and class IB. Class IA PI3K form dimers consisting of either one of the catalytic subunits p110α, p110β or p110δ, and the common regulatory subunit p85 [2] [3] [4] [5]. They typically act downstream of receptor tyrosine kinases and are important regulators of cell growth, division and survival [6]. In contrast, class IB PI3K (also termed PI3Kγ) comprises only one catalytic subunit, p110γ, which associates with the regulatory subunits p101 or p84 [7] [8] [9] [10] [11]. PI3Kγ signals downstream of G-protein coupled receptors (GPCR) such as chemokine receptors or receptor tyrosine kinases [12]. Both class IA and PI3Kγ can be activated by ras [13] [14]. Classes II and III PI3K are ubiquitously expressed and mainly involved in regulation of protein trafficking and cell homeostasis. PI3Kγ on the other hand is preferentially expressed in hematopoietic cells, although expression was also shown in peribronchial epithelial cells, the endothelium, the brain and the heart [15] [16].
Several groups have addressed the role of PI3Kγ in immune responses using specific inhibitors or p110γ-deficient mice. Neutrophils and macrophages, which are p110γ-deficient, exhibit reduced migration in vitro in response to chemotactic stimuli such as IL-8 and MIP-1α as well as the GPCR agonists C5a and fMLP [17]. Consistently, in vivo recruitment of neutrophils and macrophages to inflamed peritoneum is severely impaired in p110γ-/- animals upon peritoneal infection with Listeria monocytogenes [18]. In addition to the defects observed in innate immune cells, PI3Kγ-deficiency results in impaired adaptive immune responses. PI3Kγ-signaling in conjunction with PI3Kδ, plays a minor role in thymocyte as well as B cell development and the absence of PI3Kγ leads to a small reduction of peripheral CD4+ but not CD8+ T cells [19]. Addressing the migration capacity of lymphocytes, it was shown that PI3Kγ is superfluous for T cell homing in steady-state conditions [20]. Under inflammatory conditions however, PI3Kγ-/- mice display a reduced recruitment of CD8+ T cells. Peritoneal infection with Vaccinia virus or Lymphocytic Choriomeningitis virus infection into the footpads result in decreased numbers of CD8+ T cells at the site of inflammation in PI3Kγ-/- mice [21] [22]. In both studies, PI3Kγ-/- CD8+ T cells exert normal effector functions in terms of Interferon-γ (IFN-γ) production and cytoxicity. In contrast to T cells, B cells develop normally in PI3Kγ-/- mice and do not show any deficiency in migration [20]. More recently, we could show that PI3Kγ is required for development of lung CD11b+ DC and CD103+ DCs in particular by regulation of signaling downstream of Flt3, while it is dispensable for DC development in many other tissues [23]. In line with this data, several reports have revealed a central role for PI3Kγ in murine models of human immune-mediated inflammatory diseases such as rheumatoid arthritis and airway inflammation as well as autoimmune diseases such as systemic lupus [24] [25] [26]. Therefore, PI3Kγ is considered a promising target for the treatment of inflammatory disorders [27].
Using chemical inhibitors such as Wortmannin, PI3K family members or effectors downstream such as Akt kinases were shown to be required by influenza virus for infection of lung epithelial cells in vitro [28] [29], in particular through interactions with the viral protein NS1 [30]. Furthermore, Influenza virus strains carrying mutations rendering them unable to activate PI3K signaling were shown to lead to attenuated infection in vitro and in vivo [30]. However, the importance of PI3K signaling for host defense as well as the specific roles of individual PI3K subunits for influenza virus infection in vivo, remain poorly understood. In this context PI3Kγ itself has not received much attention in the context of influenza infection. Given the defects in innate and adaptive immunity in PI3Kγ-deficient mice and its potential direct involvement in influenza virus pathogenesis, we investigated the role of PI3Kγ-signaling upon infection with influenza virus in vivo. We found that PI3Kγ-deficiency led to greatly enhanced susceptibility to influenza virus infection due to delayed viral clearance. This was caused by impaired T cell priming by lung resident dendritic cells due to a pre-existing developmental deficiency in the lung dendritic cell compartment of PI3Kγ-deficient animals. We thus describe a novel role of PI3Kγ in regulating host responses against respiratory viral infections.
To address the role of p110γ in Influenza A virus infection (IAV) in vivo we infected p110γ kinase–dead (p110γ-KD) animals with a sub-lethal dose of the highly pathogenic strain IAV PR8. These animals carry an inactivating mutation in the kinase domain of p110γ and thus allow us to delineate the role of p110γ kinase function during IAV infection in vivo. Monitoring weight and temperature loss over time, we observed a much more severe disease course in p110γ-KD animals as opposed to WT controls with a more pronounced temperature and weight loss (Fig 1A and 1B). Furthermore, at a fourfold higher dose (i.e. 200 pfu), half of the p110γ-KD mice succumbed to infection by day 15, while 100% of WT animals survived (Fig 1C). The enhanced morbidity observed in p110γ-KD mice infected with 50 pfu was paralleled by a delayed viral clearance at later time-points of infection, where p110γ-KD animals had higher viral titers at day 9 and 11 p.i, while there was no difference at day 5 p.i (Fig 1D). In addition, the high viral load in lungs of p110γ-KD animals at later points during the infection correlated with an increased proteinosis and cell death in the alveoli, exemplified by higher levels of total protein and higher number of dead cells in the bronchoalveolar lavage (BAL) of p110γ-KD mice at day 11 (Fig 1E and 1F). Examining the immune cell infiltrate in the lungs of infected animals at an early time point of infection it was apparent that recruitment of monocyte-derived dendritic cells (moDC) natural killer (NK) cells and neutrophils was completely intact despite p110γ kinase-deficiency (Fig 1G and 1H). Similarly, numbers of tissue-resident alveolar macrophages (AM) were comparable between WT and p110γ-KD animals at day 3 p.i. (Fig 1H). Finally, also levels of hallmark inflammatory cytokines TNFα and IL-1β were similar in the BAL of both mouse strains (Fig 1I). Taken together these results suggested that p110γ plays an important role in host defense against IAV but that the early antiviral response is largely intact.
To characterize the adaptive immune response against IAV we again infected p110γ-KD animals with a sub-lethal dose of PR8 IAV and then characterized immune cell infiltration into the BAL and lung at day 7 p.i., which represents the initiation phase of the adaptive immune response. Both CD4+ and CD8+ T cells were readily detected in BAL and lung at this time point (Fig 2A). Quantifying the total number of T cells in the lungs of infected mice it was evident that p110γ-KD mice exhibited a significantly reduced number of both CD4+ and CD8+ cells present (Fig 2B). Most strikingly, nucleoprotein—34-specific (Tet+) CD8+ virus-specific cells were virtually absent in the lungs of p110γ-KD animals. Similarly, the proportion of IFNγ-producing CD4+ and CD8+ cells in the BAL was lower in mice deficient for p110γ kinase function (Fig 2C and 2D). Furthermore, in the lung draining lymph node (dLN) CD4+, CD8+ and virus specific CD8+ T cells were all considerably reduced in the p110γ- defective condition compared to WT animals (Fig 2E). Conversely, no difference was observable in the fold increase in the number of activated CD4+ and CD8+ T cells in the lungs of infected compared to naïve animals (Fig 2F). Similarly, inflammatory cytokines IL-1β and IL-6 in the BAL of infected mice (Fig 2G) were comparable between WT and p110γ-KD mice. In addition, the antiviral B cell response appeared to be completely intact as no difference in antibody titres of different isotypes in the BAL could be observed between WT and p110γ-KD mice (Fig 2H) at day 11 p.i. To address the possibility that the deficient T cell response observed in p110γ-KD mice was due to a T cell defect in naïve animals, T cells were examined in blood and lung in the steady-state. This analysis revealed that the number of T cells in the lung as well as the frequency in the blood of naïve mice was comparable between WT and p110γ-KD animals (S1A and S1B Fig). Furthermore, the frequency of CD44+CD62L- activated T cells was similar regardless of p110γ-deficiency (S1C and S1D Fig). To exclude the possibility that p110γ generally regulates development of hematopoietic cells, the composition of blood of p110γ-KD animals was carefully analysed. Granulocytes such as neutrophils are present at normal levels in p110γ-KD mice (S1E Fig) and similarly the number of red blood cells is also comparable to WT animals in p110γ-KD mice (S1F Fig). Overall, these results suggested an important and specific role for p110γ in regulating the antiviral T cell response, in particular the CD8+ T cell component but that p110γ is largely dispensable for T cell development and activation in the periphery.
To address a potential direct involvement of p110γ in influenza virus propagation in epithelial cells, A549 cells, a lung epithelial cell line, were infected with PR8 IAV and were also treated with inhibitors against p110γ (i.e. AS605240), p110δ (i.e. IC-87114) or all PI3K subunits (Wormannin). Viability of the cells was not affected by inhibitor treatment (Fig 3A), however the frequency of infected cells as well as the viral titre was significantly reduced in cells treated with Wortmannin, while AS605240 or IC-87114 showed a minor reduction (Fig 3B and 3C). Consistently, p110γ protein was undetectable in A549 cells and mRNA expression of Pik3cg and its regulatory subunit Pik3r5 was barely detectable in sorted lung epithelial cells compared to lung CD103+ DCs (Fig 3D and 3E), while significant expression of another PI3K subunit, Pik3cd, could be detected (Fig 3F). To further address the relative importance of p110γ during IAV infection in vivo in structural and hematopoietic cells, criss-cross bone marrow chimeras were generated. After reconstitution mice were then infected with IAV. Mice which had received WT bone marrow (BM) mounted a potent anti-viral T cell response, while animals receiving p110γ BM exhibited a significantly reduced number of CD4+, CD8+ and virus specific CD8+ T cells at day 7 p.i. (Fig 3B–3E). Overall, these results suggested that p110γ is required in the hematopoietic compartment for mounting an effective T cell response against IAV.
To further dissect the underlying mechanism of the impaired antiviral T cell response in the absence p110γ-kinase function, p110γ-KD mice were infected with a sub-lethal dose of IAV and the T cell response was evaluated at the peak of the response at day 10 p.i. At this time-point the number of CD8+ and Tet+CD8+ T cells in the lung was still strongly reduced in p110γ-KD animals compared to WT controls (Fig 4A). However, CD4+ T cells in p110γ-KD mice were now present in similar numbers to WT animals (Fig 4A). Similarly, no significant differences were visible in the number of CD4+, CD8+ and virus specific Tet+CD8+ T cells in the lung dLN between p110γ-KD and control mice (Fig 4B). In addition, the enhanced morbidity at this time-point post infection of p110γ-KD animals correlated with a more pronounced inflammation in the lung exemplified by a higher number of neutrophils (Fig 4C). To determine whether the reduced number of CD8+ T cells in particular was due to a reduced proliferative capacity of these cells in vivo, WT and p110γ-KD mice were infected with IAV and subsequently injected with EdU to measure the proportion of proliferating cells. At an early time-point of the antiviral T cell response the frequency of EdU+ cells among CD8+ T cells but not CD4+ T cells was reduced in p110γ-KD animals (Fig 4D–4E). To shed further light on the kinetics of the T cell response in p110γ-KD mice the same experimental set up was repeated at a later time-point at of infection. During this peak phase of the infection no deficiency in EdU incorporation could be observed for T cells localized in the lungs or dLN of infected p110γ-KD mice compared to controls (Fig 4F). Overall, these results suggested that there is a defect in the priming of CD8+ T cells lacking a functional p110γ kinase domain during IAV infection in vivo.
The T cell response against IAV is strongly dependent on T cell priming by dendritic cells (DC) [31]. To address a potential functional role of p110γ kinase activity in DC-mediated priming of T cells we cocultured bone marrow-derived DCs (BMDCs) of WT and p110γ-KD origin with OTII CD4+ T cells with different concentrations of the OVA323-339 peptide. Examining the number of CD4+ T cells after 4 days of culture as a readout for T cell proliferation it was apparent that there is no difference between using WT and p110γ-KD BMDCs (Fig 5A). Furthermore, looking at the T cell polarization in terms of Th1 cytokine production, which is the predominant T helper response during IAV infection, it was evident that p110γ-KD BMDCs have no deficiency in inducing IFNγ, GM-CSF or TNFα producing CD4+ T cells (Fig 5B). To address the possibility that p110γ is required for antigen processing by DCs and that it may play a role in the priming of CD8+ T cells, we transferred efluor-670 labeled OT-I T cells into WT and p110γ-KD recipients and injected ovalbumin in alum one day later. 7 days after transfer we then evaluated the efluor-670 staining as well as the number of TCRVα2+ cells in the inguinal lymph-node, which is the TCRα chain used by all OT-I cells. It was clearly evident that in both WT and p110γ-KD recipients OT-I cells had strongly proliferated as most cells had indeed diluted out the dye completely by day 7 (Fig 5C). Similarly, the number of CD8+CD44+CD62L-TCRVα2+ cells was much higher in the animals, which had received transferred OT-I cells, compared to the non-transferred controls (Fig 5C). Furthermore, there was no statistically significant difference between WT and p110γ-KD recipients. p110γ has been classically associated with regulating migration of immune cells towards chemokines [22] but this has thus far not been thoroughly examined in DCs. CCR7 is a critical regulator of migration of DCs towards the lung dLN and has been shown to be critical for the T cell response against IAV[32]. To test a possible role of p110γ in regulating CCR7-mediated migration of DCs we seeded WT and p110γ-KD BMDCs in a trans-well system and quantified migration towards CCL21, the principal ligand for CCR7. The frequency of migrating BMDCs was higher in p110γ-KD cells compared to WT although generally only 4–8% of cells migrated at all (Fig 5C). To account for the limitations of using BMDCs as a model for lung-resident DCs, a similar experimental set up was repeated using lung DCs. Due to the strong reduction of lung-resident CD103+ DCs in particular in p110γ-KD animals, lung CD103+ and CD11b+ DC subsets were sorted from the lungs of WT animals and then the same migration assay was done in the presence of p110γ-specific inhibitor AS650240 or DMSO as a control. A significant number of both CD103+ and CD11b+ DCs migrated, regardless of p110γ inhibition (Fig 5D). Overall these results suggested that p110γ is not generally required for DC-mediated priming of T cells in vitro and in vivo as well as for DC migration in vitro.
To elucidate the potential role of p110γ in lung DC-mediated antiviral T cell responses in vivo, we characterized the lung DC compartment of p110γ-KD mice in the naïve state and during IAV infection. As described by our group recently[23] p110γ-KD animals have a pronounced deficiency in lung-resident conventional DCs in particular the CD103+ subset (Fig 6A, 6C and 6D), while moDCs were present at comparable levels to WT mice in the steady state (Fig 6A and 6D). During IAV infection this picture changed dramatically. moDCs were recruited to the lung in high numbers regardless of p110γ-kinase deficiency (Fig 6B). Conversely, lung CD103+ DCs decreased strongly in number in both WT and p11t0γ-KD animals until day 7 p.i., where the numbers start to increase a later time-points of infection (Fig 6C). For CD11b+ DCs the situation mirrors the kinetics of the CD103+ DCs, although by day 7 p110γ-KD CD11b+ DCs manage to reach numbers similar to WT (Fig 6D). p110γ–deficient CD103+ DCs were always present in lower numbers than WT, although the difference to WT mice became somewhat smaller at the peak of IAV infection (Fig 6C). Overall, these results suggested that the conventional lung-resident CD103+ DCs in p110γ-KD mice are severely deficient at the beginning of an IAV infection and remain so for most of its course, although also in the p110γ–deficient situation the number of CD103+ DCs began to increase again at day 7 p.i. IAV To address the question whether potentially the inflammatory environment can partially overcome the pronounced developmental deficiency of lung-resident DCs in the steady-state in p110γ–deficient animals, we instilled either LPS or PolyI:C intratracheally as a single inflammatory stimulus into the lung and examined how the DC compartment changed over time. Intratracheal injection of a single inflammatory stimulus did not lead to an increase in lung CD103+ DCs in p110γ-KD mice (Fig 6F and 6G). By contrast activation of DCs by these mediators induced a reduction in the cell number, possibly due to their migration do the dLN. Thus the lack of functional p110γ cannot be compensated for by a single inflammatory signal.
To address the underlying mechanism of how the absence of p110γ kinase activity in DCs would impact on DC mediated immune responses in vivo we administered Cy-5 labeled ovalbumin intratracheally and then observed the migration of DCs to the dLN. After 24 hours cells carrying OVA-Cy5 in the dLN were predominantly DCs in both WT as well as p110γ-KD mice (Fig 7A). However, the proportion of CD103+ DCs carrying OVA in the dLN was considerably reduced in p110γ-KD compared to wild-type mice (Fig 7B). Conversely, the frequency of CD11b+ DCs was somewhat higher in p110γ-KD mice, suggesting a compensatory effect in a situation where pulmonary CD103+ DCs are virtually absent (Fig 7C). To gain a better understanding of the consequences of the strong reduction in pulmonary CD103+ DCs, we sought to evaluate CD103+ DC-specific functions. Lung CD103+ DCs were recently shown to be essential for phagocytosis of apoptotic cells and cross-presentation of antigens in the dLN [33]. To address the efferocytic capacity of pulmonary DCs in p110γ-KD mice we instilled labeled apoptotic thymocytes into the trachea and examined migration of DCs from the lung to the dLN 24h p.i. As previously described, apoptotic cells were almost exclusively transported by CD103+ DCs to the dLN, as opposed to other DC subsets (Fig 7D). Furthermore, while in WT mice a significant amount of CD103+ DCs in the dLN could be found carrying apoptotic thymocytes, these cells were almost completely absent in the dLN of p110γ-KD mice (Fig 7E). Overall, these results suggest that in p110γ-KD animals transport of apoptotic-cell related antigens by cross-presenting DCs is severely deficient.
To further investigate the notion that a deficiency in cross-presenting lung CD103+ DCs in p110γ-KD animals is responsible for their enhanced susceptibility to IAV infection, we generated BM chimeras using a mixture of either WT or p110γ-KD with Batf3-/- BM at a ratio of 1:4. Batf3 is a transcription factor strictly required for development of lung CD103+ DCs [34]. This led to a situation where at least 80% of T cells were WT, however 100% of lung CD103+ DCs were of either WT or p110γ-KD background. This set up allowed us to address the role of p110γ in lung CD103+ DCs specifically in the context of an IAV infection. Upon Infection animals that had p110γ-deficient CD103+ DCs exhibited more pronounced loss of weight and temperature (Fig 8A and 8B) compared to animals with WT CD103+ DCs. Furthermore, analyzing the T cell response at day 7 p.i. showed that CD4+, CD8+ and virus-specific CD8+ T cells were significantly reduced in the lungs of infected animals (Fig 8C). In lung dLN virus-specific CD8+ T cells were also reduced (Fig 8D). Examining the activation state of T cells in the lung it was evident that the frequency of CD44+CD62L- of CD4+ T cells was reduced in animals, which had received p110γ-KD BM (Fig 8E). However this was not the case for CD8+ T cells (Fig 8F). Overall these results suggested that p110γ in lung CD103+ DCs is required for potent antiviral T cell responses against influenza virus.
In this study we describe PI3Kγ as novel factor, which plays a key role in successful host defense against respiratory infection with influenza virus. We show that the highly increased susceptibility of p110γ-kinase dead animals stems from a defective T cell response leading to higher viral titers and more pronounced morbidity. This phenotype is due to a deficiency in the lung-resident DCs, which are essential for the initial priming of the adaptive immune response against influenza [31]. We could show that PI3Kγ is functionally not required in DCs to activate T cells in vitro and in vivo, however, the pronounced reduction of lung DCs in naïve PI3Kγ-deficient animals leads to an impaired transport of antigen to the draining lymph node and thus to a defective antiviral T cell response and consequently a delayed viral clearance.
To date the role of PI3Ks in IAV pathogenesis has mainly been analysed from the perspective of the virus, where host PI3Ks were shown to be important for viral replication, in particular of the highly pathogenic PR8 strain [30]. To our knowledge this is the first study reporting an important role of a PI3K family member for the response of the host rather than for viral pathogenesis itself. The well established pro-viral of PI3Ks in general and the novel anti-viral role of PI3Kγ described in this manuscript appear at first conflicting, however, in fact they probably just represent temporally different roles of PI3Ks in the interaction between the virus and its host. PI3Ks have been reported to be important early on during massive viral replication in lung epithelial cells [30] and our results using global inhibition of PI3K signaling support this notion. Furthermore, in this context it appears that PI3Kγ only plays a minor role, which is likely due to its very low expression in lung epithelial cells. The important antiviral role of PI3Kγ only starts to be visible once the PI3Kγ-dependent-DC-mediated T cell response commences with clearance of the virus from the lung. The fact that p110γ-KD mice do not exhibit protection from virus replication earlier on during the infection could reflect that the virus more strongly relies on other PI3K subunits for replication in vivo. The virus could also exploit redundancies between the PI3K subunits in epithelial cells, allowing it to continue replicating, despite the functional lack of PI3Kγ. How other PI3K subunits regulate both viral pathogenesis as well as the host immune response, remains to be addressed in future endeavors.
Interestingly, in p110γ-KD animals both CD4+ and CD8+ T cell responses after IAV infection are impaired at an early time-point. The strong defect in CD8+ virus-specific T cells likely stems from the pronounced deficiency in cross-presenting CD103+ DCs in naïve animals, which is similar to that observed in BATF3-/- animals [35]. The minor effect on CD4+ T cells can be explained by the moderate reduction in CD11b+ conventional DCs in the lungs of p110γ-KD animals. This subset has been associated with activating mainly CD4+ T cells [36], although recently it was shown that these cells too can stimulate CD8+ T cells in the context of influenza infection [37]. The deficiency in CD11b+ DCs however, is then overcome by the massive influx of monocyte-derived DCs as well as a novel CD11b+ DCs that likely rescue the CD4+ T cell response at later time-points during the course of the infection. The virus-specific CD8+ T cell response in the lungs of p110γ-KD animals is still strongly deficient by day 10 post infection, underlining the idea that CD103+ DCs are indeed the key players in generating a potent cytotoxic T cell response [31], despite the described capacity of CD11b+ DCs to be able to as well [37]. The deficiency in transport of apoptotic cells observable in p110γ-KD mice fits very well into this picture, as it was recently shown that this capacity was indeed crucial for the ability of CD103+ DCs to cross-present antigen derived from dying cells to CD8+ T cells [37]. Nonetheless, when one examines the frequency of CD8+ T cells which have proliferated in the lungs of infected animals the moderate reduction in EdU+ cells in p110γ-KD animals likely does not completely explain the massive defect in the virus-specific CD8+ T cell compartment. Most probably, lung-resident DCs are not only required for initial priming of the CD8+ T cell response in dLN but also for maintenance of CD8+ effector T cells once they are have reached the pulmonary compartment. This notion has been previously suggested for monocyte-derived DCs [38] and is likely also true for conventional lung-resident DCs. Furthermore, CD4+ T cells proliferate more in p110γ-KD than WT mice at late time-points of infection, which can be likely explained by the continuous presence of virus in the p110γ deficient situation. WT animals have cleared the virus by this stage of the infection and thus antigen is no longer present to stimulate T cell proliferation.
From a developmental point of view the results presented in this paper also bear some novel insights into how DC development changes from steady-state to inflammation. The pronounced developmental organ-specific deficiency of lung-resident CD103+ DCs in the lungs of p110γ-KD animals, which we recently described [23], is partially overcome by the prolonged inflammation during a respiratory viral infection. The numbers of CD103+ DCs decrease strongly until day 4 p.i. but then start to recover at later time-points, also in the p110γ- deficient situation. These results suggest the potential presence of an additional developmental pathway overcoming the requirement for Flt3 signaling in lung CD103+ DC development during inflammation. Possibly this mechanism is also regulated by IL-12 and IFNγ, which has been shown to overcome deficiency of BATF3 in chronic infection with Mycobacterium tuberculosis allowing development of lung CD103+ DCs [39]. Unsurprisingly, we could show that for the rerouting of DC development during pulmonary infection, one inflammatory stimulus alone such as LPS or poly I:C in the lung is not sufficient. These observations are also reflected in GM-CSF deficient animals where CD103+ DCs are moderately reduced in adults but this deficiency is then also completely rescued during inflammation [35].
Despite the pre-existing deficiency in lung DCs in p110γ-KD animals, we could also demonstrate that migration of DCs lacking a functional kinase domain of in p110γ is completely normal in vitro and in vivo, thus showing that p110γ is dispensable for chemokine receptor mediated migration to the dLN. This corrects a notion, which has been suggested for DCs by older publications [40].
Overall our results define PI3Kγ as a new key host factor in the defense against influenza virus infection, establishing a more complex picture about the role of PI3Ks in respiratory viral infection and the interplay between the host and its pathogen.
C57BL/6J mice were either bred in-house or purchased from Charles River (Germany). p110γ-KD mice were generated and provided by E. Hirsch, University of Torino and were back-crossed to C57BL/6 for at least 15 generations. BATF3-/- mice were obtained from Jackson (USA). All animals were housed in individually ventilated cages under specific pathogen free conditions at the ETH Phenomics Facility (Zurich, Switzerland) and used for experiments at between 6 and 14 weeks of age unless otherwise stated. The number of mice (n) indicated in the figure legends always refers to the number of animals per group.
Mice were euthanized by an overdose of sodium pentobarbital by intraperitoneal injection. Lungs, spleens or lung draining lymph nodes were removed and then processed as described previously [23] Bronchoalveolar lavage was obtained by inserting a catheter into the trachea and subsequently flushing the lungs with PBS for a total volume of 1ml. BAL cells were then obtained by centrifugation and the supernatant was analysed for inflammatory cytokines using ELISA.
Flow cytometry analysis was performed on a FACSCanto II or LSR Fortessa (BD) and analyzed with FlowJo software (Tree Star). Fluorochrome-conjugated or biotinylated monoclonal antibodies specific to mouse CD11c (N418), CD11b (M1/70), Ly-6C (HK1.4), Siglec-F (E50-2440, BD Biosciences), CD103 (2E7), CD45 (30-F11), CD45.1 (A20), CD45.2 (104), CD4 (GK1.5), CD8α (53–6.7), MHC class II (M5/114.15.2, eBioscience), CD64 (X54-5/7.1), CD19 (6D5), CD3e (145-2C11), NK1.1 (PK136, eBioscience), Ly-6G (1A8), podoplanin (8.1.1., eBioscience), Ly-6C (HK1.4), CD49b (DX5), TNF-α (eBioscience) IFN-γ (eBioscience) and GM-CSF (BD) were purchased from Biolegend unless otherwise stated. Dead cells were gated out using the live/dead marker eFluor780 (eBioscience) before analysis. PE-conjugated peptide-MHC class I tetramers (H- 2Db/NP34) with the NP34 peptide (NP366-374; ASNENMETM) from the nucleoprotein of influenza virus A/PR/8/34 were generated as described [41]. Prior to all flow cytometry stainings, FcγIII/II receptors were blocked by incubating cells with homemade anti-CD16/32 (2.4G2).
For bone marrow chimeras, C57BL/6 CD45.2+ mice were lethally irradiated (9.5 Gy, using a caesium source) and reconstituted with 5-10x106 BM cells of the background and with the ratio indicated for each experiment. Mice were used for virus experiments weeks post-reconstitution.
Influenza virus strain PR8 (A/Puerto Rico/34, H1N1) was originally provided by J. Pavlovic, University Zurich. At the age of 6 to 14 weeks, mice were infected intratracheally with varying doses of influenza virus, depending on the experiment. The mice were anaesthetized and inoculated with 50μL virus in endotoxin-free PBS. Temperature and weight of animals was monitored daily and animals were euthanized if they fulfilled severity criteria set out by institutional and cantonal guidelines.
To determine influenza viral titers in the lungs, samples were collected on various days after infection, homogenized (Polytron PT 1300 D), and serially diluted with MDCK cells as described [42].
The day before mice were sacrificed for analysis, 1.5x105 bone marrow-derived dendritic cells (BMDC) [43] were incubated overnight with 1.6x105 pfu UV-inactivated virus (PR8) in 96-well plates. 12h later, these BMDC were pulsed with 1μg/mL NP34 peptide for 2h before BAL cells from individual mice were added. After 2h of incubation at 37°C, Monensin (2μM, Sigma-Aldrich) was added to retain cytokines in the cytoplasm, and cells were again incubated at 37°C for another 3h. Cells were then harvested and stained for flow cytometry analysis.
At the indicated time points, BAL fluid was measured for virus-specific IgA and IgG antibody isotype levels. Ninety-six well plates (Maxisorp; Nunc) were coated with UV-inactivated influenza virus (PR8) in PBS overnight at 4°C. Plates were washed and incubated with PBS-1% BSA for 2h at RT for blocking. BAL fluids from individual mice were serially diluted in PBS-0.1% BSA starting with a 1:2 dilution for BAL fluids and a 1:50 dilution for sera, followed by incubation at RT for 2h. Plates were washed five times and incubated with alkaline-phosphate-labeled goat anti-mouse antibodies to IgG1, IgG2c or IgA (Southern Biotech Technologies, Inc.) at a 1:1000 dilution in PBS-0.1% BSA at RT for 2hrs. Thereafter, plates were washed five times and substrate p-nitrophenyl phosphate (Sigma-Aldrich) was added. Optical densities were measured on an enzyme-linked immunosorbent assay reader (Bucher Biotec) at 405nm.
For these experiments the Cy5 labeling kit (Axxora) was used with ovalbumin (Invitrogen) to produce OVA-Cy5 according to the manufacturer’s instructions. Mice were anesthetized using isofluorane and injected intratracheally with 40μg OVA-Cy5 and 100ng LPS (Sigma) and tissues were analyzed using flow cytometry one-day post injection.
Bone marrow cells were harvested from the femurs of donor animals using a syringe and PBS. Cells were subsequently incubated in complete RPMI-1640 medium (Life Technologies) with 20ng/ml GM-CSF for 7–9 days depending on the experiment. On the day of use non-adherent cells were harvested by gently pipetting and subsequently used for further assays.
DCs were seeded into Costar 5um polycarbonate membrane trans-well plates (Corning) and medium containing 100nM CCL21 was added to the bottom of the well. The plates were incubated for 6h at 37°C and subsequently the migrated cells were harvested and analysed using flow cytometry.
BMDCs were obtained as described above. T cells were obtained with CD4 MACS-bead (Miltenyi) sorting from naïve splenocytes. Cells were then cultured together for 4 days in complete IMDM medium (Life Technologies) with the addition of varying concentrations of OVA323-339 peptide (Mimotopes Australia). Before flow cytometric analysis, cells were restimulated with PMA (Sigma-Aldrich) ionomycin (Sigma-Aldrich) and monensin (Sigma-Aldrich).
Mice were anesthetized using isofluorane and injected intra-tracheally with either 100ng LPS (Sigma-Aldrich) in PBS or 50μg Poly (I:C) (Invivogen) in PBS respectively. Control mice were just injected with PBS. At the indicated time-points mice were sacrificed and tissues were analyzed using flow cytometry.
Thymocytes were obtained from C57BL/6 mice and subsequently smashed through a 70um cell strainer to get a single cell suspension. Subsequently apoptosis of these cells was induced by a 120mJ exposure to UV-light. Cells were incubated at 37°C for 2h and were then labeled with efluor 670 (eBioscience) according to the manufacturer’s instructions. After the labeling procedure 1x 107 cells were injected intra-tracheally into recipient mice and tissues were analyzed one day later.
A549 cells were treated with inhibitors for 30 min in infection medium (DMEM with 50 mM Hepes and 0.2% BSA, pH 6.8) at a final concentration of 5μM. They were then infected with MOI 0.5 PR8-NS1-GFP virus [44]. 0.2μg/ml TPCK-Trypsin was added and infection was allowed to proceed for 10h. Cells were then analyzed for GFP production using flow cytometry and supernatants were collected for analysis of the virus titre, as described above. The following inhibitors were used: Wortmannin (Sigma), AS605240 (Sigma) and IC87114 (Sigma)
Cells were lysed in laemmli buffer containing 10% β-mercaptoethanol and boiled for 15mins at 95°C. The samples were then separated by SDS-PAGE and transferred to a nitrocellulose membrane. Thereafter the membranes were blocked in 10% milk powder in TBST. Membranes were then incubated in the primary and secondary antibodies at RT for 1h each respectively, with washing in TBST in between. After washing the Supersignal West Pico Chemiluminescent substrate was added and the membranes were imaged on Chemidoc MP imaging system (Biorad). The following antibodies were used: goat anti-mouse p110γ (Santa Cruz Biotechnology), bovine anti-goat HRP (Santa Cruz Biotechnology), goat anti-mouse actin (Santa Cruz Biotechnology).
For analysis of PI3K subunit expression RNA was isolated from cells with TRIzol reagent (Invitrogen) and was reverse-transcribed with GoScript reverse transcriptase according to the manufacturer’s instructions (Promega). Quantitative real-time RT-PCR was performed with KAPA SYBR FAST. The expression of PI3K subunits was normalized to that of Tbp.
Mean values, SD, SEM, and Student’s t test (unpaired) and One-way ANOVA with CI 95% were calculated using Prism (GraphPad Software, Inc). p < 0.05 (*), p < 0.01 (**), p < 0.001 (***), p < 0.0001 (****).
All animal experiments were approved by the local animal ethics committee (Kantonales Veterinärsamt Zürich, licenses ZH270/2014 and 113/2012), and performed according to local guidelines (TschV, Zurich) and the Swiss animal protection law (TschG).
|
10.1371/journal.pntd.0003699 | Effectiveness and Safety of Short Course Liposomal Amphotericin B (AmBisome) as First Line Treatment for Visceral Leishmaniasis in Bangladesh | Bangladesh is one of the endemic countries for Visceral Leishmaniasis (VL). Médecins Sans Frontières (MSF) ran a VL treatment clinic in the most endemic district (Fulbaria) between 2010 and 2013 using a semi-ambulatory regimen for primary VL of 15mg/kg Liposomal Amphotericin-B (AmBisome) in three equal doses of 5mg/kg. The main objective of this study was to analyze the effectiveness and safety of this regimen after a 12 month follow-up period by retrospective analysis of routinely collected program data. A secondary objective was to explore risk factors for relapse.
Our analysis included 1521 patients who were initially cured, of whom 1278 (84%) and 1179 (77.5%) were followed-up at 6 and 12 months, respectively. Cure rates at 6 and 12 months were 98.7% (1262/1278) and 96.4% (1137/1179), respectively. Most relapses (26/39) occurred between 6 and 12 months after treatment. Serious adverse events (SAE) were recorded for 7 patients (0.5%). Odds of relapse at 12 months were highest in the youngest and oldest age groups. There was some evidence that spleen size measured on discharge (one month after initiation of treatment) was associated with risk of relapse: OR=1.25 (95% CI 1.01 to 1.55) per cm below lower costal margin (P=0.04).
Our study demonstrates that 15mg/kg AmBisome in three doses of 5mg/kg is an effective (>95% cure rate) and safe (<1% SAE) treatment for primary VL in Bangladesh. The majority of relapses occurred between 6 and 12 months, justifying the use of a longer follow-up period when feasible. Assessment of risk of relapse based on easily measured clinical parameters such as spleen size could be incorporated in VL treatment protocols in resource-poor settings where test-of-cure is not always feasible.
| Visceral Leishmaniasis (VL) is a parasitic disease which is endemic in more than 80 countries, although 90% of cases occur in India, Bangladesh, Sudan, South Sudan, Ethiopia and Brazil. Most treatments are complex, expensive and require long application periods. AmBisome is one of the newest treatments available, but evidence for its safety and effectiveness under routine program conditions in resource-poor endemic areas remains sparse. Médecins Sans Frontières (MSF) ran a VL clinic from 2010 until 2014 in Fulbaria District, Bangladesh. Our retrospective study was based on all available data from this clinic, comprising 1521 patients diagnosed with primary VL who were treated with AmBisome 15mg/kg in three equal doses of 5mg/kg. We found that this treatment was safe (less than 1% of patients experienced a severe adverse event) and effective (more than 95% of patients were cured with one treatment) after 12 months. The youngest and oldest patients, and patients with large spleen size at the end of treatment, were more likely to experience a relapse. More than half of the relapses occurred between 6 and 12 months after treatment, therefore we recommend that clinical trials and treatment protocols adopt a minimum 12-month follow-up period.
| Visceral Leishmaniasis (VL), also known as Kala Azar, is a vector borne disease caused by parasites of the genus Leishmania, L. donovani—L. infantum complex, which are transmitted through the bite of an infected sand fly (mainly genus Phlebotomus, old world, and Lutzomya, new world), Leishmaniasis infection in humans presents as cutaneous, muco-cutaneous and visceral. The visceral form is fatal if untreated. VL progressively affects the immune system of the patient, and opportunistic infections are frequently the final cause of death [1,2]. VL is endemic in around 80 countries, with over 90% of cases found in India, Bangladesh, Sudan, South Sudan, Ethiopia and Brazil [3].
The Fulbaria district of Bangladesh reported an average annual VL incidence rate of 17.8 per 10,000 people between 2008 and 2013 [4]. This far exceeds the government’s goal of one reported case per 10,000 habitants by the year 2015, as set in the Bangladesh VL Elimination Program in 2005 [4]. Until 2010, VL was treated in Bangladesh with Sodium Stibogluconate (SSG) or Miltefosine. SSG is given by intramuscular (IM) injections daily for 28 days. This course of treatment is painful, toxic, cumbersome for patients, and costly for the health system. In 2008, the national protocol for VL treatment in Bangladesh shifted to first line treatment with oral Miltefosine for 28 days. Miltefosine has gastro-intestinal side effects, is teratogenic (requiring 5 months of effective contraception for women of child bearing age, continuing for 3 months post-treatment [5]), and is prone to poor adherence if not administered under direct observation [6]. In 2013, AmBisome single dose (10mg/kg) was adopted as first line treatment, and this regimen is currently being rolled out across the country.
Médecins Sans Frontières (MSF) has run a VL control project in Bangladesh since 2010 by agreement with the Ministry of Health (MoH). In May 2010, MSF introduced an innovative first line therapy for primary VL comprising 15mg/kg liposomal amphotericin B (AmBisome) intravenous infusion, divided into three doses of 5 mg/kg given over a 5 day period, with only one night of hospitalization. Such a regimen had shown good efficacy and safety in a small clinical trial in India [7]. MSF introduced this regimen because the Bangladesh national VL guideline [8] allowed its use as second line treatment, although its effectiveness and safety had not been evaluated in routine clinical practice in south Asia. It was considered at that time that introduction of other regimens, such as a single dose [9][10], would require further evidence from clinical trials.
The aim of this study was to explore the effectiveness and safety of a short course AmBisome regimen of 15 mg/kg for primary VL under routine program conditions in Bangladesh, and to investigate predictive factors for relapse, with a 12 month follow-up period to assess final outcomes. Clinical studies in VL typically use a follow-up period of 6 months to establish final cure [11]. However, there is some evidence that most relapses occur later than 6 months post-treatment [12] [13], and wetherefore adopted a 12 month follow-up to allow the assessment of relapse rates up to 12 months post-treatment.
This retrospective analysis used routine patient data from the MSF VL clinic in Fulbaria, Bangladesh, collected from January 2010 to April 2014. We included all patients diagnosed with primary VL who received AmBisome 15 mg/kg in three separate doses (5mg/kg), and who were not referred to other institutions for treatment or follow-up. Relapse VL cases require treatment with a higher total dose of AmBisome, and were excluded from our analysis.
In accordance with the MoH protocol [8], primary VL was suspected in patients with fever >2 weeks duration, weight loss, splenomegaly or lymphadenopathy, and no history of previous VL. VL was confirmed by rK39 rapid diagnostic test (RDT) (IT-LEISH, Bio-Rad Laboratories, USA). The performance of rK39 RDTs in South Asia has been shown to be consistently very high [14–16]. In a setting with a 50% prevalence of VL among clinical suspects this results in a very high positive predictive value. The patient’s general clinical condition, height, weight, spleen size and hemoglobin level (checked with HemoCue, HemoCue AB, Sweden) were recorded on admission to the clinic and when the patient was assessed for discharge (one month after admission). Pregnancy testing was done on admission. As the prevalence of malaria in this part of Bangladesh is extremely low, screening for malaria was restricted to vulnerable groups: age <2 or > 60 years and pregnant women, and to suspected malaria cases: patients with acute high fever (>39°C) or low hemoglobin (<8 g/dL). Routine HIV testing was not conducted due to the very low prevalence of HIV infection in this population.
Patients with confirmed primary VL were treated with intravenous liposomal amphotericin B (AmBisome, Gilead Pharmaceuticals, Foster City, CA, USA) using a total dose of 15mg/kg divided into three individual doses of 5 mg/kg: at admission (day 0); 24 hours after this dose, and five days after the first dose. Patients were routinely hospitalized for one day, or longer if severely ill. Patients showing a good clinical response (feeling well, resolution of fever, regression of spleen, and improvement of Hb [11]) were discharged from treatment after an assessment at 1 month. If there had not been a good clinical response, the possibility of treatment failure was checked by test-of-cure. Test-of-cure was by microscopy of Giemsa-stained smear of splenic aspirate. Patients who had missed any of the three doses were registered as “defaulted”. No parasitological test of cure was done routinely for primary VL cases. Patients presenting with VL relapse were treated with AmBisome (3 doses of 15mg/kg if their primary VL had not been treated with AmBisome, 5 doses of 25mg/kg if previously treated with AmBisome) followed by a test-of-cure after 28 days. Clinical presentation of relapse VL tends to be less severe but, in non-HIV-infected relapse patients, signs and symptoms remain typical (prolonged undulating fever, splenomegaly, anemia, weight loss). Diagnosis of VL relapse was confirmed parasitologically (by microscopy of Giemsa-stained smear of splenic aspirate).
All patients were invited to attend a clinical follow-up appointment at 6 and 12 months after treatment or if ill health occurred. At each follow-up appointment a general clinical examination was done, including recording of temperature, spleen size and hemoglobin level. Outcome at each visit was recorded as: clinically cured (no signs or symptoms of systemic disease); sick but not VL (patients that were sick at the moment of the follow-up, but did not present signs and symptoms or clinical history suggestive of VL relapse); relapse (all patients with signs and symptoms of VL, confirmed by spleen aspirate microscopy if there were no contra-indications for splenic aspiration: spleen size >3cm below left costal margin, signs of active bleeding, increased bleeding and clotting time, severe anemia, jaundice, pregnancy, patient unable to remain still); or death. If not attending for follow-up, community health workers actively searched for the patient; if unsuccessful, the patient was classified as "lost to follow-up". Deaths during the follow-up period were determined by health staff using verbal autopsy to gather information on the cause of death and whether it could be related to VL. Molecular methods were not available for typing Leishmania strains in order to differentiate relapse from re-infection. PKDL was systematically looked for during follow-up, but was not considered a VL treatment failure.
For the measurement of effectiveness we classified treatment outcome at one month as ‘initial cure’ for patients whose fever had subsided and who presented with spleen regression, increased hemoglobin level and weight, and improved general condition. Treatment outcomes at 6 and 12 months were categorized as: ‘cured’ (patients without sign and symptoms of VL); and ‘treatment failure’ (patients who had symptomatic VL relapse or who had died). For the measurement of safety we analyzed the frequencies of reported clinical complications that could be caused by VL treatment (vomiting, bleeding, and other). Severe adverse events (SAE) were defined as events leading to suspension or the cessation of VL treatment. Potential determinants of relapse for which there was evidence of an association in univariate analysis were carried forward to a multivariable logistic regression model that was adjusted (as a priori confounders) for age and sex. Predictive factors measured at discharge (one month after initiation of treatment) were adjusted for their values at baseline (admission) and all variables were added one by one to the final model. Continuous variables measured in adults versus children, and on admission versus discharge, were compared using Student’s t test. Data were analyzed with SPSS (SPSS Inc. Released 2008. SPSS Statistics for Windows, Version 17.0. Chicago: SPSS Inc.) and Stata (StataCorp. 2013. Stata Statistical Software: Release 13. College Station, TX: StataCorp LP)
This analysis met the Médecins Sans Frontières International Ethics Review Committee criteria for a study involving the analysis of routinely collected program data. The program utilized a recognized treatment for VL in Bangladesh, and was run in coordination with the Bangladesh Ministry of Health and Family Welfare through a memorandum of understanding, which is the usual procedure for NGOs operating in this context. All electronic data were analyzed anonymously.
After applying the inclusion criteria, a total of 1521 patients who completed the treatment were included in our study (Fig 1). The first and last admissions occurred on May 17th 2010 and January 12th 2013, respectively. The last patient to complete their 12 month follow-up attended the clinic on February 10th 2014. Of the 1653 patients diagnosed with primary VL during this period, 40 complicated cases were transferred to tertiary care at Mymensingh Medical College. Of the 1613 patients treated at Fulbaria during the study period, 89 received other regimens: 18 patients with 5 x 3mg/kg had (suspected) impaired renal function and were treated with a lower dose more frequently over a longer period in accordance with protocol; 61 patients were treated with a single 10mg/kg dose as a new national protocol began to be introduced in 2013; and 10 patients were treated outside protocol for reasons which were not recorded in the database. The proportions of patients whose status was known at 6 and 12 months follow-up were 84.0% (1278/1521) and 77.5% (1179/1521), respectively. Data with which to analyze factors associated with risk of relapse at 12 months were available for 1106 patients.
The population was young (58.2% were <18 years old) and had similar age distributions in both sexes. The male:female ratio was approximately 1:1 across all age groups except among patients age 40+ years (3:1). The study population appeared to be under-nourished relative to WHO standards, with a mean BMI of 17.5kg/m2 for patients age ≥18 years, and a mean weight-for-height Z-score of −1.8 for patients <121 cm tall (see Table 1). No differences in baseline characteristics were observed between patients from different geographical areas. Adult and pediatric patients presented with similar Hb levels and spleen sizes, and were analyzed as a single patient group.
During the treatment period, complications were registered as follows: bleeding 3.4% (52/1521); vomiting 9.5% (144/1521); and other complications (including fever, diarrheas, abdominal pain and local rash at the site of the injection) 0.4% (6/1521). Serious adverse events (SAE) were recorded for 7 patients (0.5%). The remainder had no SAE data, because there was no zero-reporting of adverse events. We cannot analyze the specific details of the SAE because no extra information was recorded in patient files. One of the SAE patients died before attending their first follow-up; of the remainder, none relapsed after 12 months of follow-up.
Improvements in key clinical parameters were noted at the one month follow-up: mean Hb levels increased from 8.99 to 10.87 g/dL (mean difference 1.90 g/dL (95% CI 1.80 to 1.95), P<0.001); mean spleen size decreased from 4.6 to 0.5 cm (mean difference −4.0 cm (95% −3. 9 to −4.2), P<0.001); and mean weight increased from 29.7 to 31.3 kg (mean difference 1.5 kg (95% CI 1.4 to 1.6), P<0.001).
Cure rates at 6 and 12 months were 98.7% and 96.4%, respectively (Table 2). A ‘best case scenario’ sensitivity analysis, in which we coded all patients lost to follow-up as ‘cured’ (16% (243/1,521) at 6 months, 22.5% (342/1,521) at 12 months), yielded a 6 month cure rate of 98.9% and a 12 month cure rate of 97.2%.
At 6 months follow-up, a total of 13 patients were diagnosed as relapse and 3 deaths were registered. After completing the 12 month follow-up, a total of 39 patients were diagnosed as relapse (including the previous 13 cases): 29 with parasitological confirmation; 10 without (because splenic aspiration was contra-indicated). Of these 10, none presented with lymphadenopathy. Local staff were not trained to perform bone marrow aspirates hence, lymph or bone marrow aspirates were not taken. No further deaths were registered.
Patients followed-up at 12 months did not differ in age, sex and clinical measures on admission and at discharge compared to patients not lost to follow-up, but patients lost to follow-up were more likely to reside in further away sub-districts, which we classified as “distant sub-districts” (27.5% versus 5.9%).
The three deaths occurred among male patients age >45 years: one entered treatment with a diagnosis of complicated advanced tuberculosis that was probably the main cause of death (after completing VL treatment he remained as an inpatient and died before the first follow-up); one died between the second and third follow-up from social violence (according to information provided by family members); and one suffered a severe adverse event (registered as acute jaundice) after he completed treatment and died at home before the first follow-up. This was the only case where the death could probably be attributed to VL and/or its treatment.
In Table 3 we summarize the results of a logistic regression model with relapse at 12 months as the outcome. We included only those patients with known status at 12 months (cured or relapsed) and who had complete data for potential risk factors (N = 1106). Odds of relapse were highest in the youngest and oldest age groups. There was no association with sex. Anthropometric indices for nutritional status (BMI in adults, weight-for-height Z-score in children under 121cm) were not associated with risk of relapse: for adults (n = 463), OR = 0.87 (95% CI 0.63 to 1.20) per kg/m2, P = 0.41; for children (n = 475), OR = 0.97 (95% CI 0.65 to 1.44) per weight-for-height Z-score, P = 0.88. Larger spleen size at discharge was associated with increased risk of relapse: OR = 1.25 (95% CI 1.01 to 1.55) per cm below lower costal margin, P = 0.04). We found tentative evidence of interaction (Likelihood ratio test P = 0.08) between discharge measurements of Hb and spleen size, which suggested an amplified combined effect of low Hb and large spleen size (S1 Fig).
Our study has shown that 15mg/kg AmBisome in three doses of 5mg/kg is an effective (96.4% cure rate) and safe (<1% SAE) treatment for primary VL in Bangladesh, when judged against internationally accepted parameters for effectiveness ≥95% and safety (SAE<5%) for VL treatment [1]. We demonstrated that VL treatment studies in this setting require a follow-up period longer than 6 months if they are to capture the majority of relapses. Whether routine follow-up of discharged patients for this length of time is necessary depends on ease of access to re-treatment for patients who relapse. In settings where access to re-treatment is problematic, routine follow-up may be equally difficult (and costly), but it could help to achieve the VL elimination target (of <1 case per 10,000 people at upazila level in Bangladesh) set in the Regional Strategic Framework for Elimination of Kala-azar from South East Asia Region [17]. We have also shown that residual splenomegaly is predictive of an increased risk of VL relapse, particularly in conjunction with low levels of hemoglobin.
The main strengths of our study are its size (>1500 patients), inclusivity (92% of primary VL cases during the period of the study), and extended follow-up (12-month). The main limitation of any cohort study is loss to follow-up. These were low (23% at 12 months) for a study of this design in a resource-poor setting, in part because outreach workers were employed to seek patients who did not attend follow-up appointments. A ‘best case scenario’ analysis showed a maximum cure rate of 97.2%. Although we cannot be certain about the validity of this scenario, ease of access to VL treatment in a relatively stable (non-migratory) population with multiple means of communication and transport (and given that apart from distance from facility, we found no differences between patients who were/were not lost to follow-up) suggest that the true cure rate is unlikely to be lower than 95%. Another limitation was the completeness of SAE reporting because non-occurrence of events was not recorded. Therefore we cannot verify that blank SAE values truly correspond to “treatment complete without SAE”. Although not all milder adverse events will have been reported and we are confident that no severe adverse events were missed, this is a limitation which we would seek to address in future studies by implementing a zero-reporting protocol for SAE.
AmBisome has been used to treat VL for more than 10 years [18], on the basis of evidence from clinical trials. However, evidence for the safety and effectiveness of AmBisome-based regimens under routine program conditions in endemic resource-poor areas remains sparse [19]. This has mainly been due to the high cost of the drug and the neglected status of the disease, factors which MSF has worked to address for many years [20]. In the meantime, studies based on routine data collection from VL treatment programs are an important source of evidence.
Our 6-month cure rate (98.7%) is consistent with rates reported from other studies of AmBisome [21], including several conducted in Bihar, India (a neighboring endemic region): Sinha et al reported 98.8% with 20mg/Kg in 4 doses over 10 days [22]; Sundar et al reported 95.7% with 10mg/Kg in single doses [9], 98.4% with a single 15mg/Kg dose [23], and 97.5% with combined doses of Liposomal Amphotericin B and Miltefosine [24]. In Bangladesh, in the neighboring sub-district of Muktagacha, Mondal et al reported a cure rate of 98% with a single dose 10mg/kg [25].
Our study provides further evidence that the standard 6 month follow-up period for VL studies or treatment programs [1] is insufficient. Several other studies have shown that most relapses occur after 6 months [13] [12] [26], and we would recommend that clinical trials and treatment programs adopt a 12-month follow-up period if relapse rates are to be measured accurately. However, even this extended length of follow-up is arbitrary, and we cannot discount the possibility that some patients relapsed after 12 months [13].
VL is associated with progressive weight loss and poor nutritional status. Conversely, malnutrition results in impaired immunity and is an important risk factor for severity of clinical VL [27] [28]. The very low anthropometric indices which we found in our patients are consistent with other VL patient cohorts in this population [25]. Whether these low indices are entirely a consequence of VL or whether they also indicate that chronic under-nutrition is a factor in perpetuating the endemicity of VL warrants further research.
That spleen size on discharge is a risk factor for VL relapse is consistent with findings from India [13] and South Sudan [29]. We also found tentative evidence of an amplified combined effect of low Hb and large spleen size, which is entirely plausible given that low Hb can be indicative of severity of VL disease (effect on bone marrow), and a massive spleen traps red cells and further depresses Hb [30–32]. Assessment of risk of relapse based on these easily-measured clinical parameters could be incorporated in VL treatment protocols in resource-poor settings where test-of-cure procedures cannot be routinely implemented, with closer and/or extended monitoring of at-risk patients.
Our study has shown that 15mg/kg AmBisome in three doses of 5mg/kg is a safe and effective treatment for primary VL in Bangladesh, and could be an alternative to the current first line regimen of single dose 10mg/kg AmBisome. We have also shown that a follow-up period of 12 months is required to capture the majority of VL relapse cases, and that VL relapse is predicted by low Hb and large spleen size at the end of treatment. Until more evidence is gathered, we would recommend that follow-up be extended to 12 month for all patients or, where this is not feasible, 12-month follow-up could be targeted at patients who present with large spleen at the end of the treatment.
|
10.1371/journal.pcbi.1003126 | Reconstruction and Validation of a Genome-Scale Metabolic Model for the Filamentous Fungus Neurospora crassa Using FARM | The filamentous fungus Neurospora crassa played a central role in the development of twentieth-century genetics, biochemistry and molecular biology, and continues to serve as a model organism for eukaryotic biology. Here, we have reconstructed a genome-scale model of its metabolism. This model consists of 836 metabolic genes, 257 pathways, 6 cellular compartments, and is supported by extensive manual curation of 491 literature citations. To aid our reconstruction, we developed three optimization-based algorithms, which together comprise Fast Automated Reconstruction of Metabolism (FARM). These algorithms are: LInear MEtabolite Dilution Flux Balance Analysis (limed-FBA), which predicts flux while linearly accounting for metabolite dilution; One-step functional Pruning (OnePrune), which removes blocked reactions with a single compact linear program; and Consistent Reproduction Of growth/no-growth Phenotype (CROP), which reconciles differences between in silico and experimental gene essentiality faster than previous approaches. Against an independent test set of more than 300 essential/non-essential genes that were not used to train the model, the model displays 93% sensitivity and specificity. We also used the model to simulate the biochemical genetics experiments originally performed on Neurospora by comprehensively predicting nutrient rescue of essential genes and synthetic lethal interactions, and we provide detailed pathway-based mechanistic explanations of our predictions. Our model provides a reliable computational framework for the integration and interpretation of ongoing experimental efforts in Neurospora, and we anticipate that our methods will substantially reduce the manual effort required to develop high-quality genome-scale metabolic models for other organisms.
| Few organisms have been as foundational to the development of modern genetics and cellular metabolism as Neurospora crassa. Given the wealth of knowledge available for this filamentous fungus, the effort required to manually curate a high-quality genome-scale metabolic reconstruction would be daunting. To aid the reconstruction process, we developed three optimization-based algorithms. The first algorithm predicts flux while linearly accounting for metabolite dilution; the second algorithm removes blocked reactions with one compact linear program; and the third algorithm reconciles differences between in silico predictions and experimental observations of mutant viability. We have used these algorithms to develop the first genome-scale metabolic model for Neurospora. We have validated the accuracy of our model against an independent test set of more than 300 growth/no-growth phenotypes, and our model displays 93% sensitivity and specificity. Simulating the biochemical genetics experiments originally performed on Neurospora, we comprehensively predicted essential genes, nutrient rescues of auxotroph mutants and synthetic lethal interactions. With these predictions, we provide potential mechanistic insight into known mutant phenotypes, and testable hypotheses for novel mutant phenotypes. The model, the algorithms and the testable hypotheses provide a computational foundation for the study of Neurospora crassa metabolism.
| First discovered as an orange mold infestation of Paris bakeries in 1843 [1], the filamentous fungus Neurospora crassa has become a model organism for eukaryotic biology and the cornerstone of a vibrant research community [2]. Work on Neurospora has led to essential discoveries in circadian rhythms [3], epigenetics [4], genome defense [5], mitochondrial biology [6], post-transcriptional gene silencing [7] and DNA repair [8]. Most famously, work in the 1940's by Beadle and Tatum led to the Nobel Prize-winning ‘one-gene-one-enzyme’ hypothesis that established the fundamental link between genes and proteins in all organisms [9], [10]. Work on Neurospora thus paved the way for modern genetics and molecular biology.
Of equal consequence, the work by Beadle and Tatum ushered in a new era in the study of biochemistry and cellular metabolism. The genetic facility of Neurospora, coupled with its ability to grow on minimal media, simplified the isolation of mutants with additional nutrient requirements. The first such auxotrophic mutants established the universal link among genes, enzymes, and the ordering of reactions in biosynthetic pathways. Work over subsequent decades led to a compilation of hundreds of such mutants, shedding light on most major biosynthetic pathways [11]–[13]. With the sequencing and annotation of the Neurospora genome [14], [15], these genetic data could be organized on a physical scaffold, genetic markers could be assigned to specific genes with predicted biochemical functions, genes could be assigned to previously orphaned biochemical reactions, and a global map of Neurospora metabolism could begin to emerge.
Genome-scale metabolic models have been constructed for over 100 organisms spanning bacteria to mammalian cells [16]. These network models capture information about all known metabolic reactions and the genes that encode enzymes for these reactions in a computationally structured manner. More than simply a catalog of reactions, network models capture biochemical relationships between reactions and pathways, afford a framework for integrating genomic measurements, and provide constraints for computational inference. One widely used method for computational inference using metabolic network models is Flux Balance Analysis (FBA) [17]. FBA calculates the flux of metabolites through a network under the assumption that metabolism is at steady state on the time-scales of interest. Using constraint-based modeling methods like FBA [18], it is possible to predict the growth rate of organisms under different conditions [19], the rate of production of metabolites of interest [20], the phenotypic consequences of gene knockouts, and the metabolic impact of different gene expression programs [21], [22]. Constraint-based methods are also being used to guide metabolic engineering efforts by calculating the modifications required to optimize the production of desired metabolites [23]–[26].
The wealth of genetic and metabolic data available for Neurospora, along with ongoing efforts to knock-out and phenotypically characterize all ∼10,000 genes in the genome [27], provides a strong foundation for the development of a genome-scale metabolic model. A metabolic model would, in turn, complement experimental efforts by integrating data from experiments on single genes into a coherent genome-wide metabolic framework, providing potential mechanistic insight into experimental phenotypic observations, and enabling the comprehensive modeling of perturbations that could not be feasibly performed in the lab. A genome-scale model is also a requirement for the rational and efficient use of Neurospora as a potential biofuels organism [28]–[32].
We report here the construction and validation of a high-quality genome-scale metabolic model for Neurospora crassa. To guide the process of model construction, we developed a novel suite of algorithms called Fast Automated Reconstruction of Metabolism (FARM). We validated the model against an independent gene essentiality test set, and achieved 93% sensitivity and specificity. We applied the validated model to comprehensively predict nutrient rescue of essential genes and synthetic lethal interactions. With these predictions, we provide potential mechanistic insight into known mutant phenotypes, and testable hypotheses for novel mutant phenotypes. More generally, the model provides a framework for integrating and interpreting ongoing experimental efforts that continue extend the rich history of biochemical research on Neurospora.
We reconstructed, validated and performed computational predictions with the Neurospora metabolic network model in a process consisting of four stages, as shown in Figure 1. Below we summarize the steps of the process, then we describe the optimization-based algorithms we developed to guide the process.
A number of significant challenges remain in the reconstruction of high-quality genome-scale metabolic models [38]. Although bioinformatic methods exist that can automate the generation of draft metabolic models [39], extensive manual adjustment and literature curation remains essential for generating high-quality models. The assessment of model accuracy through independent empirical validation is also critical if the predictions of the model are to be trusted. Although a number of methods have been developed to aid in this task [40]–[51], substantial manual effort is also still required.
To facilitate the automation of metabolic network reconstruction, we developed three optimization-based algorithms, which together comprise Fast Automated Reconstruction of Metabolism (FARM). These algorithms are: LInear MEtabolite Dilution Flux Balance Analysis (limed-FBA), which predicts flux while linearly accounting for metabolite dilution; Consistent Reproduction Of growth/no-growth Phenotype (CROP), which reconciles differences between in silico and experimental gene essentiality faster than previous approaches; and One-step functional Pruning (OnePrune), which removes blocked reactions with a single compact linear program.
Here we describe the results of our genome-scale metabolic reconstruction and its application as a predictive steady-state model. Following the accepted nomenclature [60] for naming metabolic models, we call the model N. crassa iJDZ836. N. crassa iJDZ836 is available in the Systems Biology Markup Language (SBML) in Dataset S1.
The model contains 836 genes that encode 1027 unique enzymatic activities. Of these enzyme-catalyzed reactions, 694 are supported by experimental evidence from 491 publications addressing Neurospora-specific enzymes. This level of evidence compares favorably with other highly curated models [61]–[63] as shown in Table 1. In addition, 16 spontaneous reactions and 331 orphan reactions were included based on the literature [64]–[66]. Our model contains 737 chemically unique metabolites. Of these metabolites, 673 have a defined structure that permits estimates of Gibbs free energy [67], [68]. Using these Gibbs free energy estimates, 1046 biochemical reactions were thermodynamically constrained to be irreversible, while the remaining 328 were assumed to be reversible. Of the 294 biochemical reactions that were associated with multiple proteins, 47 were manually curated as being catalyzed by an enzyme complex; we assumed the rest were catalyzed by isozymes. There are 257 metabolic pathways in the model. Of these, 134 are biosynthesis pathways, 96 are degradation/utilization/assimilation pathways, and 27 are pathways involved in the generation of precursor metabolites and energy. An overview of the pathways is displayed in Figure 3 and a zoomable metabolic map of the pathways is shown in Figure S3.
Cellular compartments in the model include the cytosol, the extracellular space, and 4 organelles: the glyoxysome, the vacuole, the nucleus, and the mitochondrion. The 299 transport reactions of the model enabled not only uptake and export of 137 metabolites, but also the exchange of metabolites between the cytosol and each organelle.
The model's growth objective was based on a modular biomass composition [69], [70]. Biomass modules were separately defined for DNA, RNA, amino acids, cell wall, lipids, sterols, essential cofactors, and secondary metabolites. This modular decomposition allowed for different goals in different applications of the model. For example, wild-type biomass contains substantial amounts of secondary metabolites, such as sphingolipids, ergosterol and carotenoids (which give Neurospora its characteristic orange color), so we included the secondary metabolites module in the biomass composition when predicting wild-type fluxes. On the other hand, secondary metabolites are not strictly required for viability, so we removed this module from the biomass composition when predicting viability.
The model quantitatively captures the growth rate of Neurospora. To illustrate this, we have plotted a range of glucose uptake rates against the model's predicted doubling times, and several data points extracted from the experimental literature [71]–[73] (Figure S4). The figure shows that our predictions closely match the experimental data.
To validate the accuracy of the N. crassa iJDZ836 model, we manually curated a collection of mutant viability phenotypes from the literature. We split this collection into a training set, which we used with FARM to construct the model; and an independent test set, which we used to validate the final model. Both of these collections are available in Table S1. To simulate gene knockout experiments, we removed reactions from the model that depend on the gene, applied limed-FBA, and predicted whether or not the perturbed model could grow. We then compared experimental observations to the model's in silico predictions. Accuracy was measured as two quantities: sensitivity and specificity. Sensitivity was defined as the proportion of experimentally viable mutants that were predicted to be viable in silico. Specificity was defined as the proportion of experimentally inviable mutants that were predicted to be inviable in silico.
The final model's predictive accuracy using limed-FBA is shown in Figure 4A. On the training set, we correctly predicted growth in 107 of 108 experimentally viable gene knockouts (sensitivity = 99.1%), and we correctly predicted no-growth in 44 of 47 experimentally lethal mutants (specificity = 93.6%). On the test set, we correctly predicted 270 of 289 experimentally viable gene knockouts (sensitivity = 93.4%), and we correctly predicted 13 of 14 experimentally lethal mutants (specificity = 92.9%).
The final model's predictive accuracy on the test and training sets using FBA and MD-FBA is shown in Figure 4B. Both of these methods are generally outperformed by limed-FBA.
The differences between FBA and limed-FBA can reveal subtle issues with existing experimental data. For example, because FBA allows recycling of coenzyme A and the biosynthesis of coenzyme A requires pan-2 [74], the essentiality of pan-2 is missed by FBA. In contrast, limed-FBA correctly predicts essentiality of pan-2. Similarly, three other genes involved in coenzyme A biosynthesis (NCU08925 and pan-3) and mitochondrial transport (mic-30) were predicted to be essential by limed-FBA, but not by FBA. Surprisingly, these three genes were in our test set of non-essentials. A potential explanation for these inconsistencies is that the function of these genes can be performed by an isozyme that was not captured in our model. In addition to these three genes, there is one more gene from our test set of non-essentials where limed-FBA predicted essentiality, but FBA predicted non-essentiality. This gene is pab-1. Like arg-14 (Figure 2C), pab-1 serves as input to a metabolic cycle, so pab-1 is not required by FBA. However, consistent with limed-FBA's prediction, the pab-1 gene was reported to be essential by Beadle and Tatum's original publication [9], [10].
In the left panel of Figure 4C, we compared the N. crassa iJDZ836 gene essentiality accuracy statistics on the training set to the reported accuracies Escherichia coli and S. cerevisiae models that were optimized using experimental observations of gene essentiality [41], [42]. In the right panel of Figure 4C, we compared the N. crassa iJDZ836 gene essentiality accuracy statistics on the test set to the accuracy statistics reported in the most recently published models of E. coli [63] and S. cerevisiae [62]. In all cases, the N. crassa iJDZ836 model prediction accuracies compare favorably, outperforming the specificities of the extensively trained models for E. coli and S. cerevisiae.
The prediction of gene essentiality for all genes in the model is available in Table S2.
To validate the ability of the N. crassa iJDZ836 model to predict nutrient supplements that would rescue auxotroph mutants, we manually curated a collection of nutrient rescue conditions from the literature. We split this collection into a training set, which we used with FARM to construct the model; and an independent test set, which we used to validate the final model. Both of these collections are available in Table S1. The predictions of nutrient rescues are available in Table S2. To simulate nutrient rescue experiments, we took a mutant that was predicted to be inviable on minimal media, supplemented the media with different nutrients, and applied limed-FBA to predict whether or not the mutant could grow in the supplemented media. We then compared experimental observations to the model's in silico predictions.
Of the 77 experimentally observed nutrient rescue conditions that we used as a training set, the model correctly predicted 74 (sensitivity = 96.1%)(Figure 5A; left panel). On the independent test set of 19 nutrient rescue conditions, the model correctly predicted 17 (sensitivity = 89.5%)(Figure 5A; right panel).
Pairwise synthetic lethality arises when two mutants with single gene knockouts are viable, but the double knockout mutant is inviable. Synthetic lethality can reveal cross-pathway dependencies that provide valuable insights into metabolism at the genome scale, but an experimental approach to comprehensively perform double knockouts to identify synthetic lethals is currently infeasible for Neurospora. Computational models provide a mechanism to rapidly and comprehensively test all interactions as a way to prioritize subsequent experimental verification. To predict synthetic lethality using the model, we simulated all pairs of knockouts of non-essential genes and predicted viability on minimal media.
Of the roughly 700,000 double knockouts in the metabolic model, the model predicted 230 to be synthetically lethal on minimal media. The mechanisms underlying these predicted synthetic lethal interactions fall into three classes: those that encode isozymes of a common reaction, those that encode enzymes of a common pathway, and those that encode enzymes of interacting pathways. This list contains 22 isozyme pairs, 4 gene pairs in the same pathway, and 204 gene pairs in interacting pathways. All of these pairs are tabulated in Table S2. The non-isozyme gene pairs and a previously known isozyme pair are displayed in a symmetric interaction map in Figure 9. This interaction map classifies each synthetic lethal pair by whether or not the two genes are in a common pathway or interacting pathways.
Building on its long history as a genetic model organism for biochemical genetics and cellular metabolism, we report here the first genome-scale metabolic network model for Neurospora crassa. We assessed the Neurospora metabolic model's ability to predict the impact of gene deletions, nutrient supplements that would rescue essential gene deletions, and synthetic lethal interactions. In each case, computational predictions were validated against a curated dataset of experimentally observed mutant viability phenotypes. Importantly, to ensure that our model was not over-fit, we separated the experimental data into a training and test set. Whereas training data was used in the development of the model, the test set was reserved to assess the accuracy of the final model. The final accuracy assessment was thus independent of any data used during model training.
The prediction of the growth phenotype of gene deletions is a canonical test of metabolic model accuracy and a useful benchmark for assessing the quality of different models [92]. The accuracy of our model compares favorably to extensively curated models such as S. cerevisiae and E. coli. Moreover, at 93% sensitivity and specificity on a test set of 303 phenotyped gene knockouts, the Neurospora model displays high absolute accuracy that lends confidence to the ability of the model to make accurate novel predictions.
The N. crassa iJDZ836 model also demonstrates high accuracy in predicting the ability of different nutrients to rescue essential gene knockouts and in predicting synthetic lethal interactions. In the former case, the model displays nearly 90% accuracy on an independent test set of nutrient rescue experiments. In the latter case, we were only able to curate a handful of experimentally verified synthetic lethal interactions. Nonetheless, although no synthetic lethal data was used during model training, the model was able to correctly identify four out of five known synthetic lethal interactions.
Genome-scale metabolic models complement experimental investigations, and one role of metabolic modeling is to rapidly generate and prioritize testable predictions that can be used to guide subsequent experimentation. As important as the predictions themselves, metabolic models also provide potential mechanistic explanations for the results. The explanations provide an important check on the overlying predictions. During the validation of models, these explanations ensure that not only are correct answers given, they are given for valid underlying reasons. For novel predictions, mechanistic explanations can provide potential insight into the results as well as tangible avenues to experimental validation. To illustrate the last point, in Text S4 and Figure S5 we simulate the observed physiological effect of oxygen limitation on ethanol production when grown on xylose. Therefore, our model can be used to simulate perturbations that optimize ethanol yield, which can then be verified experimentally.
As with all previous modeling efforts, errors in predicting known experimental results highlight limitations in either the model itself or the modeling framework. In terms of the model, the quality will only be as good as the information that was used to develop it. In the case of Neurospora, the extraordinarily rich literature for this well-studied model organism was the foundation that enabled a model to be generated that performed with high accuracy. Nonetheless, certain areas of the model remain less well developed, and one value of model construction is the objective measure it can provide on the relative information available for different aspects of metabolism. This can be used to target areas that are less well understood. For example, the substrates of certain reactions in the thiamin diphosphate and neurosporaxanthin biosynthesis pathways and the fate of the end-product in the histidine degradation pathway cannot be included with confidence in any metabolic model, because they are open biochemical questions [93].
More generally, the constraint-based modeling framework we used here is known to suffer from certain limitations. As with similar models, this accounts for a significant portion of the prediction errors in the Neurospora model. In particular, our model does not account for regulation of either enzyme expression or activity. These factors sometimes acted in combination. An illustrative example is gln-1 and gln-2, which code for the alpha and beta subunit, respectively, of glutamine synthetase [94]. Our model requires both subunits for enzyme catalysis. However, it was experimentally shown that concentration of extracellular ammonium regulates this enzyme's subunit composition, which can include both subunits, only alpha subunits, or only beta subunits [95]. This metabolic complexity highlights the need for the future incorporation of kinetics and regulation.
In one instance, however, a prediction initially thought to be an error provided the means to identify an issue with an experimentally observed knockout. The viability phenotype experiment for Δerg-14 was performed on a knockout strain originally designated as a homokaryon. Experimental observations of this strain revealed a normal growth phenotype. In contrast, the model predicted that the Δerg-14 mutant was blocked in the production of mevalonate, which is a necessary precursor for the sterol component of biomass. Moreover, previous efforts to phenotype temperature-sensitive mutants of erg-14 revealed severe morphological defects that were expected to be lethal in the full knockout [81]. Driven by these inconsistencies, a re-examination of the Δerg-14 knockout revealed that the mutant used was in fact a heterokaryon. This prediction, in effect, served as a blind control that highlighted the predictive value of the model.
The construction of genome-scale metabolic models remains a daunting task. Even aided by sophisticated tools for the management and visualization of pathway knowledge, a metabolic reconstruction still requires substantial manual review of the corresponding literature [38]. Moreover, it is desirable that the model construction process be guided by objective and quantitative measures of predictive accuracy. Incorporating this requirement into the model generation process increases the complexity of the task by requiring iterative cycles of data curation, model improvement, and accuracy assessment. To facilitate the process of model improvement, a number of tools have been developed [39]–[47], [49], [51]. We contribute to this set of tools with the development of a set of optimization-based algorithms, which together comprise Fast Automated Reconstruction of Metabolism (FARM).
Two of the three FARM algorithms specifically facilitate the process of model construction. Consistent Reproduction Of growth/no-growth Phenotype (CROP) assists in automating the process of adding and subtracting reactions from a model to improve predictive accuracy. CROP integrates diverse evidence for pathways into a probabilistic framework that assigns a weight to each reaction associated with the likelihood that the reaction is present in the network. These weights are then used to guide the selection of reactions to add or remove. Previous methods to achieve in silico growth used mixed integer linear programming (MILP), and thus required substantial compute time [39], [42], [44]. For CROP, we utilize LP relaxation, which is faster by orders-of-magnitude. Additional details for this algorithm along with comparisons to GrowMatch [41], [42] and Model SEED [39], [40] are available in Text S3.
OnePrune was developed to efficiently solve the problem of removing reactions that are blocked. OnePrune utilizes the goal programming optimization framework to achieve multiple competing objectives. The advantage of this framework is that once an individual objective is achieved, the optimization can pursue other objectives. OnePrune's goals are to send flux through as many reactions as possible, so once a reaction has achieved a nonzero flux, OnePrune is free to pursue flux through other reactions. Thus, OnePrune identified which reactions could carry flux with a single compact linear program.
The third FARM algorithm, limed-FBA, is an enhancement to the FBA method that improves predictive accuracy. limed-FBA accounts for the dilution of active metabolites that is ignored by FBA (Figure 2B), so limed-FBA is able to correctly identify the essentiality of reactions that are typically missed by standard FBA. For example, FBA predicted that pab-1 was not essential, because pab-1 serves as input to a metabolic cycle; however, limed-FBA predicted that pab-1 was essential. In fact, essentiality of pab-1 was shown by the original experiment of Beadle and Tatum on Neurospora crassa [9], [10].
The metabolic reconstruction of Neurospora crassa was performed in accordance with previously described protocols [96], [97], and is detailed in Text S2.
From the Neurospora crassa genome assembly NC10, gene boundaries were predicted using the Calhoun annotation system [14]. For all enzymes, we obtained the probability that the enzyme catalyzes a particular biochemical reaction, characterized by its Enzyme Commission (EC) number, from the enzyme function prediction software EFICAz [33].
We used the Pathway Tools software suite [93] to create the NeurosporaCyc Pathway/Genome Database (PGDB) and to manage curated data. We added functional gene annotations with associated Gene Ontology (GO) terms [36] and literature citations that were manually curated by the Community Annotation Project [27]. EC numbers, GO terms, and functional annotations were used as input to Pathologic [35] to automatically infer pathways from MetaCyc. These pathways comprised the initial NeurosporaCyc PGDB. Reaction directions were based on the Gibbs free energy predictions using the group contribution method [67]. Enzyme complexes were manually curated using the Pathway Tools protein complex editor and evidence from the Neurospora literature [93]. Transporters were automatically predicted from the genome sequence by the Transporter Automatic Annotation Pipeline (TransAAP) from TransportDB [98], and from the genome annotation using the Transport Identification Parser (TIP) [99]. Cellular compartment information was described using the Cellular Component Ontology (CCO), available at http://bioinformatics.ai.sri.com/CCO/. Neurospora-specific enzyme kinetics, allosteric regulation, biomass composition, and growth media were added during manual curation of the experimental evidence in the literature.
Before the NeurosporaCyc PGDB could be used to generate a working model, a number of data-cleaning steps were performed. Reactions were mass balanced, and the individual metabolites were protonated to the intracellular pH of 7.2 [100]. For reactions containing compound classes (e.g. “an alcohol”), the compound class was replaced by its instances (e.g. “ethanol”) that render the equation mass-balanced. Polymerization pathways such as fatty acid beta oxidation were either lumped into a summary reaction, or instantiated into a chain of individual polymerization steps. For polymerization reactions, we specified an arbitrary maximum polymer size n and created a lumped reaction that was stoichiometrically equivalent to n steps of the polymerization. The NeurosporaCyc Pathway/Genome Database can be downloaded from the PGDB registry at http://biocyc.org/registry.html and is available online at http://neurosporacyc.broadinstitute.org.
To curate our growth rate data, we identified several manuscripts that included glucose concentrations for Neurospora grown on glucose minimal medium where doubling times could be inferred from growth curves [71] or were given [72], [73]. The glucose concentrations were converted to glucose uptake rates using data for the derepressed system in Figure 1 of Schneider and Wiley [101].
To curate our viability phenotype data, we primarily relied on two resources. The first was The Neurospora crassa e-Compendium, curated by Alan Radford. The e-Compendium contains >2,400 citations and >3,000 gene entries. Numerous gene entries have associated mutant phenotypes extracted from the literature. In some cases, these phenotypes also include supplements that rescue no-growth mutants. Because this resource primarily lists phenotypes from mutants that are not knockouts, viability of a mutant could be due to non-essentiality of a mutated gene or to partial efficacy of a mutated enzyme. Thus, this resource did not clearly identify non-essential genes. The second resource was knockouts from the Neurospora Genome Project [27]. Knockouts that did not germinate and grow in a short time were not further evaluated to determine whether they showed low growth or no growth, so this project did not clearly identify essential genes.
To collect the essential gene sets, we first identified inviable mutants in the e-Compendium, and only retained genes for which we could manually verify their essentiality in the literature. To split the essential genes into a training set and a test set we intersected the list of inviable mutants in the e-Compendium with the list of heterokaryon knockouts from the Neurospora Genome Project (http://www.dartmouth.edu/~neurosporagenome/knockouts_completed.html). Genes that were in the intersection became the test set, and the rest remained in the training set.
To collect the non-essential gene sets, we used the homokaryon knockouts from the Neurospora Genome Project [82], all of which were experimentally observed to be viable in Vogel's minimal media [102]. A subset of these homokaryons were extensively phenotyped, and these were available at the Neurospora crassa Database at the Broad Institute (http://www.broadinstitute.org/annotation/genome/neurospora/Phenotypes.html). This subset became our non-essential training set, while the rest of the homokaryon knockouts became our non-essential test set. Homokaryon knockouts that were in the essential gene set were discarded.
Both the supplemental nutrient rescue training and test sets were initially identified from the e-Compendium and from the book Neurospora: Contributions of a Model Organism [1]. They were then confirmed through manual curation of the associated citations. We also used this protocol to identify and confirm known synthetic lethal mutations.
We simulated Vogel's minimal media (http://www.fgsc.net/methods/vogels.html) [13] with sucrose by including exchange reactions for each metabolite in the media, and limiting sucrose uptake to 1.5 mmol/(gram Dry Weight * hour). Alternative media were formulated in silico by adding/removing exchange reactions. Supplemental nutrients were limited to a flux of 3 mmol/(gram Dry Weight * hour).
We simulated gene knockouts by removing reactions that require knocked-out genes. Biomass flux was predicted using limed-FBA, except where stated otherwise. In silico growth phenotypes were regarded as viable if the biomass flux exceeded 0.02, and inviable otherwise. FBA was run using the optimizeCbModel function from the COBRA Toolbox 2.0.5 [103] using Gurobi 5.0 in Matlab (MathWorks, Natick, MA). MD-FBA was run using Matlab code downloaded from Tomer Shlomi's Research Group's website (http://www.cs.technion.ac.il/~tomersh/methods.html) with Tomlab v7.9 and CPLEX 12.2. This code was modified for our model: all MD-FBA constraints could be satisfied by wild-type on Vogel's minimal media, so we did not need to apply their addNewExReactions function. This code allowed 400 seconds per optimization.
We exported the model to the COBRA-compatible subset of the Systems Biology Markup Language (SBML) [106], [107]. All SBML identifiers were based on NeurosporaCyc Frame IDs. To avoid using characters disallowed in SBML identifiers, we implemented a substitution scheme. We substituted disallowed characters with their ASCII equivalent, demarcated on both sides by two underscores. For example, “[” has ASCII value 91, so it was substituted by “__91__”. Identifiers beginning with a number were given a prefix of underscore; e.g. “1diacyl” became “_1diacyl”. So that our SBML file can be reversed-transformed into its original character encoding, we wrote an extension to the COBRA toolbox in Matlab. This extension is available at http://code.google.com/p/fast-automated-recon-metabolism. SBML Notes fields contain COBRA-compliant gene associations, pathways, EC numbers, Pubmed IDs, chemical formulae, charge, and NeurosporaCyc and KEGG identifiers [107]. SBML Annotation fields contain MIRIAM-compliant links to InChI identifiers [108], [109].
The metabolic model (Dataset S1) has been deposited at the BioModels Database [110] with accession MODEL1212060001, and is available on the web at http://neurosporacyc.broadinstitute.org.
|
10.1371/journal.pntd.0000967 | Immunological and Viral Determinants of Dengue Severity in Hospitalized Adults in Ha Noi, Viet Nam | The relationships between the infecting dengue serotype, primary and secondary infection, viremia and dengue severity remain unclear. This cross-sectional study examined these interactions in adult patients hospitalized with dengue in Ha Noi.
158 patients were enrolled between September 16 and November 11, 2008. Quantitative RT-PCR, serology and NS1 detection were used to confirm dengue infection, determine the serotype and plasma viral RNA concentration, and categorize infections as primary or secondary. 130 (82%) were laboratory confirmed. Serology was consistent with primary and secondary infection in 34% and 61%, respectively. The infecting serotype was DENV-1 in 42 (32%), DENV-2 in 39 (30%) and unknown in 49 (38%). Secondary infection was more common in DENV-2 infections (79%) compared to DENV-1 (36%, p<0.001). The proportion that developed dengue haemorrhagic fever (DHF) was 32% for secondary infection compared to 18% for primary infection (p = 0.14), and 26% for DENV-1 compared to 28% for DENV-2. The time until NS1 and plasma viral RNA were undetectable was shorter for DENV-2 compared to DENV-1 (p≤0.001) and plasma viral RNA concentration on day 5 was higher for DENV-1 (p = 0.03). Plasma viral RNA concentration was higher in secondary infection on day 5 of illness (p = 0.046). We didn't find an association between plasma viral RNA concentration and clinical severity.
Dengue is emerging as a major public health problem in Ha Noi. DENV-1 and DENV-2 were the prevalent serotypes with similar numbers and clinical presentation. Secondary infection may be more common amongst DENV-2 than DENV-1 infections because DENV-2 infections resulted in lower plasma viral RNA concentrations and viral RNA concentrations were higher in secondary infection. The drivers of dengue emergence in northern Viet Nam need to be elucidated and public health measures instituted.
| Dengue is estimated to affect 50 million people each year and can occur as explosive outbreaks that overwhelm health systems. Despite significant advances the available knowledge is not sufficient to predict the outcome of individual infections or the occurrence of epidemics. Studies from low dengue transmission settings are lacking but offer the potential to better understand the contribution of age, primary versus secondary infection and serotype because there are likely to be more adult and primary infection patients and fewer serotypes circulating compared to high transmission settings. This is the first reported study of clinical dengue in Ha Noi, the largest urban area of Northern Viet Nam. Records kept by the Preventive Medicine Center indicate that <2500 clinical dengue cases attended government health care facilities in Ha Noi each year from 1999 until 2007. Patients in Ha Noi were older than in high transmission settings, the contribution of primary infection to overt and severe illness was greater and associations between serotype, plasma viral RNA concentration and overt and severe illness were distinct. The dengue situation in Ha Noi provides an opportunity to further examine the roles of serotype and prior immunity in dengue severity and epidemic emergence.
| Dengue virus (DENV) infections range in severity from asymptomatic to a syndrome characterized by a haemorrhagic tendency and vascular permeability [1]. The events that precipitate endothelial cell dysfunction and vascular leak are incompletely understood. Numerous studies including several prospective cohorts [2], [3], [4] demonstrate that the risk of severe dengue is higher during a secondary infection with a new serotype in children [5], [6], [7]. Severe dengue has also been associated with high viral loads [8], [9], [10], [11], [12], prolonged viremia [13] and high NS1 antigen levels [12], [14]. At sub-neutralizing concentrations, dengue specific antibodies can enhance dengue virus infection of mononuclear phagocytes [15]. In addition, antibodies to pre-membrane protein appear to enhance infection of all serotypes, even when present at high concentration [16]. It has therefore been proposed that antibody-dependent enhancement (ADE) of viremia is a risk factor for severe dengue [1]. However, ADE may not fully account for severe dengue as several studies have found no association between severity and secondary infection in adults [7], [17], [18], [19], or severity and viral load [20], [21], [22], [23]. In one study dengue virus titers were higher prior to defervescence in patients with secondary infection [8] but most other studies have found that titers are either similar or higher in primary infection [10], [17], [20], [24]. Demonstration of a link between enhancing antibody levels, viral load and disease severity in humans also remains elusive.
The emerging picture is that multiple factors including prior immunity, viral load, age of the patient and infecting serotype and genotype may contribute to the severity of dengue infection [2], [3], [25], [26] but the nature of these interactions remains unclear. Dengue pathogenesis has largely been studied in dengue hyper-endemic regions where analysis is “confounded by a multiplicity of preexisting immunity patterns coupled with co-circulation of multiple serotypes” [3], [27]. Studies in low transmission settings, where few dengue serotypes circulate and primary infection in adults is common, potentially offer an opportunity to better identify factors associated with severity across serotype and immunity groups. We conducted a prospective study in Ha Noi, Viet Nam, to examine the association between primary and secondary infections, serotype, plasma viral RNA concentration, and the development of dengue haemorrhagic fever (DHF) in a low transmission setting.
The protocol for this study was approved by the scientific and ethical committees at the National Hospital of Tropical Diseases and The Oxford University Tropical Research Ethics Committee (OXTREC). Patients provided written informed consent to participate in this study.
Patients were eligible for recruitment if they were admitted to the National Hospital of Tropical Diseases (NHTD) in Ha Noi, Viet Nam between September and November 2008 with a clinical diagnosis of dengue according to the WHO criteria [28]. These criteria were fever plus two or more of the following: headache; retro-orbital pain; myalgia/arthralgia; rash; bleeding or leukopenia. The study protocol included children but the number attending the National Hospital of Pediatrics (NHP) in Ha Noi during the study period was too low to warrant investigation. The NHTD is a 160 bed tertiary care center for adult patients with infectious diseases and also serves as a referral center for dengue in Northern Viet Nam. The NHP is the coordinating center for pediatric care in the country, receives patients from all Northern provinces and sees on average 40,000 in-patients and 350,000 out-patients per year.
Patients were examined daily during hospitalization by a dedicated team of physicians with experience in dengue diagnosis and treatment. Signs and symptoms of hemorrhage, capillary permeability and shock along with other relevant clinical data were prospectively recorded using standardized case record forms. Ultrasound of the chest and abdomen was performed at study enrolment and additionally when clinically indicated. Full blood counts were performed daily for at least 6 days during admission, as well as at discharge and approximately 10 days after discharge. Other investigations and clinical management were at the discretion of the attending physicians. CRFs and laboratory results were reviewed to identify patients that fulfilled the WHO criteria for dengue hemorrhagic fever (DHF), i.e. plasma leakage, plus thrombocytopenia and hemorrhagic signs [28]. Plasma leakage was said to be evident if pleural effusion and/or ascites were detected by ultrasound, or if a haematocrit during admission was ≥20% higher than at follow-up. If the patient did not attend follow-up, the average of follow-up values for males (n = 60, average = 44) or females (n = 56, average = 38) was used. This compares to a normal haematocrit value of 38% with a range of 35–41% set by the Ministry of Health of Viet Nam. Thrombocytopenia was defined as a platelet count less than 100,000/mm3.
An in-house IgM & IgG capture ELISA using antigens from DENV 1-4 and monoclonal antibodies provided by Venture Technologies (Sarawak, Malaysia) was performed as previously described [29]. A sample was considered IgM or IgG positive if the units were at least 6 times higher than the negative control sera. An internally controlled, serotype-specific, real-time reverse-transcriptase polymerase chain reaction (RT-PCR) assay [30] was used to identify the infecting serotype and determine viral RNA concentrations expressed as cDNA equivalents/ml of plasma. The sequences of the dengue serotype-specific primers and probes have been published previously [30]. They amplify and detect parts of the NS5 coding region that were first identified by Laue et al as being highly conserved within each dengue serotype [31]. The assay limit of detection was 10 cDNA equivalents per reaction. Dengue NS1 antigen was detected using a commercial ELISA (BIO-RAD Platelia™ Dengue NS1 Ag) according to the manufacturer's instructions. A diagnosis of confirmed dengue was made using a previously described reference algorithm [32] that has been adapted to include NS1 ELISA and remove the indirect recombinant membrane protein ELISA, which was not used in this study (Figure S1). Using this algorithm a patient is considered to have confirmed dengue if either RT-PCR or NS1 ELISA is positive, if there is an increase in the level of IgM detected by ELISA or an IgG ELISA conversion in the presence of a positive IgM ELISA. Serology was considered to be consistent with primary dengue infection if on or after day 6 of illness IgM levels were at least 1.78 times higher than IgG levels [33], or with secondary infection if IgM levels were less than 1.2 times higher than IgG levels.
Illness day was calculated from the first date that the patient recalled having fever, which was assigned as day 1. Proportions were compared using odds ratios and Chi-Square or Fishers exact test when any expected cell count was less than 5. Continuous variables were presented as medians and interquartile ranges (IQR) and compared using Kruskal-Wallis and Mann Whitney tests. Data for study patients was compared to records kept by the Ha Noi Preventive Medicine Center (Ha Noi PMC), which includes age, gender, province and district of people attending government health care facilities with a clinical diagnosis of dengue by month and year.
We modelled dengue severity as depending on serological definition of prior infection/immunity and infecting serotype (and a potential interaction) using simple and multiple logistic regression analyses. The probability of a positive NS1 result and the log-10 plasma viral RNA measurements on day 5 of illness (i.e. the median illness day when patients were admitted) were modelled using logistic and linear regression models, respectively, depending on serological definition of prior infection, infecting serotype and dengue severity. In a sensitivity analysis, the model was additionally adjusted for age and gender. The time from illness onset to the first undetectable NS1 and viral RNA measurement, respectively, were modelled using Weibull accelerated failure time regression models for interval censored data, i.e. patients were treated as reaching undetectable levels in the interval between their last positive and their first undetectable measurement and patients for whom the first measurement was undetectable were treated as first reaching undetectable levels between illness day 1 and the day of this first measurement.
Analyses were performed with the statistical software R version 2.9.1 [34] and SPSS for Windows, Rel. 14.0.0.245, 2005 (SPSS Inc. Chicago IL.).
158 of 206 eligible patients consented and enrolled in an 8 week period commencing on September 16 2008. During this time approximately 1240 clinical dengue cases attended government health care facilities in Ha Noi out of a total of 2371 for the whole of 2008 of which 975 (41%) were admitted to NHTD. 139 patients were from Ha Noi province, 11 were from 4 neighboring provinces and 6 were from more distant provinces, the farthest being Nhge An, Son La and Quang Ninh which are more than 100 km from Ha Noi. None of the patients were from provinces north of Ha Noi. Laboratory diagnosis of dengue was made for 130 patients. A further 18 with probable dengue were not included in this analysis as these were recruited significantly later in their illness and 61% had already defervesced. 26% of all enrolled patients and 23% of those with confirmed dengue had been transferred from another hospital. The age and gender distribution of confirmed dengue patients (Table 1) were similar to that for dengue cases attending any government health care facility in Ha Noi in 2008 (median age 23 years, IQR 18–31, 52% male). The geographic distribution was also similar to that for all reported cases in Ha Noi (data not shown). Amongst confirmed cases serology was indicative of secondary infection in 61% and of primary infection in 34% (Table 1). The infecting serotype could be defined by real-time-RT-PCR for 81 patients of which 52% had DENV-1 and 48% had DENV-2 (Table 1). Viral RNA could not be detected by RT-PCR in 49 confirmed dengue patients. The median admission day for these patients was 1 day later than for those with virus RNA detectable by RT-PCR (Table 1). The E genes of 9 DENV-2 (GenBank: GU908512- GU908520) and 20 DENV-1 (GenBank: HQ591537-HQ591556) viruses were sequenced and belong phylogenetically to the Asian 1 genotype and Genotype I, respectively (unpublished findings). Secondary infection was significantly more common in DENV-2 patients (79%) compared to DENV-1 patients (36%, p<0.001). Age, sex and course of illness were similar across serotype and serology subgroups (Table 1). Most patients (91, 70%) were classified as dengue fever (DF) and 36 (28%) developed DHF, of which 5 were classified as grade I, 30 as grade II and 1 as grade III. All patients were well at discharge except one patient who was transferred to the surgical hospital.
In patients with DHF platelet counts fell to their lowest levels and haematocrits increased by the greatest percentage over baseline on days 5–7 (Figure 1). DHF rates were similar for DENV-1 and DENV-2 patients (Table 1). There was a non-significant trend of higher DHF rates in patients with secondary compared to primary infection (Odds Ratio 1.96, 95% CI 0.80 – 4.85, p = 0.14). Results were consistent when including both serotype and serological definition of primary versus secondary infection in a logistic model and after adjusting for age and sex (data not shown), and there was no evidence of an interaction between serotype and primary/secondary infection status in the development of DHF (Likelihood ratio p = 0.48).
The proportion NS1 positive on day 5 of illness was significantly higher for DENV-1 compared to DENV-2 patients and the time to undetectable NS1 was shorter for DENV-2 (Table 2, Figure 2). The proportion NS1 positive on day 5 was similar for patients with primary and secondary infection but the time to undetectable was shorter for secondary infection. There was no evidence of an interaction between serotype and primary/secondary infection status (Likelihood ratio = 0.24) and effects were consistent when the analysis was additionally adjusted for age and gender. There was also no evidence of any association between NS1 and DHF (Table 2).
Log10 viral RNA concentration in plasma on day 5 was estimated to be 0.96 lower for DENV-2 than for DENV-1 patients and the time taken until viral RNA was undetectable in plasma was estimated to be 0.89 times shorter for DENV-2 (Table 3, Figure 3). Plasma viral RNA concentration on day 5 was significantly higher in patients with secondary infection, but the time taken until viral RNA was undetectable in plasma was not significantly different. Results were consistent after adjusting for age and sex (data not shown) and likelihood ratio tests showed no evidence of interaction between serotype and primary/secondary infection status for either viral RNA concentration on day 5 (p = 0.96) or time taken until viral RNA was undetectable (p = 0.85). We couldn't establish a clear association between DHF and plasma viral RNA concentration on day 5 or time taken until viral RNA was undetectable in plasma.
Hospitalized dengue patients in Ha Noi in 2008 were predominantly adults with high rates of primary infection compared to Southern Vietnam [35] and other hyper-endemic regions [36], [37]. DENV-1 and DENV-2 were the only serotypes identified, consistent with national dengue surveillance data for Ha Noi (Le Quynh Mai, personal communication), whereas in Southern Viet Nam all serotypes were detected during clinical dengue surveillance in 2008 but DENV-1 was predominant [38]. Overall case numbers and clinical presentation were similar for DENV-1 and DENV-2. However, while primary infection predominated amongst DENV-1 patients suggesting that a substantial proportion of adults are dengue naïve, secondary infection predominated in DENV-2. This suggests that primary DENV-2 infections may be less likely to present clinically.
A limitation of this study is that patients were recruited from only one hospital in Hanoi. However, the patients studied represented ∼12% of dengue cases seeking treatment for dengue at government health care centers in Ha Noi during the study period and had similar epidemiology. We did not include children as originally intended because only 5 were admitted to the National Hospital of Pediatrics with clinically suspected dengue while the study was being conducted. This may be expected if transmission is low such that secondary infections mainly occur in adulthood. However it is not clear why children with primary infection do not present given that 34% of adult patients had primary infection. Others report that DHF rates during primary infection are higher for adults and suggest that primary infection is more severe in adults [39]. Patients that admitted late were not precluded from this study in order to obtain a comprehensive description of clinical dengue, however this limited our ability to detect viral RNA in plasma and determine peak concentrations. The infecting serotype was unknown for 38% of confirmed dengue patients most of whom admitted 5 to 6 days after illness onset, and by day 6 viral RNA could be detected in only 39 of 100 confirmed patients tested. We suspect that DENV-2 will be the predominant infecting serotype in this group because viral RNA clearance from plasma was faster for DENV-2 than for DENV-1 patients in our study.
Similar to the findings of this study, records kept since 1988 indicate that Dengue case numbers have been low in Ha Noi and that dengue has predominated in adults (Horby P. et al, in preparation). This may not reflect transmission because the proportion of infections that present clinically can be low and the proportion and age-distribution depends on the prevalent serotypes and the age-associated prevalence of past infection with each serotype [3], [40], [41]. However, the age of clinical dengue cases generally increases with decreasing transmission intensity [42], and the epidemiology of hospitalized dengue in this study is similar to that in Singapore where transmission has decreased due to effective vector control but the age and proportion of cases with primary dengue has increased, presumably because adults are more prone to present clinically upon primary infection [40].
Others have reported that primary DENV-2 infections are rarely symptomatic [6], [36], [37], [43], [44]. The reason for this has not been established but our data shows that viral RNA concentration is low and NS1 detection brief in the plasma of DENV-2 compared to DENV-1 patients, factors that could be considered important in disease pathogenesis leading to severe dengue. Furthermore, the relatively high prevalence of secondary DENV-2 coincided with higher plasma viral RNA concentrations in secondary infection. It is important to note that the DENV-2 viruses sequenced in this study belonged to the Asian I genotype, which has been associated with more severe disease compared to the American genotype [26] and with higher plasma viral RNA concentrations compared to Asian/American genotype [38]. While the results suggest that primary infection with DENV-1 is more likely to lead to clinically overt disease than with DENV-2, we can not exclude the possibility that secondary infection contributes to overt DENV-1 or the possibility that DENV-2 infections are more likely to be enhanced than DENV-1 infections. The latter has been suggested elsewhere [3] and is supported by several studies showing that in children with secondary infection DHF is more common for DENV-2 compared to DENV-1 [4], [36]. DHF was approximately twice as common in secondary compared to primary infection in our cohort, but the number of patients with DHF was small and this did not reach statistical significance or permit analysis within each serotype. However, in a study of hospitalized patients in Thailand the association between secondary infection and DHF was greater for DENV-2 than DENV-1 because DHF was less common in DENV-2 compared to DENV-1 during primary infection [34]. As in our study, this suggests that primary DENV-2 infections may be less virulent than DENV-1. As discussed above, we suspect that DENV-2 would have been the predominant infecting serotype amongst confirmed dengue patients in which the infecting serotype was unknown. NS1 measurements and prior infection status were similar for serotype-unknown and DENV-2 patients and distinct from DENV-1 patients providing further indication that DENV-1 may be distinct in terms of virulence during primary infection.
There was a non-significant trend of increased DHF in secondary infection, but 22% of DHF cases had primary infection. Thus secondary infection was not essential for DHF in this cohort. Secondary infection was associated with higher viral RNA concentration in plasma on day 5 of illness, but we did not find an association between viral RNA concentration and DHF. Interpretation of the effect of viral RNA concentration on DHF in our patients is limited by the relatively low proportion that developed DHF and perhaps also the high proportion that presented after day 3 of illness. Several of the studies that find an association between viremia and DHF recruit children within the first 3 days of illness and suggest that DHF is positively associated with peak viremia [8], [9], [11]. It remains controversial whether virus clearance times also differ. In one study clearance of infectious virus determined by mosquito inoculation was faster in children with DHF compared to DF [8] but studies of adults with DENV-2 or DENV-3 infection [10], [13] and children with DENV-2 infection [24] have found longer times to virus RNA or virus-RNA containing immune complex clearance amongst those with DHF[10], [13]. It is also possible that we did not detect an association between DHF and plasma viral RNA concentration because a relatively high proportion of our patients had primary DENV-1 and studies where DHF has been associated with viremia rarely include primary DENV-1 [9], [10], [11], [13], [14], [24] or reported that there was no association in patients with primary DENV-1 [8]. The contribution of plasma viral RNA concentration to the development of DHF may be not be discernable in primary DENV-1 because plasma viral RNA concentration is generally high in DENV-1, but this would imply that high viral RNA concentration alone is not sufficient to cause DHF.
The time until NS1 was undetectable was longer for DENV-1 compared to DENV-2, similar to findings of an earlier study in Southern Viet Nam [32]. We previously suggested that this reflected a predominance of primary infection in DENV-1 and that NS1 clearance is faster in secondary infection due to sequestration by IgG [32], [45]. In the current study there were sufficient primary cases for a stratified analysis, which demonstrated that serotype is the main determinant of the sensitivity of NS1 tests, and this should be considered when interpreting NS1-based diagnostics.
In conclusion our results indicate an association between secondary infection and clinically overt DENV-2 infection. Higher plasma viral RNA concentration in secondary infection may underlie the association between secondary infection and overt DENV-2. We could not detect an association between DHF and secondary infection or plasma viral RNA concentration but this may be due to the relatively high proportion of patients with primary DENV-1, a situation that may change if dengue emerges and the proportion and age of the population that is dengue naïve declines.
The number of countries affected by dengue has increased six-fold in the last 30 years with potential for further spread through temperate, subtropical and tropical areas [46]. The Ha Noi Preventive Medicine Centre reported a 7-fold increase in the number of clinical dengue cases from 2008 to 2009 and this unforeseen epidemic overwhelmed the health system. A similar problem is faced in regions where dengue has been endemic for decades due to large multi-annual peaks in severe disease incidence [47]. Current understanding of the drivers of dengue epidemics is inadequate to predict their occurrence and inform public health prevention and preparedness measures. The dengue situation in Ha Noi provides an opportunity to further examine the roles of serotype infection sequence and prior immunity in dengue severity and emergence.
|
10.1371/journal.pgen.1000095 | Trends in Selenium Utilization in Marine Microbial World Revealed through the Analysis of the Global Ocean Sampling (GOS) Project | Selenium is an important trace element that occurs in proteins in the form of selenocysteine (Sec) and in tRNAs in the form of selenouridine. Recent large-scale metagenomics projects provide an opportunity for understanding global trends in trace element utilization. Herein, we characterized the selenoproteome of the microbial marine community derived from the Global Ocean Sampling (GOS) expedition. More than 3,600 selenoprotein gene sequences belonging to 58 protein families were detected, including sequences representing 7 newly identified selenoprotein families, such as homologs of ferredoxin–thioredoxin reductase and serine protease. In addition, a new eukaryotic selenoprotein family, thiol reductase GILT, was identified. Most GOS selenoprotein families originated from Cys-containing thiol oxidoreductases. In both Pacific and Atlantic microbial communities, SelW-like and SelD were the most widespread selenoproteins. Geographic location had little influence on Sec utilization as measured by selenoprotein variety and the number of selenoprotein genes detected; however, both higher temperature and marine (as opposed to freshwater and other aquatic) environment were associated with increased use of this amino acid. Selenoproteins were also detected with preference for either environment. We identified novel fusion forms of several selenoproteins that highlight redox activities of these proteins. Almost half of Cys-containing SelDs were fused with NADH dehydrogenase, whereas such SelD forms were rare in terrestrial organisms. The selenouridine utilization trait was also analyzed and showed an independent evolutionary relationship with Sec utilization. Overall, our study provides insights into global trends in microbial selenium utilization in marine environments.
| Selenium (Se) is an essential micronutrient due to its requirement for biosynthesis and function of the 21st amino acid, selenocysteine (Sec). Sec is found in the active sites of selenoproteins, most of which exhibit redox function, in all three domains of life. In recent years, genome sequencing projects provided a large volume of nucleotide and protein sequence information. Identification of complete sets of selenoproteins (selenoproteomes) of individual organisms and environmental samples is important for better understanding of Se utilization, biological functions of this element, and changes in Se use during evolution. Here, we describe a comprehensive analysis of the selenoproteome of the microbial marine community derived from the Global Ocean Sampling (GOS) expedition. More than 3,600 selenoprotein gene sequences belonging to 58 protein families were detected and analyzed. Our study generated the largest selenoproteome reported to date and provided important insights into microbial Se utilization and its evolutionary trends in marine environments.
| Selenium (Se) is an essential trace element that exerts a number of health benefits yet is required only in small amounts [1]–[3]. It is incorporated into selenoproteins, many of which are important antioxidant enzymes, in all three domains of life, and occurs in these proteins in the form of selenocysteine (Sec), the twenty-first amino acid in the genetic code [4]–[6]. Sec insertion is specified by a UGA codon, which is normally read as a stop signal. The decoding of UGA as Sec requires a translational recoding process that reprograms in-frame UGA codons to serve as Sec codons [5]–[8]. The mechanisms of selenoprotein biosynthesis have been the subject of numerous studies [5], [7]–[12]. The translation of selenoprotein mRNAs requires both a cis-acting selenocysteine insertion sequence (SECIS) element, which is a hairpin structure residing in 3′-untranslated regions (3′-UTRs) of selenoprotein mRNAs in eukaryota and archaea, or immediately downstream of Sec-encoding UGA codons in bacteria [7], [13]–[16], and several trans-acting factors dedicated to Sec incorporation [7],[17].
In recent years, an increase in the number of genome sequencing projects combined with the rapidly emerging area of microbial metagenomics provided vast amount of gene and protein sequence data. However, selenoprotein genes are almost universally misannotated in these datasets because of the dual function of UGA codon. To address this problem, a variety of bioinformatics approaches have been developed and used for selenoprotein searches in both prokaryotes and eukaryotes [18]–[24]. With these programs, researchers successfully identified complete sets of selenoproteins (selenoproteomes) of individual organisms and environmental samples [20]–[26].
In early 2007, three papers from the J. Craig Venter Institute were published reporting the results of the first phase of the large-scale metagenomic sequencing project – Global Ocean Sampling (GOS) expedition, a comprehensive survey of bacterial, archaeal and viral diversity of the world's oceans [27]–[29]. The general objective of this project was to expand our understanding of the microbial world by studying the gene complement of marine microbial communities. A metagenomics approach was used to sequence DNA isolated from selected sites of the aquatic microbial world. The previous Sargasso Sea project [30], which reported environmental shotgun sequencing of marine microbes in the nutrient-limited Sargasso Sea, was considered as a pilot study for the subsequent GOS project. The GOS dataset encompasses 44 sequenced samples from diverse aquatic (largely marine) locations which contain a total of ∼7.7 million sequencing reads and more than 8 billion nucleotides [29]. These data not only provide opportunities for the identification and characterization of genes, species and communities, but have potentially far-reaching implications for biological energy production, bioremediation, and creating solutions for reduction/management of greenhouse gas levels.
Within this framework, identification and characterization of selenoproteins in such a huge metagenomic dataset can shed light on the roles of Se in marine microbial communities. Previously, we examined the microbial selenoproteome of the Sargasso Sea via searches for Sec/cysteine (Cys) pairs in homologous sequences [25]. This method performed well and further research has shown that it is reliable in identifying selenoproteins in both organism-specific and environmental genomes [24],[26],[31].
In this study, we utilized a similar approach to analyze the distribution and composition of marine selenoproteins in the GOS shotgun dataset. More than 3,600 selenoprotein genes were detected, which is ten times the number of selenoproteins in the Sargasso Sea study. Several novel prokaryotic selenoprotein families were predicted. We further analyzed the dataset in various ways deriving insights into global trends in Se utilization.
Computational analysis of 44 sequenced GOS samples identified 3,506 selenoprotein sequences that belonged to previously described selenoprotein families (Table 1, all sequences are available in supplemental Dataset S1). We also identified 58,225 Cys-containing homologs of these selenoproteins in the GOS sequences. Canonical correlation analysis of their occurrence based on sample size (i.e., total number of sequenced reads for each sample) showed a strong correlation between the number of Cys-containing homologs and sample size (correlation coefficient, CC, is 0.98), but selenoproteins showed a weak correlation (CC is 0.59), suggesting widely different utilization of Sec in GOS samples (Figure 1). The samples were then clustered in various ways based on geographic location, water temperature (tropical or temperate), and salinity (sea water, fresh water, estuaries, or hypersaline lake). GS00c (Sargasso Sea Station 3, 425 selenoproteins, 12.1% of all detected selenoproteins), GS31 (coastal upwelling near Galapagos Islands, 269 selenoproteins, 7.7%) and GS17 (Yucatan Channel in Caribbean Sea, 257 selenoproteins, 7.3%) had the highest numbers of selenoproteins (Figure 2A). Normalized occurrence of selenoproteins is shown in Figure 2B (on average, GOS samples had 0.047% reads containing selenoprotein genes). We designated samples as selenoprotein-rich (6 samples) if they contained 1.5 times the average level and selenoprotein-poor (11 samples) if they had twice less the average level of selenoproteins (Figure 2B). Geographically selenoprotein-rich and -poor samples did not cluster with each other, arguing against significant geographic differences in Sec utilization within the areas examined by the GOS project (Figure 3). It should be noted that except for the Sargasso Sea samples (GS00a–GS01c), all other samples were collected in daytime between August 2003 and May 2004, but most of them were collected in a narrow time period (November 2003∼March 2004, see Table 1 for sample date and time). Therefore, seasonal and yearly shifts in microbial community were considered to be small. However, it would be of interest to examine the contribution of seasonal factors to changes in the detected microbial selenoproteome once sufficient sampling becomes available.
It has been reported that GOS samples grouped based on sequence similarity and taxonomy correlate with environmental parameters of GOS sites, particularly with regard to water temperature and salinity [29]. We found that except for sample GS09, all selenoprotein-rich samples belonged to the marine “tropical & Sargasso” group which had an average sampling temperature at 25.5°C. Also, all samples from the Gulf of Mexico and Caribbean Sea (GS15–GS19) showed elevated levels of selenoproteins (Figure 2B), suggesting an active utilization of Sec in this area. In contrast, more than half of selenoprotein-poor samples (6 out of 11) were derived from temperate water area (12.1°C in average). This observation is consistent with our previous hypothesis that the use of Sec increases at higher temperature [32]. Besides, 5 of 7 nonmarine aquatic samples were selenoprotein-poor and the remaining two were borderline selenoprotein-poor (Figure 2B). These nonmarine samples were geographically distant (Figure 3) and located in different temperature zones. Further analysis of these samples with regard to habitat and environment suggested one likely factor, salinity, which was different between marine (including both open ocean and coastal areas) and nonmarine environments. Except for GS33 which was sampled from a hypersaline lagoon (salinity is 63.4 ppt) and showed low species richness [29],[33], all nonmarine aquatic samples were characterized by very low salinity (<4 ppt) [29]. This observation suggested that fresh water or low-salinity aquatic environments may work against Sec utilization. Although more extensive sample classification was difficult because of their number, water depth, fresh water input, proximity to land and filter size all appeared to affect Sec abundance to some extent. For example, the filter for most samples was 0.1∼0.8 µm, which concentrated mostly bacterial and archaeal microbial populations [29]. However, among the Sargasso Sea samples, GS01a, GS01b and GS01c were three subsamples from the same site, representing three distinct size fractions (3.0–20, 0.8–3.0, and 0.1–0.8 µm, respectively). This feature explains the fact that GS01a was relatively poor in selenoproteins even though it was located in the area rich in selenoproteins. Similarly, GS25, another selenoprotein-poor sample, was collected using a larger filter (0.8–3.0 µm). No conclusion could be made regarding the relationship between nutrients (such as carbon, nitrogen and phosphorus) and Sec utilization. For example, in the nutrient-limited Sargasso Sea, both selenoprotein-rich (GS00c) and selenoprotein-poor (GS00a) samples were found. Similar observations were observed for coastal waters and estuaries where nutrients are more abundant, and for the open ocean where nutrients are limited. Additional factors, such as organism density, ecosystem complexity, light for phototrophs and fixed carbon/energy for chemotrophs may ultimately affect Sec utilization in microbial communities and warrant further studies once additional sequences become available.
Selenoproteins detected through the homology-based procedure (see details in Materials and Methods) belonged to 51 previously described selenoprotein families (Table 2, details are shown in Table S1). Most of these families had much more Cys-containing homologs than selenoproteins in the GOS dataset. All selenoprotein families previously detected in the Sargasso Sea were identified in the current GOS dataset, including prominent selenoproteins: SelW-like, selenophosphate synthetase (SelD), proline reductase PrdB subunit, peroxiredoxin (Prx), thioredoxin (Trx), glutaredoxin (Grx) and a variety of Prx-like/Trx-like/Grx-like proteins [25]. Other selenoproteins included a UGSC-containing protein (one of the major selenoprotein families in GOS samples, U is a one letter designation for Sec) and several selenoproteins identified in various metagenomic sequencing projects [26],[31]. In addition, we identified a large number of distant homologs of Prx-like/Trx-like selenoproteins. In order to analyze them against previously identified Prx-like/Trx-like proteins, we clustered these proteins into different subfamilies based on conserved domain classification (Pfam/COG), motif features and phylogenetic analysis. Several selenoproteins were represented by single sequences only, e.g., glycine reductase selenoprotein A (GrdA) and heterodisulfide reductase subunit A (HdrA). In this case, sequencing errors that generated in-frame TGA codons could not be excluded; however, the fact that they corresponded to known selenoproteins and also possessed strong SECIS elements strongly suggested that they were true selenoproteins. 20 selenoprotein families were represented by more than 40 selenoprotein sequences and accounted for more than 94% of all selenoprotein sequences. Similar to the selenoproteome of the Sargasso Sea, the most abundant selenoprotein families were SelW-like, SelD, UGSC-containing protein, Prx, PrdB, and different subfamilies of Prx-like/Trx-like/Grx-like proteins. The current version of GOS selenoproteome has become the largest selenoproteome identified to date, and its analysis greatly expands our understanding of Sec utilization in microbial marine communities.
Most selenoproteins with known function are oxidoreductases, and among 51 selenoprotein families detected in GOS samples, 33 (2887 sequences, 82.3%) were homologs of known thiol oxidoreductases or possessed Trx-like fold (Table 2). Many of these selenoproteins contained a conserved UxxC/UxxS/CxxU/TxxU redox motif. In a small number of known selenoprotein genes, new Sec positions were identified. For example, a new redox motif (CxxU) was detected in Trx-like 1 family (COG0526, TrxA, thiol-disulfide isomerase and thioredoxins) which normally contains a UxxC motif (i.e., in all previously identified sequences) (Figure 4A). In addition, several UxxU-containing sequences were detected in a Prx-like 2 family (low similarity to pfam04592, Selenoprotein P N-terminal region), which is a very distant homolog of known Prxs and has no strong homolog in any of the sequenced prokaryotic genomes (Figure 4B).
To further investigate the relationship between occurrence of selenoprotein families and sample features (e.g., marine versus nonmarine), we analyzed the most abundant selenoprotein families in each GOS sample separately (Table 3). Excluding the samples containing a small number of selenoproteins (≤15), the majority selenoprotein families showed a similar occurrence in marine and nonmarine aquatic samples. In contrast, several selenoprotein families appeared to be differentially distributed. For example, SelW-like protein was generally the most abundant selenoprotein family in marine samples, whereas the UGSC-containing protein was most frequently utilized in nonmarine samples. As discussed above, salinity appears to be a factor that influences (perhaps indirectly) selenoprotein utilization. Figure 5 shows the occurrence of 20 most abundant selenoprotein families based on habitat. T-test was used to assess occurrence of each of these families in marine and nonmarine habitats. This analysis showed that occurrence of selenoprotein families in group I (selenoproteins with lower occurrence in nonmarine samples, Figure 5) and II (selenoproteins with lower occurrence in marine samples) were statistically different between marine and nonmarine samples (p<0.01).
Besides known selenoproteins, we identified 7 new selenoprotein families (Table 4, all sequences are available in supplemental Dataset S2). They were represented by 2–11 individual TGA-containing sequences except for a hypothetical protein GOS_C which had 74 selenoprotein sequences. Among 7 new families, four either contained a domain of known function or were homologous to protein families with known/predicted functions. Particularly interesting was identification of ferredoxin-thioredoxin reductase (FTR) catalytic subunit and trypsin-like serine protease homologs. FTR is a key enzyme of the ferredoxin/thioredoxin system, which catalyzes reduction of thioredoxins with light-generated electrons [34]–[36]. Two Cys residues constitute a redox-active disulfide bridge functional in the reduction of Trx [37]. We identified two FTR selenoprotein sequences, including one (JCVI_READ_1093012271142) which contained two predicted Sec residues exactly corresponding to the two redox-active Cys residues (Figure 6A). Location of these Secs indicates functionality of these residues.
Trypsin is a well-known serine protease which catalyzes the hydrolysis of peptide bonds. No redox function has been reported for members of this family. We found 9 selenoprotein sequences containing the trypsin-like domain (COG5640, secreted trypsin-like serine protease) and the predicted Sec corresponded to a conserved Cys residue within this domain, suggesting a potential redox function for this Cys (Figure 6B). No functional evidence could be obtained for hypothetical proteins GOS_A∼GOS_C. However, a Trx-like fold and a conserved UxxC motif were present in GOS_C, suggesting that this protein may have a thiol-based redox function. Multiple alignments of these new selenoproteins and their Cys-containing homologs (Figure 6A–6G) highlight sequence conservation of Sec/Cys pairs and their flanking regions. New selenoproteins contained stable bacterial SECIS-like elements downstream of Sec-encoding TGA codons (Figure 7). In addition, we detected several TGA-containing sequences, which showed similarity neither to known and new selenoproteins nor to each other. Some of them contained candidate SECIS elements. However, no definitive conclusions could be made regarding these sequences because of the possibility of sequencing errors. Future experimental verification is needed for these selenoprotein candidates.
Previous analyses revealed that several selenoprotein families occur in both prokaryotes and eukaryotes, e.g., SelW-like, GPx and deiodinase [25]. Recently, additional such selenoprotein families were identified, e.g., methionine sulfoxide reductase A (MsrA), Prx, SelL (a Prx-like protein), arsenite S-adenosylmethyltransferase (PRK11873, arsM) and several Prx-like/Trx-like proteins [31], [38]–[40]. Most eukaryotic species containing these selenoproteins are aquatic organisms (such as green algae and fish). In the GOS sequence dataset, more than 90% sequences are derived from bacteria whereas only 2.8% could be definitively assigned to the eukaryotic domain [27]. To distinguish bacterial and eukaryotic selenoproteins, we employed several approaches including phylogenetic analyses and investigation of eukaryotic SECIS elements. Our results suggested that all detected new and known selenoproteins that occur in both prokaryotes and eukaryotes could be assigned to the bacterial domain.
In addition, several eukaryotic selenoproteins were detected in different GOS samples by homology analysis using known eukaryotic selenoproteins, including protein disulfide isomerase (PDI), SelM, SelT, SelU and thioredoxin reductase (data not shown). Although most of the reads containing these selenoprotein genes were too short to investigate the presence of eukaryotic SECIS element in 3′-UTR, phylogenetic analyses and the absence of bacterial SECIS elements suggested that these sequences are eukaryotic.
Interestingly, a new eukaryotic selenoprotein family, gamma-interferon-inducible lysosomal thiol reductase (GILT), was also detected. GILT is a key enzyme to facilitate complete unfolding of proteins destined for lysosomal degradation by releasing structural constraints imposed by intra- and inter-chain disulfide bonds [41],[42]. No homologs of this protein are known in prokaryotes. In this study, we identified three selenoprotein sequences for this family. A eukaryotic SECIS element predicted by SECISearch [18] was found in the 3′-UTR of one selenoprotein gene, providing additional evidence that they are eukaryotic GILT selenoproteins. Multiple alignment of GILT sequences and the predicted eukaryotic SECIS element are shown in Figure 8.
We identified novel domain fusions in several selenoprotein families. One example involved Prx that was fused with a distant homolog of PP2C-type phosphatase (smart00331, PP2C_SIG, Figure 9A). The PP2C-type phosphatase superfamily includes several subgroups, such as RsbU that contains an additional N-terminal domain (pfam08673, RsbU_N) and acts as a positive regulator of the activity of σB, the general stress-response σ factor of gram positive microorganisms [43],[44]. Other PP2C-type phosphatase subfamilies include PrpC, SpoIIE, RsbP and RsbX [45]–[49], in which the PP2C-type phosphatase domains are fused with different domains (Figure 9B). We further checked the occurrence of this distant PP2C-type phosphatase in all sequenced bacteria and found orthologs only in a limited number (no more than 20) of organisms in different bacterial phyla and fused with different domains (Figure 9C). Phylogenetic analyses suggested that the Prx-fused phosphatases form a separate group within the PP2C-type phosphatase superfamily (Figure 10). Multiple alignments showed that several conserved residues are specific for this subgroup, especially a Cys residue which is present in all members of the Prx-fused subgroup but absent in other PP2C-type phosphatase subfamilies (Figure 11). This conserved Cys may also have a redox function. Surprisingly, one marine gliding bacterium, Microscilla marina, the only organism containing the Prx-fused phosphatase domain in Bacteroidetes, possessed a large number of such proteins. Compared to other organisms which contained only 1–2 members, 159 individual sequences containing this phosphatase subfamily were identified in M. marina, all of which had the conserved Cys residue and were fused with different domains, suggesting a particular importance of this distant PP2C-type phosphatase subfamily in this marine organism.
Additional examples of domain fusions are shown in Figure S1. Functions of most of these domains are not clear. However, as a rule, at least one conserved Cys was present in these sequences, suggesting a potential redox activity. For example, the UGSC-containing protein which likely has a Trx-like fold was fused with a conserved domain (designated Unknown_1, Figure S1A). Unknown_1 protein was also present in a limited number of aquatic organisms. Another example involved the fusion of a Prx-like 3 and Unknown_3 domain (Figure S1D). There were three conserved Cys residues in Unknown_3, including a conserved CxxC motif which may have a thiol-based redox function.
Previously, we detected two fusions of SelD: (i) NADH dehydrogenase (COG1252, Ndh, FAD-containing subunit) fusion [32] and (ii) Cys sulfinate desulfinase (COG1104, NifS) fusion (unpublished data). The Ndh-SelD fusion proteins were detected in several bacteria most of which were aquatic organisms. Such fusions were also observed in several lower eukaryotes, such as in Ostreococcus. In all detected fusion sequences, a conserved CxxC motif was present in the predicted active site of the SelD domain. However, this motif is very rare (<5%) in single-domain SelD proteins. The NifS-SelD fusion was only detected in Geobacter sp. FRC-32 (an anaerobic, iron- and uranium-reducing deltaproteobacterium), and a CxxU motif was present in the active site of the SelD domain. Functions of the two fusion SelDs are not fully clear, but are expected to be involved in selenophosphate synthesis. In the GOS dataset, we detected hundreds of Ndh-SelD fusion proteins (all containing the CxxC motif), which accounted for approximately 40% of all detected Cys-containing SelDs. In contrast, no NifS-SelD fusion was detected. Interestingly, we found that ∼5.6% of single-domain selenoprotein SelDs contained a CxxU motif. Figure 12 shows a multiple alignment of Ndh-SelD fusion proteins and other Sec/Cys-containing SelDs in both sequenced organisms and GOS samples.
We also found several sequence reads containing two neighboring selenoprotein genes, including ten Prx/SelW sequences, one Prx/Prx-like 2 and one Prx-like 1/AhpD-like 2 sequences. Phylogenetic analysis showed that these Prx and SelW sequences were clustered in a small phylogenetic group, suggesting that they come from closely related organisms. Further analyses are needed to examine a possible functional link between these selenoproteins.
In some prokaryotes, Se (in the form of selenophosphate) is also used for biosynthesis of a modified tRNA nucleoside, 5-methylaminomethyl-2-selenouridine (mnm5Se2U), which is located in the wobble position of the anticodons of tRNALys, tRNAGlu, and tRNA1Gln [50]–[52]. The proposed function of mnm5Se2U involves codon-anticodon interactions that help base pair discrimination at the wobble position and/or translation efficiency [52],[53]. A 2-selenouridine synthase (YbbB) is necessary to replace a sulfur atom in 2-thiouridine in these tRNAs with selenium [54]. Here, we investigated the occurrence of YbbB to assess the selenouridine utilization trait in the GOS samples.
A total of 865 YbbB genes were identified in GOS sequences. Occurrence of YbbB in individual samples is shown in Figure 13A. In most GOS samples, the number of reads containing YbbB gene was proportional to the sample size (CC is 0.87). However, several samples appeared to have a significantly different distribution of YbbB. Similarly to selenoprotein classification of GOS samples, we clustered them into selenouridine-rich and selenouridine-poor. Previously, we have suggested a relatively independent relationship between Sec and selenouridine utilization [32]. In the current study, we examined correspondence between selenoprotein-rich/poor samples and selenouridine-rich/poor samples (Figure 13B). Two selenoprotein-poor samples (GS00a and GS33) were selenouridine-rich, whereas one selenoprotein-rich sample (GS51) appeared to be a selenouridine-poor sample, implying no strong relationship of the two Se utilization traits in GOS samples. Also, no significant difference was observed for the occurrence of the selenouridine utilization trait in other selenoprotein-rich/poor samples, further suggesting a relatively independent relationship between them. Considering that Se supply in the sea water should be equal to co-occurring Sec-utilizing and selenouridine-utilizing organisms, substantial microbial taxonomic diversity might explain differences in Se utilization in different areas of the sea. No clear relationship was also found between selenouridine utilization and habitat types or geographic location. Except for GS01a (a sample collected with a large filter), GS12 (from the estuary close to Chesapeake Bay, MD) was the only sample in which both Se utilization traits were limited. We also found high utilization of both traits in GS17 (Caribbean Sea, Yucatan Channel).
In recent years, a number of metagenomic sequencing projects were carried out that enabled researchers to identify genes in both abundant and non-abundant microbes in a particular environment, providing a powerful tool to examine community organization and metabolism in natural microbial communities [30], [55]–[57]. Similarly, identification of selenoprotein genes in these datasets may help in understanding the role of Se in microbial populations. In this study, we have used shotgun data from a recent GOS expedition [27]–[29] to characterize the distribution and composition of the selenoproteome in this largest to date marine metagenomic dataset. Our results highlight importance of Se utilization within marine microbial communities and provide a comprehensive analysis of Se-dependent proteins which are utilized by marine microorganisms.
The GOS project produced a total of 7.7 million random sequence reads from the North Atlantic Ocean, the Panama Canal, and East and central Pacific Ocean gyre. In order to identify all selenoproteins in the GOS dataset we employed a procedure that analyzed Sec/Cys pairs in homologous sequences. A total of 3,506 sequences which belonged to 51 previously described prokaryotic selenoprotein families, and 102 sequences that corresponded to 7 new selenoprotein families were identified. Compared to smaller scale non-aquatic metagenomic projects, such as whale fall community and Waseca County farm soil metagenome [56] and human distal gut microbiome [57], the GOS project produced hundreds of times more selenoproteins. Our current study generated by far the largest selenoproteome reported to date.
If selenoproteins and their Cys-containing homologs are randomly used in marine microbes, the number of selenoproteins would be expected to be proportional to the number of sequence reads in GOS samples. However, whereas the correlation was good for Cys homologs, it was weak for selenoproteins. We normalized the occurrence of selenoproteins in each sample and found that all selenoprotein-rich samples originated from the sea water and almost all from the tropical sea areas. In contrast, half of the selenoprotein-poor samples were obtained from nonmarine aquatic environments (including fresh and hypersaline water), and half of the marine selenoprotein-poor samples came from temperate sea areas. Thus, our data suggest that water salinity and temperature may influence Sec utilization. However, the fact that the occurrence of selenoproteins in some samples collected from sites with similar temperature and salinity was somewhat different suggests that additional factors may also affect Sec utilization. Moreover, other features of GOS samples (e.g., water depth, fraction filter and light intensity) may also result in bias when comparing the samples.
Among 51 previously characterized selenoprotein families, most were homologs of known thiol oxidoreductases or possessed Trx-like fold, consistent with the idea of redox function for selenoproteins in marine microorganisms. Twenty selenoprotein families, including SelW-like, SelD, Trx-like 1 and UGSC-containing proteins, were found to be the major selenoprotein families in GOS samples and represented approximately 95% of all detected selenoprotein sequences. Except for SelD, FdhA and UshA-like (COG0737, UshA, 5′-nucleotidase/2′,3′-cyclic phosphodiesterase and related esterases), all of these families contained conserved Cys-based redox motifs which are involved in a variety of redox functions. Comparison of the distributions of these major selenoprotein families in marine and nonmarine environments showed that a small number of selenoproteins exhibited significantly different occurrence in the two types of habitat. For example, SelW-like, DsbA 1, Prx-like 2, Prx-like 3 and Trx-like 3 were much more abundant in marine samples whereas UGSC-containing, AhpD-like 2 and Prx-like (UGC-containing) proteins were more abundant in nonmarine samples. Therefore, salinity and other factors affected the use of Sec, but this influence is not necessarily unidirectional and depends on specific selenoproteins affected.
Seven new selenoprotein families were identified. Except for hypothetical protein GOS_C, which was represented by 74 selenoprotein sequences in the GOS dataset, occurrence of other new selenoprotein families was limited. Among these new families, FTR is a well-characterized enzyme involved in disulfide reduction in Trx. However, previous studies could not detect any Sec-containing form for this enzyme. In addition, several Sec-containing sequences were predicted for a trypsin-like family, suggesting a potential redox function for a particular Cys residue in this well-known serine protease family. Although functions of other new families are unclear, the fact that a CxxU motif was present in both FmdB putative regulatory protein family and putative secreted serine protease MucD, and that a UxxC motif was present in a hypothetical protein GOS_C, implied a thiol-related redox function.
It has been reported that a small fraction (less than 3%) of reads in the GOS dataset is of eukaryotic origin (e.g., small-sized green algae). We did detect several eukaryotic selenoproteins, including a new selenoprotein family, GILT. Homologs of this protein family were only detected in eukaryotes. A eukaryotic SECIS element was detected in the 3′-UTR in one selenoprotein sequence. Although eukaryotic organisms containing the Sec-containing GILT are not known, future studies will likely identify such organisms.
Domain fusions could help identify functionally-related proteins. We identified several new fusion events involving selenoproteins. Compared to their more common forms present in most organisms, these selenoproteins contained additional upstream or downstream domains fused into a single protein chain. Fusion events were observed for a variety of Trx-fold-containing selenoproteins, including Prx, Prx-like 2, Prx-like 3 and UGSC-containing protein. Function of most of these fused domains is not clear; however, single or multiple conserved Cys residues were present in these domains, suggesting a potential redox function of these residues. In addition, almost half of the Cys-containing SelDs detected in the current GOS dataset were Ndh-SelD fusion proteins, all of which contained a conserved CxxC motif in the active sites. The abundance of Ndh-SelD fusion proteins in GOS samples suggests that this fusion SelD plays an important role in selenophosphate biosynthesis in marine/aquatic organisms.
Given that Se is also utilized for biosynthesis of selenouridine in bacteria, distribution of the selenouridine trait was assessed by analyzing occurrence of YbbB in GOS samples. We identified selenouridine-rich and selenouridine-poor samples, which were not the same as Sec-rich/poor samples, suggesting that the two Se utilization traits are functionally independent (but of course both depend on supply of Se). This observation is consistent with the previous hypothesis that Sec and selenouridine utilization traits are relatively independent even though both traits require SelD for selenophosphate biosynthesis [32]. In addition, no strong relationship was found between selenouridine utilization and habitat types (marine or nonmarine) or geographic location. Although both Se traits require Se supply or thus could influence evolution of each other, additional factors appear to play more important roles in the evolution and utilization of individual Se utilization traits.
In this study, we report a comprehensive analysis of Sec utilization in marine microbial samples of the GOS expedition by characterizing the GOS selenoproteome. This is the first time that the microbial selenoprotein population is described in a global biogeographical context. Our analysis yielded the largest selenoprotein dataset to date, provided a variety of new insights into Sec utilization and revealed environmental factors that influence Sec utilization in the marine microbial world.
Shotgun reads of the GOS project were downloaded from the CAMERA (Community Cyberinfrastructure for Advanced Marine Microbial Ecology Research and Analysis) website at http://camera.calit2.net. This dataset contains a total of 7,709,422 genomic sequences derived from 57 samples (13 samples are not fully sequenced), which cover a wide range of distinct surface marine environments as well as a few nonmarine aquatic samples [29]. The genomic sequences combined had 8.148 billion nucleotides. In addition, we downloaded the non-redundant (NR) protein database from the NCBI ftp server. It contained a total of 4,644,764 protein sequences. BLAST [58] was also obtained from the NCBI.
Previously, we developed and employed a set of programs for automated selenoprotein searches [24]–[26]. However, since this approach is based on an exhaustive search of all possible Cys/Sec pairs for each Cys-containing sequence in the NR database, the computation procedure can become very expensive when the target sequence dataset is very large, as is the case in the GOS database. Therefore, we utilized an Open Science Grid (OSG) management system which is dedicated to supporting scientific research through the use of advanced networking technology and high performance computing [59]. We employed Condor-G software [60], a powerful and full-featured task broker, to manage such a high throughput computing project on large collections of distributively owned computing resources. In addition, we used the Prairiefire, a 128-node, 256-processor Beowulf cluster supercomputer at the Research Computing Facility of the University of Nebraska – Lincoln.
We used a strategy which we had successfully used in selenoprotein searches in other metagenomic datasets: Sargasso Sea and symbiotic microbial consortium of the marine oligochaete Olavius algarvensis [24]–[26]. Briefly, each Cys-containing sequence in the NR protein database was searched against the GOS dataset for top 1000 homologs using TBLASTN with E-value below 10 (this step is the most time-consuming and was performed completely on the OSG system). Cys/TGA pairs were then selected and a minimum open reading frame (ORF) for each TGA-containing nucleotide sequence (TGA was translated to Sec, U) was obtained. After that, BLASTX and RPS-BLAST programs were used to analyze the conservation of TGA-flanking regions in all six reading frames as well as the presence of conserved domains. Remaining sequences were clustered based on sequence similarity and location of predicted Sec using BL2SEQ with an E-value below 10−4. All clusters were further searched against NCBI NR protein and microbial genomic databases to identify conserved Cys-containing homologs. Sequences in the remaining clusters were manually analyzed for occurrence of bacterial SECIS elements using bSECISearch program [21], and were classified into known selenoproteins and candidate selenoproteins (i.e., clusters having at least two Sec-containing sequences). In addition, an independent BLAST homology search for selected Sec-containing representatives of all previously identified prokaryotic selenoprotein families was performed. Finally, distinct representatives of all identified selenoprotein sequences were used to iteratively search against the GOS dataset for identification of additional distant Sec-containing homologs.
We used CLUSTALW [61] with default parameters for multiple sequence alignments. Phylogeny was analyzed by PHYLIP programs [62]. Neighbor-joining (NJ) trees were obtained with NEIGHBOR and the most parsimonious trees were determined with PROTPARS. Robustness of these phylogenies was evaluated by two additional algorithms: maximum likelihood (ML) analysis with PHYML [63] and Bayesian estimation of phylogeny with MrBayes [64]. |
10.1371/journal.pntd.0000994 | Evaluation of the Effectiveness of Insecticide Treated Materials for Household Level Dengue Vector Control | To assess the operational effectiveness of long-lasting insecticide treated materials (ITMs), when used at household level, for the control of Aedes aegypti in moderately infested urban and suburban areas.
In an intervention study, ITMs consisting of curtains and water jar-covers (made from PermaNet) were distributed under routine field conditions in 10 clusters (5 urban and 5 suburban), with over 4000 houses, in Trujillo, Venezuela. Impact of the interventions were determined by comparing pre-and post-intervention measures of the Breteau index (BI, number of positive containers/100 houses) and pupae per person index (PPI), and by comparison with indices from untreated areas of the same municipalities. The effect of ITM coverage was modeled.
At distribution, the proportion of households with ≥1 ITM curtain was 79.7% in urban and 75.2% in suburban clusters, but decreased to 32.3% and 39.0%, respectively, after 18 months. The corresponding figures for the proportion of jars using ITM covers were 34.0% and 50.8% at distribution and 17.0% and 21.0% after 18 months, respectively. Prior to intervention, the BI was 8.5 in urban clusters and 42.4 in suburban clusters, and the PPI was 0.2 and 0.9, respectively. In both urban and suburban clusters, the BI showed a sustained 55% decrease, while no discernable pattern was observed at the municipal level. After controlling for confounding factors, the percentage ITM curtain coverage, but not ITM jar-cover coverage, was significantly associated with both entomological indices (Incidence Rate Ratio = 0.98; 95%CI 0.97–0.99). The IRR implied that ITM curtain coverage of at least 50% was necessary to reduce A. aegypti infestation levels by 50%.
Deployment of insecticide treated window curtains in households can result in significant reductions in A. aegypti levels when dengue vector infestations are moderate, but the magnitude of the effect depends on the coverage attained, which itself can decline rapidly over time.
| An estimated 40% of the world's population lives at risk of contracting dengue, and it inflicts a significant health, economic and social burden on the populations of endemic areas. In the absence of a vaccine, vector control is the only available strategy to prevent transmission. Some control methods against Aedes aegypti (the main dengue vector) have been successful in reducing vector infestation levels, but rarely sustained the reductions for a prolonged period. We report here on the first effectiveness trial of insecticide treated curtains and jar covers against A. aegypti implemented under ‘real-life’ conditions. The coverage of tools was high at distribution, but declined quickly over the 18 months of follow up. The vector infestation levels showed a sustained 55% decrease in the intervention clusters, while no discernable pattern was observed at the municipal level. At least 50% curtain coverage was needed to reduce A. aegypti infestation levels by 50%. We concluded that deployment of insecticide treated window curtains in households can result in significant reductions in dengue vector levels, which are related to dengue transmission risk. The magnitude of the effect depends on the curtain coverage attained, which itself can decline rapidly over time.
| An estimated forty percent of the world's population lives at risk of contracting dengue, which currently is the most important mosquito-borne viral disease worldwide, responsible for 24,000 deaths, 250,000–500,000 hemorrhagic fever cases and up to 50 million dengue infections annually [1], [2]. The public health importance of dengue has grown rapidly in recent years, with a 30-fold increase in incidence since the 1960s. This has coincided with the expansion of the geographical range of its main vector, the mosquito Aedes aegypti [2], [3], and co-circulation of multiple dengue serotypes, which elevates the risk of sequential infections and severe disease [4].
No curative treatment is available and the prevention of a fatal outcome in severe dengue cases hinges on early case detection and appropriate supportive treatment. To decrease the burden of disease, prevention of transmission is crucial. As there is no vaccine yet, this is possible only by vector control. Existing A. aegypti control tools can reduce vector infestation levels, but very few have succeeded in sustaining reductions for a prolonged period [5], [6] or in impacting on dengue transmission [7], [8]. The national routine dengue vector control programmes in endemic countries are facing variable and often disappointing results, which are among others due to inadequate implementation processes, lack of community participation or poor user acceptance of chemical-based vector control methods [9], [10]. Programs integrating chemical or biological based strategies with community involvement are having better results, but rarely eliminate the vector [11]–[14], though there have been notable successes in recent years [8].
Insecticide treated materials (ITMs) have recently shown promise in reducing household level dengue vector infestations [15]–[17]. Unlike most dengue vector control strategies, ITMs target the adult mosquito, which is epidemiologically the most important vector stage. It is postulated that the likelihood of adult vectors contacting an ITM during host seeking reduces their life expectancy, effectively altering the age structure of the vector population, such that fewer mosquitoes live long enough to become infective with dengue [9]. Furthermore, ITMs made from long-lasting insecticide treated fabrics retain their efficacy for at least 1 year [18], which is longer than any other applied Aedes control tool. In previous trials, ITMs were shown to have an impact on vector populations and to have high acceptance levels by householders up to a few months after distribution [15], [17], though the key question of whether or not this will result in reduced dengue transmission remains to be proven before ITMs can be recommended as dengue vector control tools on a large scale. Achieving and sustaining high levels of ITM uptake and use under routine programme conditions rather than in an experimental situation also are fundamental prerequisites to success and this too requires investigation.
We report here on an intervention study in urban and suburban areas of Trujillo State, Venezuela, where insecticide treated window curtains and water jar covers were distributed by local health committees and by the existing routine vector control programme. Over a period of 18 months, we assessed the uptake and use of these tools by local householders and their effectiveness in controlling the vector population, comparing the Breteau and the pupae per person indices at several time points before and after intervention; and also comparing them with indices from untreated areas of the same municipalities.
This study received clearance from the ethical committee that oversees research of the Institute of Tropical Medicine, Antwerp and from the bio-ethics committee of the Jose Witremundo Torrealba Research Institute, Trujillo. Community representatives from each participating cluster approved the intervention and written informed consent was obtained from each individual household included in the study. The ITMs were made from material that is approved by the World Health Organization Pesticide Evaluation Scheme (WHOPES) for bed net use. The trial was registered at ClinicalTrials.gov (number NCT 00883441).
The study was conducted in the municipalities of Valera (9°19′N 70°36′W; altitude 541 m) and San Rafael de Carvajal (9°20′N 70°35′W; altitude 556 m; referred to as Carvajal) in Trujillo State in north west Venezuela. The climate is tropical with two rainy seasons (March/April and September/November), an average annual rainfall of 750 mm and temperatures ranging from 16–37°C. The city of Valera is the economic capital of the state, with 128,556 inhabitants and a population density of 534 inhabitants/km2. Carvajal is a suburban municipality located 4 km from Valera, with 44,213 inhabitants and a population density of 493 inhabitants/km2.
Dengue is endemic in Trujillo State. Between 2006 and 2008, dengue case reports ranged between 203 and 396 cases/100,000 inhabitants/year, of which 1.1 to 4.9% were hemorrhagic cases (Regional direction of epidemiology and statistics, Trujillo state health ministry). In Trujillo, dengue affects children and young adults (up to 24 years old) primarily, with cases peaking between August and December.
All routine A. aegypti vector control activities are carried out by a team of 24 persons from the department of environmental health of the Trujillo state health ministry. Activities comprise adulticiding (indoor spraying with malathion 94% ULV) and larviciding (Abate) within a 200 meter radius of a reported dengue case. When the number of clinical cases exceeds the epidemic alert level of the endemic channel, space spraying with vehicle-mounted equipment is added (Department of environmental health of the Trujillo state Health Ministry).
The study had been designed as a cluster randomized trial for comparing the efficiency of 2 ITM distribution models in terms of uptake and continued use that would at the same time permit a non-randomized before - after comparison to evaluate the effectiveness of the ITMs. Uptake rates attained at distribution and the subsequent reductions in coverage over time were equivalent in both models and are not analyzed in depth in this manuscript.
In March 2006 10 clusters (defined as distinct neighbourhoods of 300–600 houses) were recruited for an intervention study and were stratified based on location in urban or suburban areas, which differ in Aedes infestation levels and population characteristics. The 10 clusters were selected from 18 districts that had dengue notification rates of at least 40/10,000 inhabitants (2003–2005). The inclusion criteria at cluster level were: middle or low socio-economic status (the number of high socio-economic level clusters was small and they were not representative of the overall area) with fewer than 50% of the population residing in apartment blocks (for operational reasons). Rural areas, where principal land use was for agricultural activities and where dengue was not a major health problem, were excluded.
The sample size (number of clusters) was determined using calculations proposed by Hayes and Bennett [19], and had a power of 80% to detect a 5-fold decrease in the Breteau index at an alpha error level of 0.05 (assuming a between-cluster coefficient of variation of 0.50).
After collecting baseline data for 1 year, insecticide treated (IT) curtains and IT jar covers were distributed to all households in the 10 intervention clusters that had given their informed consent and agreed to use them. This was done between July and September 2007 either by the routine vector control programme or by local health committees. The ITcurtains and ITcovers were made from the same PermaNet (Vestergaard-Frandsen) polyester netting treated with a long-lasting formulation of deltamethrin (55 mg/m2), coated with an unknown protectant (not disclosed by the manufacturer) to prevent degradation of the insecticide when exposed to UV light. The manufacturer stated that this material does not require re-treatment and its insecticidal effect is expected to last for up to 2 years or 6 “standard” washes (http://www.vestergaard-frandsen.com/permanet-curtain-e-brochure.pdf, accessed 22/05/2008). The number of ITcurtains and ITcovers distributed per house depended on the number of windows in the main living area and bedrooms (up to a maximum of 5 curtains/house) and on the numbers of 150–200 liter water storage jars present in the house (no maximum of covers per house).
During distribution, at least one person in every household received information on the use and maintenance of the ITMs through person-to-person communication.
A total of 1120 houses were selected through systematic random sampling (560 houses across all urban clusters, corresponding to every 3rd house; and 560 across all suburban clusters, corresponding to every 4th house) for periodic entomological monitoring. Five independent entomological surveys were conducted by a survey team (trained and supervised by an experienced entomologist, author MO) at roughly six-month intervals: two pre-intervention surveys (October 2006, March/April 2007) and three post-distribution surveys (November/December 2007, April 2008, January 2009). In all houses, containers were inspected for the presence of larvae and pupae. If pupae were found they were counted, collected, transported to the laboratory and allowed to emerge for species identification.
As external control data, in line with Kroeger et al. [15], we used the entomological data collected in the municipalities of Valera and Carvajal as part of the routine surveillance activities. The intervened clusters represented 7% of the total number of houses in the municipalities and the populations of vectors in the municipalities were considered beyond the influence of the ITMs used in the study clusters, and expected to fluctuate naturally as influenced by seasonal parameters only. The routine entomological surveillance data were collected by the department of environmental health of the Trujillo state health ministry: Houses in a radius of 200 m around a confirmed dengue case are visited and infestation of water holding containers with immature vector stages is recorded. We report the routine data for the months when entomological surveys were conducted in the intervention clusters and correct for the differences in data collection methods in the analyses (see subsection on data analysis).
Rainfall and temperature data were obtained from the weather station at the Valera airport (located in between the two study sites, that are, themselves, 4 km apart) (http://www.wunderground.com/history/station/80426). The averaged rainfall data from the month of each entomological survey plus data from the preceding month were used in our analyses. Data on routine A. aegypti control activities that took place in the intervention clusters during the month of each entomological survey and the preceding month were retrieved from the reports of the vector control programme. For each cluster, an intensity score was calculated based on the number of houses treated per cluster (adulticiding and/or larviciding): 0 = no activities, 1 = less than 10% of houses covered, 2 = between 11 and 20% of houses covered, 3 = between 21 and 50% houses covered, 4 = more than 51% of houses covered, 5 = intensive and repeated spraying and larviciding in all houses.
A baseline socio-economic survey was conducted during June–July 2006 in a systematic random sampling of 955 households (465 urban and 490 suburban). The survey encompassed both general and dengue related household characteristics. We used the Graffar Method, adapted to the Venezuelan context by Méndez-Castellano [20] to classify the households according to socio-economic stratum based on the profession of the head of household, education level of the mother, main source of family income and housing conditions. It classifies households from stratum I (upper class) to stratum V (critical poverty). A random sub-sample of households that participated in the baseline sociological survey was revisited in September 2007, February 2008 and January 2009 to observe the presence and use of the ITMs. On the same occasions, we inquired about any adverse effects attributed to the use of the ITMs.
We developed 2 indicators for assessing ITcurtain coverage per cluster: the percentage of houses with at least 1 curtain and the median number of curtains per house. For ITcovers, the 2 indicators were the percentage of houses using at least 1 ITcover and the percentage of eligible water storage jars covered. We used the chi square test and the Mann-Whitney test to compare urban and suburban coverage proportions and medians respectively, and calculated 95% confidence intervals for their differences.
A. aegypti infestation levels were the outcome measures. We calculated the Breteau index (BI, number of containers positive with immature A. aegypti/100 inspected houses) per cluster, per setting (urban/suburban) and per survey round. We compared trends over time for the BI in the urban and suburban study clusters. The trend of the BI in the corresponding municipalities was used to control for the natural seasonal fluctuations in vector populations. We calculated the % difference between the values at each survey time point and the October 2006 pre-intervention values. This permitted to represent the trends and to allow, at the same time, for the different methodology used to measure BI in intervention clusters and the municipality. For the intervention clusters, the pupae per person index (PPI, number of A. aegypti pupae/inhabitant) - considered a more accurate proxy for adult mosquito abundance [21] - was also calculated per cluster, per setting and per survey round. 95% confidence intervals around each estimate at each time point were calculated with a negative binomial regression model taking into account the cluster design.
To estimate the independent effect of ITM coverage on A. aegypti infestation at the cluster level, we constructed two generalized linear random effect regression models with a negative binomial link function, taking into account the repeated measurements. Both outcome measures, BI and PPI, were the dependent variables. Each of the 10 clusters contributed 1 data point at each of the 5 entomological survey rounds. The models included the % ITcurtain and ITcover coverage, the setting (urban or suburban), intensity of routine vector control activities in each cluster, municipal level data on A. aegypti infestation, rainfall and temperature. Interaction between variables was assessed. Based on the corresponding model regression-coefficient, the independent effect of ITcurtains coverage on BI was graphically represented over the empirically observed coverage range.
Data were analyzed with Stata 10.0 (StataCorp, Texas, USA) and SPSS 17.0 (SPSS Inc., Chicago, IL, USA). In less than 1% of cases some essential data were missing, and these households were subsequently withdrawn from the database.
The 5 urban and the 5 suburban clusters contained 1742 and 2359 houses, respectively. All clusters completed the study protocol through January 2009 and all were included in the analysis. Households in urban clusters had a significantly higher socio-economic status (p<0.05) (Table 1). Permanent water supply was more common in urban than suburban intervention clusters (65.8% and 44.9% respectively; p<0.001), while houses in suburban clusters had more water storage jars than urban houses (averages of 1.2 and 0.5 water storage jars/house respectively; p<0.05).
All 510 pupae collected in the October 2006 survey belonged to subgenus Stegomyia and 89% were A. aegypti. Also, in November 2007, after ITM distribution, 89% of the collected pupae were identified as A. aegypti. Since the vast majority of immature stages were A. aegypti at both time moments, all immature stages observed in entomological surveys were assumed to be of this species, in line with Arredondo-Jimenez and Valdez-Delgado [22].
Immediately after ITM distribution, in September 2007, coverage with ITcurtains was similarly high (>75%) in both settings (Table 2). ITcover coverage levels differed significantly between settings, with lower coverage in urban (12.9% of houses) compared to suburban (31.1%) clusters (p<0.001). This was not surprising since, at baseline, 73.1% of jars in urban clusters were found to be fitted with a locally purchased cover as compared to 35.9% in suburban clusters (regardless of condition or use)(p<0.05).
Minor allergic reactions (temporary [less than 48 hrs] itching of palms) after handling the ITMs were reported in 5.4% of houses.
As the trial progressed, ITM coverage declined such that by the end of the 18-month follow-up period (January 2009), fewer than 40% of houses were using ITcurtains and fewer than 20% were using ITcovers (Figure 1). There were no significant differences between the settings (urban or suburban) in the rates of decline in coverage of ITcurtains (p>0.05) or ITcovers (p>0.05). It was observed that the ITcovers deteriorated as the study progressed, particularly the elasticated rim, resulting in poorly sealed jars.
Prior to intervention, the BI was 42.4 in suburban intervention clusters, which was significantly higher than the BI of 8.5 in the urban clusters; the PPI was 0.9 and 0.2, respectively (Figure 2). Both settings experienced significant declines in the BI (IRR = 0.30, 95% CI 0.22–0.49) and PPI (IRR = 0.23, 95%CI 0.14–0.37) in the months following the distribution of the ITMs (Figure 2). In November 2007, the BI fell to 15.8 in urban and 3.8 in suburban intervention settings, and the PPI decreased to 0.2 and 0.03, respectively. While the PPI gradually increased again in the suburban (but not in the urban) clusters, the BI remained consistently at 55% or more below pre intervention levels in both settings throughout the 18-month follow up period. In contrast, BI levels in urban and suburban municipalities (59.0 and 82.6 respectively in October 2006) fluctuated considerably and did not show the same patterns as the study areas (Figure 3). The differences in average BI between the pre- and post-intervention period was −63% for urban and −67% for suburban intervention clusters. For the corresponding municipal areas these differences were −35% and −26% respectively.
In the random effects negative binomial regression models (Table 3), the setting (urban or suburban) and the amount of rainfall were significantly correlated with the BI and PPI in the intervention clusters. Overall infestation levels at the municipality level and temperature had no significant effect, and it was interesting to note that the intensity of routine A. aegypti control activities in the clusters also had no effect. Temperature was not included in the final model because it did not confound the relationship between BI or PPI and ITcurtain coverage. ITcurtain coverage was highly significantly correlated with both entomological indicators, but ITcover coverage was not. Each 1% coverage increase with ITcurtains reduced the BI and PPI by 2%. Plotting this effect (Figure 4), reveals that 50% coverage or more was needed to halve the BI, and that the reduction, in absolute terms, depended on the initial BI.
The presence of insecticide treated window curtains in an environment where A. aegypti infestation levels are moderate (BI ranging between 10 and 50) can lead to substantial reductions in the Breteau index and the pupae per person index. The scale of the effect depends on the household coverage attained, and without any further intervention, curtain usage may rapidly decline over time. The demonstration of an effect on the PPI, in addition to the BI, is important, as PPI is considered a more accurate measure of local adult vector abundance, and therefore more directly related to dengue transmission risk [21].
We were unable to directly monitor adult A. aegypti populations (due to operational reasons and resource constraints) let be to measure dengue transmission. This study was also limited by the fact that it is a before and after evaluation and that we did not include randomized control clusters in the design, but used routine entomological surveillance data from the whole municipality as ‘control’ data. However, it is not likely that temporal trends in vector density should selectively affect the intervention clusters only and bias, if any, could not explain the differences that we observed. Importantly, the effects we attribute to the actual coverage with ITcurtains and ITcovers are independent of the comparison with the control data at specific time points, and of possible confounding factors such as rainfall, temperature, routine vector control activities and temporal trends that were controlled for in the analysis. Additionally, it has been demonstrated that the insecticide in the PermaNet curtains, when used by households, remains effective for at least 1 year [18]. Furthermore, the local mosquito population remained susceptible to deltamethrin, as shown in bioassays on A. aegypti collected in neighbouring municipalities where the same ITcurtains were concurrently deployed in the frame of another study (A. Lenhart, personal communication).
Major strengths of this study, from a public health perspective, are the length of the pre- and post intervention data collection periods and the fact that the ITMs were introduced into the community by the routine vector control programme and the local health committees, which mimics the reality of routine operational conditions. Both these elements markedly distinguish our approach from the one used in the only previous study on ITcurtain deployment for dengue control [15]. In that study, Kroeger et al. [15] reported no differences in entomological indices between intervention and control arms. This lack of a difference was attributed to a “spill-over” of the effect of the IT curtains in the intervention clusters into the adjacent control clusters. The authors performed therefore a before and after evaluation of Stegomyia indices, using routine surveillance data collected in nearby communities, and concluded that the changes in Stegomyia indices in the control and intervention study areas combined, could not be explained by natural fluctuations in the vector population due to seasonal parameters.
The coverage attained at ITcurtain distribution in our study is lower, but still comparable to the 87% coverage attained in the Venezuelan site of the above efficacy trial [15], in which delivery was controlled by the research team, and we achieved reductions in vector density of the same order of magnitude. However, our follow up period was much longer and our results on sustained use differ markedly. While coverage remained stable up to the final observation at month 5 in Kroeger et al. [15], we observed a 20% decline in use 6 months after distribution and over 50% at 18 months. The determinants of uptake and continued use of the ITM have been assessed [23] and it was found that uptake was linked to pre-use behaviour and contextual factors, but continued use was mainly determined by the perceived effectiveness of the tool. Furthermore, our results demonstrate that the level of coverage attained has profound implications for the effectiveness of ITcurtain interventions.
We did not find a significant effect of ITcovers on entomological indices. Kroeger et al. [15] reported on the combined efficacy of ITcurtains and ITcovers, but did not study ITcover efficacy independently. They also remarked that the covers were not durable and were easily torn and that use decreased by over 30% in the initial 5 months, which concurs with our own observations. In contrast, a trial of ITcovers in Cambodia [17] reported a much higher coverage of 3.1 ITcovers per house and a 58% reduction in PPI at 13 weeks in the intervention area as compared to the control area (dropping to 13% at 22 weeks post-intervention). Inherent differences between the studies, in A. aegypti oviposition behavior and/or variations in ITcover coverage or the quality of the materials used might explain the differences in results between studies, but further research is needed to clarify exactly how ITcovers impact on dengue vector populations.
Comparisons of our results with those obtained in controlled trials of other dengue vector control tools are difficult, certainly at quantitative level, as there are variations in study design, follow-up periods and/or outcomes measured. Although direct comparisons cannot be made, a number of issues are noteworthy. First, apart from the promising results observed with insecticide treated materials [15]–[17], and small field-laboratory studies of lethal ovitraps, (e.g. in Brazil; [24]), all published controlled intervention studies to date have targeted immature dengue vector stages. Secondly, substantial and sometimes sustainable effects have been reported with approaches combining chemical and biological control [25], chemical control and community based environmental management [13], [14], [26] , and biological control and environmental management [27]. Hence, strategies integrating multiple control measures appear to be more effective [28]. Thirdly, only very few of the control strategies managed to completely eliminate the vector [8].
Against this backdrop, we cautiously state that ITcurtains constitute a potentially effective novel tool for controlling A. aegypti, with efficacy likely to be optimized when deployed in combination with other vector control tools, and particularly when their use is embedded in a strategy that also mobilizes the community. However, before calling for the launch of large scale integrated effectiveness trials with ITcurtains, important questions remain regarding the efficacy, cost and implementation of ITM strategies: Does long lasting material remain effective beyond one year [18] when heavily exposed to sunlight and dust? How efficacious are ITMs at low or very high A. aegypti infestation levels and, ultimately, what is their impact on dengue transmission? What is their incremental cost-effectiveness at 0.93 USD per m2 of fabric plus approximately 0.5 USD for distribution [29]? And, obviously, finally, how can the high level of coverage required for effectiveness be attained and maintained under routine conditions?
|
10.1371/journal.pgen.1007630 | The uterine epithelial loss of Pten is inefficient to induce endometrial cancer with intact stromal Pten | Mutation of the tumor suppressor Pten often leads to tumorigenesis in various organs including the uterus. We previously showed that Pten deletion in the mouse uterus using a Pgr-Cre driver (Ptenf/fPgrCre/+) results in rapid development of endometrial carcinoma (EMC) with full penetration. We also reported that Pten deletion in the stroma and myometrium using Amhr2-Cre failed to initiate EMC. Since the Ptenf/fPgrCre/+ uterine epithelium was primarily affected by tumorigenesis despite its loss in both the epithelium and stroma, we wanted to know if Pten deletion in epithelia alone will induce tumorigenesis. We found that mice with uterine epithelial loss of Pten under a Ltf-iCre driver (Ptenf/f/LtfCre/+) develop uterine complex atypical hyperplasia (CAH), but rarely EMC even at 6 months of age. We observed that Ptenf/fPgrCre/+ uteri exhibit a unique population of cytokeratin 5 (CK5) and transformation related protein 63 (p63)-positive epithelial cells; these cells mark stratified epithelia and squamous differentiation. In contrast, Ptenf/fLtfCre/+ hyperplastic epithelia do not undergo stratification, but extensive epithelial cell apoptosis. This increased apoptosis is associated with elevation of TGFβ levels and activation of downstream effectors, SMAD2/3 in the uterine stroma. Our results suggest that stromal PTEN via TGFβ signaling restrains epithelial cell transformation from hyperplasia to carcinoma. In conclusion, this study, using tissue-specific deletion of Pten, highlights the epithelial-mesenchymal cross-talk in the genesis of endometrial carcinoma.
| Endometrial cancer is highly prevalent gynecological cancer in the United States. Pten is the most commonly mutated gene in endometrial carcinoma originating in the epithelium. Previous studies focused on PTEN’s role in epithelial growth regulation. Here we show that in addition to Pten mutation in the epithelium, its mutation in the stromal compartment is critical for the initiation and progression of endometrial carcinoma. We present evidence that while loss of Pten function in both uterine epithelia and stroma results in rapid development of endometrial carcinoma, its loss in epithelial cells leads to endometrial hyperplasia, but not carcinoma. Our findings highlight the critical role of stromal PTEN in the transformation of hyperplasia to carcinoma and stromal TGFβ appears to play a role in preventing this transformation. This study reveals a previously unidentified role of PTEN in influencing the microenvironment in the uterus for the initiation and generation of endometrial carcinoma.
| Endometrial carcinoma (EMC) is the most common cancer of the female reproductive organs in the United States. In 2017, about 60,000 new cases were diagnosed and about 11,000 deaths occurred related to EMC in the US [1, 2]. EMC has been categorized into two major types: type I endometrioid cancers are focused in the endometrial gland cells, and type II non-endometrioid cancers are often of serous morphology. Type I represents approximately 85% of EMCs in which Pten is commonly mutated. Other than endometrial cancer, Pten mutations are also evident in endometrial hyperplasia [3–5]; hyperplasia is a well-established precursor lesion of EMC [6]. The understanding of divergence between hyperplasia and cancer is of clinical significance. On one hand, faulty diagnosis of complex atypical hyperplasia (CAH) may lead to hysterectomy [7], a non-reversible procedure that negatively impacts women seeking to preserve fertility. Alternatively, diagnosis at early stage of hyperplasia may prevent progression to carcinoma. Although stromal invasion and histological changes are considered diagnostic standards of EMC [8], identification of the biomarkers for early stage carcinomas and the mechanism underlying cancer progression are greatly needed.
Pten homozygous null mice are embryonic lethal. Therefore, Pten heterozygous mice are widely used for cancer studies [6]. Pten heterozygous females show atypical endometrial hyperplasia phenotype, with 20% developing cancer. Using the Cre-loxP system and Pgr-Cre driver, we previously showed that Ptenf/fPgrCre/+ mice with endometrial Pten deletion develop epithelial carcinoma as early as one month of age with Pten loss in major uterine cells [9]. To study roles of Pten in different uterine cell types, we created mice with Pten deletion specifically in the stroma and myometrium using Amhr2-Cre driver [10]. Ptenf/f/Amhr2Cre/+ females showed no EMC; instead myometrial cells transformed into adipocytes [11]. Taken together, these findings suggest epithelial origin of this pathology. We thought that the epithelial origin of EMC could be tested if only epithelial-specific loss of Pten is induced in the uterus, disrupting the cross-talk between the stroma and epithelium to initiate EMC and its progression.
To address this issue, an efficient Cre mouse line is necessary to specifically delete epithelial genes. Pten was conditionally deleted in the epithelium using Wnt7a-Cre, but the mutant pups died around 10 days of age [11]. Conditional deletion of Pten using Sprr2f-Cre met with failure because of brain cancer and limited life span [11, 12]. To generate a mouse line with Cre activity specifically in the adult uterine epithelium, we generated a mouse line expressing codon-improved Cre (iCre) under a Lactoferrin (Ltf) promoter. By crossing with LacZ reporter mice, we showed that the Ltf-driven iCre expression exhibits robust Cre activity in uterine luminal and glandular epithelia beginning at puberty [13]. In contrast to Cre expression driven by promoters of Pgr, Amhr2, and Wnt7a that occur before or right after birth, Cre activity driven by Ltf promoter is activated with the beginning of estrous cycle [13].
In this study, using LtfCre/+ mice, we established the mouse model with uterine epithelial-specific Pten deletion by crossing with Ptenf/f mice. Surprisingly, Ptenf/fLtfCre/+ females rarely develop EMC, but show epithelial CAH. We also found that Ptenf/fLtfCre/+ females do not readily form stratified epithelial layers which are prevalent in Ptenf/fPgrCre/+ uteri. Ptenf/fPgrCre/+ epithelial layers show the presence of CK5, a stratified epithelial cell marker, and CK8 that is primarily expressed in simple epithelial cells. CK5 positive cells are located between CK8 positive epithelia and stroma, and the two populations of CK5 and CK8-positive cells are mutually exclusive in Ptenf/fPgrCre/+ epithelia. p63 is critical for initiation of epithelial stratification [14], and has been identified as a prognostic marker in multiple cancers [15, 16]. Interestingly, p63 is expressed in CK5 positive cells in Ptenf/fPgrCre/+ epithelia. We further identified TGFβ, which represses stratification of epithelium [17], is downregulated in the Ptenf/fPgrCre/+ stroma as compared to that in Ptenf/fLtfCre/+ mice. The differential phenotypes between Ptenf/fPgrCre/+ and Ptenf/fLtfCre/+ mice highlight the crucial role of stromal microenvironment and stromal-epithelial interactions in EMC progression.
In Ptenf/fLtfCre/+ mice, genomic deletion of Pten begins at 1 month of age (S1A Fig). By 2 months of age, rare PTEN positive signals, if any, are observed in the uterine epithelium (Fig 1A, S1B Fig). Notably, levels of PTEN expression in Ptenf/fLtfCre/+ stroma are upregulated as compared to those in Ptenf/f mice at 3 months of age. We examined whether epithelial deletion of Pten produces EMC. Pathological analysis shows that Ptenf/fLtfCre/+ uteri exhibit normal histology by 1.5 months old, but most of Ptenf/fLtfCre/+ uteri start developing CAH from 2 months of age. By 4 months, only 2 of 8 (25%) mice show focal myometrial invasion (Table 1). Ptenf/fLtfCre/+ mice at 6 and 12 months of age also show atypical glandular hyperplastic epithelia showing medium to large cysts with fluid retention. This glandular hyperplasia perhaps predisposed to carcinoma, but no epithelial invasion to the myometrium was evident (S2 Fig). Further analysis shows that PTEN expression patterns are comparable in Ptenf/fLtfCre/+ mice with or without myometrial invasion (S3 Fig). Detailed pathological analyses at different ages is presented in Table 1. The results show that progression of EMC is dramatically retarded in Ptenf/fLtfCre/+ mice as compared to that in Ptenf/fPgrCre/+ uteri [9], suggesting that stromal Pten suppresses transformation of CAH to EMC.
PTEN, a phosphoinositide 3-phosphatase, metabolizes phosphatidylinositol 3,4,5-trisphosphate (PIP3) [18, 19], and suppresses AKT activation [20, 21]. As expected, AKT activation markedly increases in Ptenf/fLtfCre/+ uterine epithelia at both 2 and 3 months of age (S4A Fig). Previously, we reported that mTORC1 is a downstream target of PTEN/AKT signaling in Ptenf/fPgrCre/+ uteri [22]. In Ptenf/fLtfCre/+ mice, mTORC1 activation is upregulated in epithelia at both 2 and 3 months of age, as evident from elevated levels of phosphorylated ribosomal protein S6 (pS6), a downstream effector of mTORC1 (S4B Fig). Heightened COX-2 expression and mTORC1 activity exacerbate EMC in Ptenf/fPgrCre/+ uteri [22]. In Ptenf/fLtfCre/+ uteri, COX-2 expression is induced in Ptenf/fLtfCre/+ epithelia (S4C Fig). Western blotting results confirmed upregulated levels of p-AKT and pS6 in 2-month old Ptenf/fLtfCre/+ uteri (S5 Fig).
We compared the expression levels of p-AKT, pS6, and COX-2 in uteri of Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ females. Levels of p-AKT and pS6 are higher in Ptenf/fLtfCre/+ uterine epithelia as compared to that in Ptenf/f uteri (Fig 2A and 2B). We have previously shown that COX-2 expression is associated with endometrial cancer progression, and inhibition of COX-2 slows down cancer development and progression [22]. Scattered signals of COX-2 are observed in Ptenf/fLtfCre/+ epithelia, whereas COX-2 positive cells are widely distributed in Ptenf/fPgrCre/+ epithelia and underneath stroma (Fig 2C). These results suggest that Ptenf/fLtfCre/+ epithelial cells are less invasive as compared to Ptenf/fPgrCre/+ epithelium. However, Ptenf/fPgrCre/+ epithelial cells do not appear to undergo epithelial mesenchymal transition (EMT) as evident from staining of EMT markers E-cadherin (S6A Fig) and Desmin (S6B Fig) [23, 24].
Previous studies showed that p63, a p53 homologue, is a marker of metaplastic differentiation, including basal/squamous differentiation and is found in stratified human tumors including EMC [25, 26]. Thus, we explored the expression of p63 in Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ uteri. As shown in Fig 3A, p63 is localized primarily in the basal layer of luminal epithelia of Ptenf/fPgrCre/+ mice and surrounding glands at 3 months of age, while no p63 signal was observed in Ptenf/fLtfCre/+ uteri at the same age. Notably, p63 positive cells express E-cadherin (S6C Fig), suggesting that these cells maintain epithelial characteristics. Trp63 encodes multiple isoforms of p63, including full length TA isoforms with an acidic transactivation domain and ΔN isoforms lacking this domain [27]. Therefore, we used Western blotting analysis to assess the isoforms of p63 in uterine lysates of Ptenf/fPRcre/+, Ptenf/fLtfcre/+ and respective littermate controls. The result shows that p63 in Ptenf/fPRcre/+ epithelia is of TA isoform (S6D Fig).
Cytokeratin can also be used to distinguish simple or stratified epithelium [28]. CK8 is produced by simple epithelia [29], while CK5 is particularly expressed in the basal layer of stratified squamous epithelium. We found that the expression pattern of CK5 is similar to that of p63 (Fig 3B). Interestingly, co-staining of p63 or CK5 with CK8 identified that the expression pattern of p63/CK5 and CK8 are mutually exclusive in Ptenf/fPgrCre/+ uteri (Fig 3A and 3B). These results suggest a potential relationship between p63 and EMC. Since two Ptenf/fLtfCre/+ mice of 4 month old showed myometrial invasion, we examined the expression of p63 in these mice. Notably, p63 positive cells were observed in Ptenf/fLtfCre/+ mice with myometrial invasion (Fig 3C). These results again suggest that expression of p63 correlates with EMC. To study the correlation between p63 and carcinoma progression, the expression of p63 in Ptenf/fPgrCre/+ uteri was examined in uteri of 1 and 2-month old Ptenf/fPgrCre/+ mice. Ptenf/fPgrCre/+ uteri at 1 month of age are negative for p63 signal, but p63-positive cells appear underneath the CK8 positive luminal epithelium at 2 months of age (Fig 3D). These results provide evidence that stromal PTEN restrains epithelial stratification, and p63 serves as an indicator of EMC.
We explored the underlying mechanism preventing epithelial carcinoma by stromal PTEN. Since Ki67 positive cells are present at the leading edge of the tumor in Ptenf/fPgrCre/+ uteri [9], we examined the distribution of Ki67-positive cells in Ptenf/fLtfCre/+ uteri. As shown in Fig 4A, strong signals for Ki67 are present in both Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ uterine epithelium. Remarkably, Ki67 staining is more intense in the CK8-negative luminal epithelium. These results were corroborated by co-staining of CK8, p63, and Ki67 staining on the consecutive sections (Fig 4C). Ki67 signals are localized in p63-positive epithelia. The staining of phosphor-Histone H3 (pHH3) in Ptenf/fPgrCre/+ uteri also showed similar expression pattern to that of Ki67 (Fig 4B and 4C). The results show that CK8 positive epithelial cells are proliferative in Ptenf/fLtfCre/+ uteri, whereas epithelial cells in the p63-positive layer show cell proliferation in Ptenf/fPgrCre/+ uteri. To better understand the turnover of epithelial cells in Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ uteri, we examined cell apoptosis by cleaved-Caspase 3 (caspase-3) immunostaining and observed increased cell population of caspase-3 positive cells in Ptenf/fLtfCre/+ epithelia; the signal is limited in Ptenf/fPgrCre/+ epithelia (Fig 4D). These results indicate that the PTEN-positive stroma in Ptenf/fLtfCre/+ uteri restricts the epithelial hyperplasia by promoting apoptosis in hyperplastic epithelia, while Pten deletion in the stroma in Ptenf/fPgrCre/+ uteri fails to prevent excessive proliferation and transform hyperplastic epithelial cells to EMC. Notably, Ptenf/fAmhr2cre/+ uteri with Pten deletion in the stroma show apparently normal proliferation and apoptosis in epithelia (S7 Fig), suggesting stromal PTEN has limited impact on epithelial growth under normal physiological conditions.
Uterine cell proliferation and differentiation is regulated by ovarian hormones through ESR1 [30] and PR [31]. We examined the expression of these two nuclear receptors. The results show that the expression of ESR1 and PR is maintained in all major cell types in both Ptenf/fPRcre/+ and Ptenf/fLtfcre/+ uteri (S8 Fig).
Given extensive apoptosis in Ptenf/fLtfCre/+ uteri, we then asked if immune cells play a role in apoptotic cell clearance in Ptenf/fLtfCre/+ uteri. First, we accessed the distribution of CD45-positive cells of hematopoietic origin. Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ uteri show increased population of immune cells in the uterus (Fig 5A); the weight of spleen, liver and thymus did not show many changes (S9 Fig). The uterine recruitment of immune cells suggests local inflammation. As previously reported that neutrophils are recruited in Ptenf/fPgrCre/+ uteri [32], the population of Ly6G-positive cells, a marker of neutrophils, is much higher in Ptenf/fPgrCre/+ uteri (Fig 5B). Interestingly, increased CD45-positive cells in Ptenf/fLtfCre/+ are not neutrophils but macrophages, as shown by F4/80 staining (Fig 5C). The infiltration of different immune cells could be due to extensive apoptotic or metaplastic cells in Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ uteri respectively.
Two types of macrophages (M1 and M2) exhibit diverse phenotypes and functions. We examined the expression of MHCII and CD206, markers for M1 and M2 macrophages, respectively, to determine which subtypes contribute to increased macrophage population in Ptenf/fLtfCre/+uteri. The results show that the number of MHCII-positive cells is much higher in Ptenf/fLtfCre/+ uteri than that in Ptenf/fPgrCre/+ (Fig 5D). However, no significant differences in CD206-positive M2 macrophages are observed (Fig 5E). The quantification of M1 versus M2 macrophages is shown in Fig 5F and 5G. Furthermore, F4/80-positive signals do not co-localize with Ki67 or caspase-3 signals (S10A and S10B Fig), suggesting that resident macrophages in the uterus do not proliferate but migrate from the circulation. These results suggest a potential role of macrophages in clearing apoptotic epithelial cells.
TGFβ signaling has been reported to have a dual function during the progression of carcinoma: cell-cycle arrest and apoptosis in the early-stage cancer and tumorigenesis at the late stage [33]. TGFβ signaling also inhibits epithelial stratification [17]. Using immunofluorescence, we observed distinct TGFβ signals in the stroma of Ptenf/fLtfCre/+ uteri, whereas the signal is much lower in Ptenf/fPgrCre/+ stroma, especially in the stroma surrounding the luminal epithelium (Fig 6A). The distribution of phosphorylated SMAD2/3 (p-SMAD2/3), a downstream effector, correlates with TGFβ signaling [34]. Consistent with TGFβ staining, the activation of p-SMAD2/3 is significantly lower in Ptenf/fPgrCre/+ uteri (Fig 6B). These results suggest that stromal Pten potentially exerts its tumor suppressive role by upregulating TGFβ signaling.
To study if our findings have any relevance to human uterine corpus endometrial carcinoma (UCEC), we compared the RNA profile from patients with UCEC and controls that are available in RNA-seq dataset from The Cancer Genome Atlas (TCGA). Our analysis shows that the UCEC group has significantly lower levels of PTEN and TGFβ RNA, as well as TGFβ’s target genes, SERPINE1 and ID1, as compared to those in control tissues (Fig 6C). This is consistent with low TGFβ levels in Ptenf/fPgrCre/+ uteri. (Fig 6C).
PTEN is considered as a tumor suppressor protein. Pten mutations are closely related to various types of tumorigenesis, especially type I EMC [35]. Our present and previous studies using cell specific deletion of Pten in the uterus provide evidence that absence of Pten in the epithelium, stroma and myometrium promptly produces EMC, while its deletion in the stroma and myometrium fail to generate EMC but transforms myometrial cells to adipocytes. Surprisingly, Pten deletion specifically in the epithelium primarily shows CAH.
The function of epithelial Pten has been studied using several approaches. Adenovirus was used to delete endometrial epithelial Pten by intraluminal injection [36], although a small percentage of endometrial stromal cells of adeno-Cre injected mice showed Pten deletion. Conditional Pten deletion using Wnt7a-Cre and Ksp-Cre in combination with Pik3ca mutation was also reported [37]. Combination of Pten deletion and Pik3ca mutation leads to carcinoma, while Pik3ca mutation alone showed no EMC or hyperplasia phenotype. Similar to the CAH phenotype in the uterus, Pten deletion leading to hyperplasia has been corroborated in several other different epithelial tissues besides the endometrial epithelium, such as urothelial cells, keratinocytes, prostatic epithelial cells, and lung epithelium [6]. Furthermore, the glandular epithelium specific Pten deletion also showed endometrial hyperplasia [38]. As Pten is deleted in both luminal and glandular epithelia in our Ltf-iCre model, definitive answers to distinguish the role of Pten in the luminal or glandular epithelium will require a luminal epithelium-specific deletion mouse model. It is also of interest to evaluate the uterine phenotype in stroma and glandular-deletion or stroma and luminal-deletion of Pten using a combination of Cre systems.
Our study with Ptenf/fLtfCre/+ uteri suggests that stromal Pten restrains transition of hyperplasia to carcinoma. In contrast, the deficiency of this gene in three major uterine cell types (Ptenf/fPgrCre/+) with rapid generation of EMC suggests that Pten-deleted stroma provides a more susceptible microenvironment for further deterioration of hyperplastic epithelium into EMC. In this regard, the role of endometrial stroma in EMC was reported using a stromal-specific Lkb1-deleted mouse model in which the loss of Lkb1 in the stroma was sufficient to initiate neoplasia [39]. A previous study also showed that EMC develops in uteri with epithelial modification in both Pten and Pik3ca [37]. In the mouse uterus, epithelial deletion of Pten alone is not sufficient to induce EMC. In human cancer specimens, PTEN is predominantly lost in the epithelium and maintained in the stroma [40]. EMC was also observed in a transplant model, in which a mixture of Pten deficient epithelial cells and WT stromal cells were transplanted under kidney capsule [40]. In spite of these findings, many questions still remain about the differences between human and mouse models of cancers.
The higher levels of pS6 signal in Ptenf/fLtfCre/+ epithelia as compared with that in Ptenf/fPgrCre/+ mice at 3 months of age suggest that activation of mTORC1 is closely associated with hyperplasia. Our previous study using mice with whole uterine deletion of Tsc1 also supports this conclusion [41]. Interestingly, mice with Tsc1 deletion in the stroma and myometrium also shows hyperplasia, suggesting the existence of unidentified paracrine signals from stromal influencing epithelial proliferation. Furthermore, our results with rapamycin (an inhibitor of mTORC1 signaling) suggest that inhibition of mTORC1 signaling could be an effective preventive strategy to combat endometrial hyperplasia and/or EMC. We have also shown that inhibition of upregulated COX-2 in the uterus of Ptenf/fPgrCre/+ mice is reduced by a COX-2 inhibitor (Celecoxib) with attenuated EMC development [22]. In the present study, COX-2 is also induced in hyperplasic Ptenf/fLtfCre/+ epithelia. By comparing the expression of COX-2 in Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ uteri, we found that the COX-2 level is much higher in Ptenf/fPgrCre/+ uteri, suggesting hyperplasic cells are less invasive than cancerous cells.
The current study demonstrates that the expression of p63 is closely associated with EMC development. However, the role of p63 in uterine luminal epithelial stratification is still not clear. p63 plays multiple roles in development depending on different contexts [42]. p63 is required for establishing stratified epithelia perhaps by maintaining stem cell populations or triggering differentiation of simple epithelia into stratified epithelia [43, 44]. In humans, p63 is expressed in hyperplastic and metaplastic endometria [25]. It is possible that p63 suppresses epithelial metaplasia and prevents epithelia from further invading into the muscle layer, since the loss of p63 in tumor tissues is associated with more aggressive EMC [26]. p63-positive cells invade the area underneath p63-negative columnar cells and push them upward, which leads to the detachment of p63-negative cells [45]. However, we cannot rule out the possibility that p63 itself promotes EMC development.
TGFβ signaling appears to constrain hyperplastic Ptenf/fLtfCre/+ epithelia from stratification toward tumorigenesis. TGFβ acts as a tumor suppressor in the epithelium [46] and restricts epithelial growth and early tumor development [47]. In mouse uteri, TGFβr1 mRNA is detected mainly in the epithelia of Ptenf/fPgrCre/+ uteri. SMAD2 is highly expressed in uterine epithelium at the proestrus phase. These results suggest a role for TGFβ in epithelial proliferation [48]. Pten and TGFβr1 double knockout mice using Pgr-Cre driver show severe endometrial lesions with disrupted myometrial layers and pulmonary metastasis [49], suggesting a role for TGFβr1 in cancer progression. Mice with uterine stromal TGFβr1-deletion using Amhr2-Cre show enhanced proliferation in both luminal and glandular epithelia [50], suggesting TGFβ signaling is involved in epithelial-stromal interactions. In this study, we observed PTEN levels are upregulated in the stroma of Ptenf/fLtfCre/+ mice (Fig 1A) and is associated with heightened stromal TGFβ and pSMAD2/3 levels. In contrast, TGFβ levels are suppressed in Ptenf/fPgrCre/+ stroma, indicating stromal TGFβ signaling may play a role in preventing epithelial tumorigenesis. Taken together, these data indicate that Pten expression in the stroma maintains stromal TGFβ expression, which perhaps limits epithelial growth.
TGFβ signaling plays a role in cell proliferation and apoptosis. In mouse uteri, reduced TGFβ signaling leads to loss of growth-inhibitory response, and constitutively activated TGFβr1 reduces glandular growth [51], suggesting an inhibitory role for TGFβ in epithelial proliferation. TGFβ signaling also plays a role in cell apoptosis. In polarized endometrial epithelial cells, TGFβ induces apoptosis via SMAD3 [52]. Pten knockdown blocks TGFβ-induced apoptosis and leads to increased cell proliferation. We observed intense cell apoptosis in Ptenf/fLtfCre/+ mice compared to Ptenf/fPgrCre/+ uteri, in which cell apoptosis is rarely seen. Interestingly, epithelial cell proliferation is not affected by changes in TGFβ signaling as evident by comparable numbers of Ki67 positive cells in Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+. Given the role of TGFβ in apoptosis, the decreased levels of TGFβ and apoptosis in Ptenf/fPgrCre/+ uteri suggest stromal PTEN-driven TGFβ prevents epithelial tumorigenesis by promoting epithelial cell apoptosis.
Pten deletion often leads to cancer in situ [53]. However, PTEN’s role in providing a microenvironment conducive to cancer progression is not clear. By comparing the phenotype of Ptenf/fLtfCre/+ and Ptenf/fPgrCre/+ mouse uteri, we show here that EMC development requires Pten deletion in both the stroma and epithelium. Our data also imply that stromal regulation of epithelial growth is mediated by TGFβ signaling. Our current study presents new insights into the role of Pten in the microenvironment for tumorigenesis.
All mice were housed in the Cincinnati Children’s Hospital Medical Center Animal Care Facility in conformity with NIH and institutional guidelines. PtenloxP/loxP mice (stock number 004597, 129S4/SvJae/BALB/cAnNTac) were obtained from the Jackson Laboratory (Sacramento, CA, USA). Ptenf/fPgrCre/+ (129S4/SvJae/BALB/cAnNTac/C57BL/6) mice and Ltf-iCre mice (129/C57BL/6/albino B6) were generated as previously described [9, 13]. Ptenf/fLtfCre/+ were generated by crossing PtenloxP/loxP and Ltf-iCre mice. Littermate floxed mice were used as controls in all experiments. Uterine tissues from the diestrous stage were collected for experiments.
For paraffin sections, tissues were fixed in Safefix (Thermo Fisher Scientific, Lafayette, CO, USA) and embedded in paraffin. After deparaffinization and hydration, sections (6 μm) were subjected to antigen retrieval by autoclaving in 0.01M sodium citrate solution (pH = 6) for 10 min. For frozen tissues, sections (12 μm) were fixed in 4% paraformaldehyde solution. Depending on the primary antibody (S1 Table), some sections were subjected to antigen retrieval by autoclaving in 0.01M sodium citrate solution (pH = 6) for 10 min. COX-2 and TGFβ antibodies were custom-made as previously described [54, 55]. For immunohistochemistry, Histostain-Plus kit (Invitrogen, Carlsbad, CA, USA) was used to visualize signals. Immunofluorescence was performed using secondary antibodies conjugated with Alexa 488 or Alexa 594 (Jackson ImmunoResearch, West Grove, PA, USA). Hematoxylin and Hoechst were used for counterstain in immunohistochemistry and immunofluorescence, respectively. For all images of pHH3 staining at lower magnification, the maximum filter of ImageJ was applied to the red staining channel for clear visibility.
Numbers of M1 and M2 macrophages were calculated by counting MHCII and CD206 positive cells according to immunofluorescence staining. Sections from 3 different mice, and 4 fields per section have been evaluated.
cRNA probes for Pten were generated by reverse RT-PCR followed by 35S-labeling using Sp6 or T7 RNA polymerases. Paraformaldehyde-fixed frozen sections (12 μm) were hybridized with 35S-labeled cRNA probes of Pten as previously described [9].
Western blotting was performed as previously described [11]. Briefly, uterine protein samples from uteri at the diestrous stage were run on 10 or 12% SDS-PAGE gels depending on the molecular weights of proteins and transferred onto PVDF membranes. After blocking in 5% BSA for detection of phosphorylated protein, or in 10% non-fat milk for detection other proteins, membranes were blotted with antibodies to PTEN, p-AKT, AKT, pS6, S6, p63 and β-ACTIN. Signals were detected using ECL reagents (GE healthcare, Pittsburgh, PA, USA).
RNAseq data were downloaded from TCGA data portal (https://tcga-data.nci.nih.gov/). RNAseq data from 176 UCEC cases and 23 controls were used for data analysis. Transcript-levels of genes were calculated using RNA-Seq by Expectation Maximization (RSEM) method.
Data were analyzed by the Mann Whitney tests. P<0.05 was considered significant. Values are mean ± SEM.
|
10.1371/journal.ppat.1000689 | Hypoxia and the Hypoxic Response Pathway Protect against Pore-Forming Toxins in C. elegans | Pore-forming toxins (PFTs) are by far the most abundant bacterial protein toxins and are important for the virulence of many important pathogens. As such, cellular responses to PFTs critically modulate host-pathogen interactions. Although many cellular responses to PFTs have been recorded, little is understood about their relevance to pathological or defensive outcomes. To shed light on this important question, we have turned to the only genetic system for studying PFT-host interactions—Caenorhabditis elegans intoxication by Crystal (Cry) protein PFTs. We mutagenized and screened for C. elegans mutants resistant to a Cry PFT and recovered one mutant. Complementation, sequencing, transgenic rescue, and RNA interference data demonstrate that this mutant eliminates a gene normally involved in repression of the hypoxia (low oxygen response) pathway. We find that up-regulation of the C. elegans hypoxia pathway via the inactivation of three different genes that normally repress the pathway results in animals resistant to Cry PFTs. Conversely, mutation in the central activator of the hypoxia response, HIF-1, suppresses this resistance and can result in animals defective in PFT defenses. These results extend to a PFT that attacks mammals since up-regulation of the hypoxia pathway confers resistance to Vibrio cholerae cytolysin (VCC), whereas down-regulation confers hypersusceptibility. The hypoxia PFT defense pathway acts cell autonomously to protect the cells directly under attack and is different from other hypoxia pathway stress responses. Two of the downstream effectors of this pathway include the nuclear receptor nhr-57 and the unfolded protein response. In addition, the hypoxia pathway itself is induced by PFT, and low oxygen is protective against PFT intoxication. These results demonstrate that hypoxia and induction of the hypoxia response protect cells against PFTs, and that the cellular environment can be modulated via the hypoxia pathway to protect against the most prevalent class of weapons used by pathogenic bacteria.
| Bacteria make many different protein toxins to attack our cells and immune system in order to infect. Amongst them, pore-forming toxins (PFTs), which punch holes in the protective plasma membrane that surrounds cells, are by far the most abundant and constitute important virulence factors. Since the integrity of the plasma membrane is fundamental to maintaining the normal intracellular environment, the breaching of the plasma membrane by PFTs results in many and dramatic intracellular responses. However, we know little about the relevance of these responses to cell survival or cell intoxication. Here, using the only genetic system for studying pore-forming toxin effects in a whole animal, we show that the same response that protects cells against low oxygen stress unexpectedly also protects cells against pore-forming toxins. Mutations in the animal that hyper-activate the low oxygen response actually make animals resistant to pore-forming toxin attack, whereas mutations that inactivate the low oxygen response make animals more susceptible. Furthermore, a low oxygen environment itself is protective against pore-forming toxins. These data show a new and powerful connection between low oxygen responses and defense against the single most common mode of bacterial attack.
| Pore-forming toxins (PFTs) are by far the most abundant and amongst the most important bacterial protein virulence factors [1]. These toxins, secreted by many pathogenic bacteria, function by binding to receptors on host cell plasma membrane, oligomerizing, inserting, and then forming holes [2]. The importance of PFTs in promoting bacterial pathogenesis has been demonstrated by numerous experiments where individual PFTs have been genetically deleted from pathogenic bacteria and the bacteria then tested for reduced virulence [3]. Examples of PFTs that contribute significantly to bacterial virulence include α-toxin by Clostridium septicum, streptolysins by Group A and B Streptococci, α-toxin by Staphylococcus aureus, Vibrio cholerae cytolysin (VCC), α-hemolysin from uropathogenic E. coli, cytolysin from Enterococcus faecalis, and crystal (Cry) proteins from Bacillus thuringiensis (Bt).
Although they are expressed by many bacterial pathogens and are broadly important as potentiators of infection, the effects of these toxins on host cells have been vastly understudied. There are several reasons for this lack of attention. First, their mechanism of action is deceptively simple. Second, most of the attention has been given to understanding how PFTs can change conformation from secreted, soluble proteins to insoluble proteins embedded in the plasma membrane. Third, because breaching of the plasma membrane is a major insult to a cell, a multitude of cellular responses to PFTs have been reported, including Ca2+ influx, K+ efflux, increased endocytosis/exocytosis, vacuolization, necrosis, and apoptosis [3],[4],[5],[6]. Because the responses are so large and extensive, it has been daunting to determine whether these responses contribute to defense, intoxication, both, or neither. Fourth, most of the studies carried out to date involved cells in isolated culture, which does not always accurately recreate the response of cells to toxins in the context of intact tissue.
To address many of these limitations, an excellent genetic system for studying PFT responses in an intact animal has recently emerged: the Bacillus thuringiensis (Bt) Crystal (Cry) PFT – Caenorhabditis elegans toxin-host interaction system [7]. C. elegans has become an important genetically tractable organism for studying immune responses to bacterial pathogens [8]. Bt is thought to be a natural pathogen of C. elegans [9],[10],[11] and is famous for the production of three-domain PFTs that are widely used in insect biocontrol [12]. The interaction of Cry proteins with C. elegans allowed for the first molecular PFT defense pathway identified, p38 mitogen-activated protein kinase (MAPK) pathway [13]. Loss of the p38 MAPK pathway was shown to result in loss of protection against Cry PFTs in C. elegans and was subsequently shown to result in loss of protection against PFTs in mammalian cells as well [13],[14]. This same system was used to discover that the unfolded protein response (UPR) is also required for PFT defenses as a downstream target of the p38 MAPK pathway [15]. Both the p38 MAPK pathway and the UPR were demonstrated to be activated by PFTs in C. elegans and mammals [15],[16]. Apart from these studies, only one other study to date has demonstrated a specific molecular pathway as involved in PFT responses [17].
Since, when studying intracellular PFT response pathways in the past, we have screened for C. elegans mutants hypersensitive to PFTs [13],[15], we reasoned that we could learn something different by screening for the opposite phenotype– C. elegans mutants resistant to PFTs. The reason for this assumption is that no intracellular pathway mutants were known that can make cells resistant to PFTs in general. Here we report on the results of a PFT resistance screen and find, unexpectedly, that resistance can be achieved by mutations that up-regulate the C. elegans low oxygen (hypoxia) response. Elimination of HIF-1 (hypoxia inducible factor 1), the main effector of the hypoxia pathway, abrogates this resistance and can lead to PFT hypersensitivity. This protection applies to multiple different PFTs and is clearly distinguished from the role of the HIF-1 pathway in other stress responses and aging. Furthermore, the hypoxia pathway is activated in response to PFTs, and low oxygen is itself protective against PFT attack. Our results indicate that the hypoxia/low oxygen response is likely to be of general importance for cellular responses to small-pore PFTs.
To identify pathways important for cellular responses to PFTs, we screened for mutants resistant to the PFT Cry protein, Cry21A. Cry21A is a three-domain Cry protein that targets nematodes and is in the same family as Cry5B [11]. Like Cry5B [18], secondary structure programs predict Cry21A contains all the α helical segments that are involved in pore-formation in three-domain Cry proteins [19]. All three-domain Cry proteins, like Cry5B and Cry21A, are believed to act as PFTs, and pore-forming activity has been demonstrated for all Cry proteins so studied to date [12]. In the past, we have screened for mutants resistant to Cry5B, which has given rise to detailed understanding of the Cry5B receptor [20],[21],[22] but not to information on intracellular responses to Cry PFTs. Our rationale for screening for Cry21A PFT resistant animals was that it could elucidate new information about how cells respond to PFTs since Cry5B resistant mutants are only weakly resistant to Cry21A.
To perform this screen, C. elegans hermaphrodites were mutagenized with EMS and allowed to self-fertilize for two generations. Sixty eight thousand F2 mutagenized hermaphrodites were fed an intoxicating dose Cry21A PFT and then screened for those that resisted intoxication. One mutant line, allele ye49, was identified that bred true and is resistant to Cry21A PFT produced either from Bt or E. coli (Figures 1A and 1B).
To identify the gene mutated in ye49, we performed standard three-factor and single-nucleotide mapping experiments, which narrowed the search to a region on chromosome V, between markers snp_F15H10 and snp_T21C9, that includes 43 genes. Mutation in one of the genes in this region, egl-9, had been previously identified as resistant to cyanide produced by Pseudomonas aeruginosa PA01 [23]. We therefore performed complementation testing between ye49 and the egl-9 null allele egl-9(sa307) and found that ye49/egl-9(sa307) animals are resistant to Cry21A PFT, indicating ye49 fails to complement egl-9(sa307) and most probably mutates the same gene (Figure 1C). Furthermore, injection of an extrachromosomal array bearing the egl-9 promoter and coding region restored wild-type Cry21A susceptibility to ye49 animals (Figure 1D). In addition, sequencing of egl-9 cDNA isolated from the ye49 mutant identified a point mutation (W508-to-stop) that upon translation is predicted to truncate the protein in the prolyl hydroxylase domain, thereby eliminating protein hydroxylase function (Figure 1E). These results demonstrate that Cry21A PFT resistance phenotype associated with ye49 is due to loss of egl-9 function mutation. As predicted from this conclusion, feeding of double-stranded RNA to cause RNA interference (RNAi) results in animals resistant to Cry21A (Figure 1F).
The EGL-9 protein is a prolyl hydroxylase and functions in the C. elegans low oxygen response (hypoxia) pathway (Figure 2A; [24]). The ability of cells to respond to hypoxia is mediated by a transcription factor called HIF-1α. Under normal oxygen (normoxia) conditions, HIF-1α (called HIF-1 in C. elegans) is hydroxylated by a prolyl hydroxylase (EGL-9 in C. elegans or PHD in mammals) that then increases HIF-1's affinity for von Hippel-Lindau tumor suppressor protein (called VHL-1 in C. elegans), part of an E3 ubiquitin ligase complex. Association of HIF-1 with VHL-1 eventually leads to HIF proteasomal degradation. When EGL-9 is disabled by mutation, HIF-1 is stabilized at constitutively high levels even under normoxic (normal oxygen) conditions [25].
Since loss of EGL-9 function confers resistance to Cry21A PFT, we hypothesized that other elements of the hypoxia pathway might be important as well. We therefore performed quantitative dose-dependent mortality assays using null or putative null alleles of all the above elements of the hypoxia pathway. L4 staged animals from each genotype and wild-type N2 were placed in numerous doses of Cry21A PFT or Cry5B PFT and scored for viability after a few days (Figures 2B and 2C).
As predicted from the above studies, we find that animals lacking EGL-9 are quantitatively resistant to Cry PFTs. At doses from 1–16 µg/mL Cry21A and 10–80 µg/mL Cry5B PFTs, egl-9(sa307) and egl-9(ye49) animals are resistant to PFT-induced mortality relative to wild-type animals, with resistance strongest at higher PFT doses (Figures 2B and 2C; Table 1). For example, 7×–10× more egl-9 mutant animals are alive at 8 µg/mL Cry21A or 40 µg/mL Cry5B PFT than wild-type animals at the same doses (P<0.001; Table 1; note, direct dose comparison between Cry21A and Cry5B toxicity is not possible since Cry21A assays are performed with Bt spore-crystal lysates and Cry5B assays with purified protein). Based on LC50 values (concentration at which 50% of the animals are dead), loss of EGL-9 results in 3–5 fold resistance over wild-type animals to Cry21A or Cry5B PFTs (Table 1). Note, since all our mortality assays are carried out in liquid medium, resistance to the PFT cannot be attributed to improved avoidance behaviors. Cry21A PFT resistance was also confirmed using a quantitative brood size assay (Figure S1).
We also found that vhl-1(ok161) mutant animals are resistant over a similarly wide range of Cry21A and Cry5B PFT doses (Figures 2B and 2C; Table 1). For example, 6.2× and 7.4× more vhl-1 mutant animals are alive at 8 µg/mL Cry21A and 40 µg/mL Cry5B, respectively, than wild-type animals. Based on LC50 values, vhl-1 mutant animals are 4× resistant to Cry5B PFT. We also tested rhy-1(ok161) mutant animals on Cry21A PFT. RHY-1 (regulator of hypoxia-inducible factor) antagonizes HIF-1 function by inhibiting expression of some HIF-1 target genes via a VHL-1 independent pathway [26]. Animals lacking RHY-1 are also resistant to Cry21A (Table 1; Figure S2). Based on LC50 values, animals lacking RHY-1 are 5.7× resistant to Cry21A PFT (Table 1). These data demonstrate that loss of function mutations in genes that normally antagonize HIF-1 function all result in resistance to Cry protein PFTs. In other words, stimulation of HIF-1 function via removal of HIF-1 inhibitory factors results in PFT resistance.
To confirm that the resistance associated with egl-9 mutants was going through HIF-1, we looked at the dose-dependent response of hif-1(ia04) and egl-9(sa307) hif-1(ia04) double mutant animals to Cry21A and Cry5B PFTs. When fed Cry21A, hif-1(ia04) animals have a response that is indistinguishable from wild-type animals (Figure 2B; Table 1). egl-9(sa307) hif-1(ia04) double mutant animals have a statistically identical response to Cry21A as wild-type animals at all doses except at 8 µg/mL (Table 1). Furthermore the LC50 values of hif-1(ia04) and egl-9(sa307) hif-1(ia04) animals on Cry21A PFT are statistically identical (P>0.05) but both statistically different than that of egl-9(sa307) (P<0.01; Table 1). These results have been qualitatively confirmed using RNAi—whereas wild-type animals subject to egl-9 RNAi are resistant to Cry21A, hif-1(ia04) mutant animals subject to egl-9 RNAi are not (Figure S3). Similarly, whereas egl-9(sa307) mutants are resistant to Cry21A, this resistance is suppressed by RNAi of hif-1. Thus, HIF-1 is required for Cry21A resistance mediated by loss of EGL-9.
The results with Cry5B PFT are similar to those of Cry21A (Figure 2C; Table 1) in that HIF-1 is required for Cry5B resistance mediated by loss of EGL-9 (i.e., loss of HIF-1 suppresses Cry5B PFT resistance associated with loss of EGL-9). There is, however, one striking difference between the hif-1 results with Cry21A and Cry5B. Both hif-1(ia04) and egl-9(sa307) hif-1(ia04) animals are hypersensitive to Cry5B PFT relative to wild-type animals. That is, animals lacking HIF-1 are more readily killed by Cry5B PFT than wild-type animals, especially at doses ∼5–10 µg/mL (P<0.05; Figure 2C; Table 1). Thus, hif-1 is required for intrinsic cellular defenses (INCED) [15] against Cry5B PFT. With regards to the different results with Cry5B and Cry21A, we speculate that perhaps Cry5B PFT intoxicates more potently than Cry21A and that, whereas increased HIF-1 activity is protective against all PFTs, loss of HIF-1 activity is more acutely felt when the stronger PFT is present. In the case of Cry21A, other INCED pathways are sufficient for full protection even in the absence of HIF-1.
Cry proteins are small-pore PFTs. To test whether or not the hypoxia pathway was more generally required for INCED against PFTs, we fed C. elegans two V. cholerae strains that differ primarily in their ability to produce another small-pore PFT, VCC. VCC is a virulence factor of V. cholerae and mutants lacking VCC are attenuated for pathogenesis in vivo, especially for strains lacking cholera toxin [27],[28]. The strains we use are CVD109(VCC+) and CVD110(VCC−) that are nearly isogenic (except for the presence of cholera toxin B subunit in CVD110, which should not matter since C. elegans lacks sialic acid that the B subunit binds to as part of its GM1 receptor [29]). Although both strains are pathogenic when fed to C. elegans, CVD109(VCC+) is more virulent than CVD110(VCC−), demonstrating that VCC is a virulence factor for C. elegans (Figure 3A).
Our results with hypoxia pathway mutants on CVD109(VCC+) and CVD110(VCC−) are striking and parallel those with Cry PFTs. When feeding on CVD109(VCC+), egl-9(sa307) animals are resistant relative to wild-type animals (Figure 3A; Table 2; median survival 4 vs. 3 days respectively; P<0.001). This resistance is dependent upon the presence of VCC since when feeding on CVD110(VCC−), egl-9(sa307) animals are not resistant (Figure 3A; Table 2). Similarly, hif-1(ia04) and egl-9(sa307) hif-1(ia04) animals are, as with Cry5B PFT, hypersensitive relative to wild-type animals on CVD109(VCC+) (median survival of 2, 1, and 3 days respectively; P<0.0001; Figure 3B; Table 2). This hypersensitivity is dependent upon the presence of VCC since these mutants are not hypersensitive when feeding on CVD110(VCC−) (Figure 3B; Table 2). It is interesting to note that egl-9(sa307) mutant animals are hypersensitive compared to wild-type animals to CVD110(VCC−) strain (median survival of 4 and 6 days respectively; P<0.0001; Figure 3A; Table 2). We speculate that while activation of the hypoxia pathway (in an egl-9 mutant or otherwise) protects the animals against VCC and PFTs (hence egl-9 mutants are resistant to the VCC+ strain), activation of the hypoxia pathway may make the animals more susceptible to other V. cholerae virulence factors. The relative contribution to these responses (protection versus susceptibility) is dependent upon which virulence factors are present and their relative contribution to virulence. In the VCC+ strain, the PFT has important function. Hence, the protective role of pathway activation can be discerned. In the VCC− strain, the PFT defense is no longer needed. Hence, the susceptible role can be discerned. It is this give-and-take interaction between the host and virulence factors that could partly explain why constitutive mutation in egl-9 is not selected in the wild. In any event, taken together, our Cry PFT and VCC data demonstrate that stabilization of HIF-1 results in resistance to VCC PFT whereas loss of HIF-1 results in hypersensitivity to VCC PFT.
Because loss of EGL-9 results in resistance to PFTs (here) and cyanide [23],[30], we hypothesized that egl-9 mutant animals might show resistance to other stressors as well. We found that, relative to wild-type animals, animals lacking EGL-9 are resistant to killing by 1) the pathogen Pseudomonas aeruginosa PA14; 2) heat stress; and 3) oxidative stress (Figure 4A, Table 2; Figures 4B and 4C). Since correlation between stress response and lifespan had previously been reported, such as in the daf-2 mutant [31],[32], we tested whether loss of EGL-9 had an effect on longevity. Indeed, egl-9(ye49) and egl-9(sa307) mutant animals live longer than N2 wild-type when feeding on the standard E. coli strain (Figure 4D, Table 2).
To study the relationship between the hypoxia response pathway and resistance to stresses in more detail, we asked if the resistance to these different stresses via loss of EGL-9 was, as for resistance to PFTs, mediated through HIF-1. Unexpectedly, we found that hif-1(ia04) loss-of-function mutant animals as well as egl-9(sa307) hif-1(ia04) mutant animals are resistant to P. aeruginosa PA14 infection, heat stress, and oxidative stress (Figure 4E, Table 2; Figures 4F and 4G). Both mutant strains are also long lived (Figure 4H; Table 2). Thus, in the case of these stresses, but unlike that of PFT response, loss of either EGL-9, HIF-1, or both results in stress resistance. We speculate that, in the case of these other stresses, hydroxylation of HIF-1 by EGL-9 may result in its activation prior to degradation. Similar results have been previously reported in that mutation of either hif-1 or egl-9 results in C. elegans resistant to pathogenic E. coli [33]. With regards to lifespan, published studies are contradictory but there is at least one published report with egl-9 mutants long lived and two with hif-1 long-lived [34],[35],[36]. In any event, our results demonstrate that role of the hypoxia pathway in PFT INCED is separable from that of other stress responses.
Bt Cry PFTs attack intestinal cells [21],[22],[37]. It is possible that the hypoxia defense pathway functions within the cells targeted by the PFTs or that the hypoxia pathway is functioning cell non-autonomously. To address this question, we expressed egl-9 under the control of various promoters including the intestinal specific cpr-1 promoter [21],[22],[38] and the unc-31 promoter, which is expressed in all neurons and in secretory cells of the somatic gonad [39]. We find that expression of wild-type EGL-9 under the cpr-1 promoter in the intestinal cells of egl-9(sa307) animals (Figure 5), but not under the unc-31 promoter in the neuronal or secretory cells (not shown), is sufficient to rescue the egl-9(sa307) Cry21A resistance phenotype. Control animals in which green-fluorescent protein (GFP) was expressed from the cpr-1 promoter did not result in rescue. Quantitative mortality assays using two independent lines of cpr-1::egl-9-transformed egl-9(sa307) mutant animals confirm that intestinal-specific expression of EGL-9 rescues Cry21A PFT resistance to a level statistically indistinguishable from N2 wild-type (not shown). These data are consistent with the hypoxia pathway acting to directly counteract the effects of PFTs and not, for example, providing protection via altered behavior.
To address how the hypoxia pathway might function in protection against PFTs, we sought in two ways to find functional downstream effectors of the pathway. First, we compared known functional targets of the hypoxia pathway in C. elegans and asked if any of these are involved in PFT defenses. One pathway immediately surfaced, the unfolded protein response or UPR [40]. It has been recently reported that the hypoxia pathway genetically functions upstream of the XBP-1 arm of the UPR with regards to longevity in C. elegans [35]. Furthermore, we have previously shown that the XBP-1 is required for PFT INCED since loss of XBP-1 leads to animals that are hypersensitive to Cry5B PFT [15]. These data suggest that the XBP-1 arm of the UPR is one downstream target of the hypoxia PFT INCED.
To test this suggestion, we examined whether or not the hypoxia pathway regulates activation of the XBP-1 UPR pathway. Activation of the XBP-1 UPR pathway can readily be discerned by examining xbp-1 mRNA, which is spliced upon activation of the pathway [41]. We indeed find that activation of the hypoxia pathway results in activation of the UPR as seen by a 1.4 fold increase in spliced xbp-1 levels in egl-9 mutant animals (P<0.001; see Materials and Methods). Thus, one functional downstream effector of the hypoxia pathway for PFT defenses is the XBP-1 UPR.
We conversely asked if any of the genes known to be involved in PFT INCED are known to be important for the hypoxia pathway. From over 100 PFT INCED genes we have identified in our lab, we found one and only one currently known to be regulated by the hypoxia pathway, nhr-57. nhr-57 was initially identified as part of the hypoxia pathway by the fact that its expression is positively regulated by hif-1 and negatively regulated by egl-9 and vhl-1 [42],[43]. In fact, nhr-57 transcriptional activation is considered the most reliable marker for activated HIF-1 function in C. elegans [26]. We confirmed using quantitative PCR that in egl-9 mutant animals, nhr-57 transcripts are induced 15 fold and that this increase is completely dependent upon HIF-1 (data not shown). However, to date no functional role of nhr-57 for any HIF-1-regulated pathway has been shown.
We find that knock down of nhr-57 results in animals slightly but statistically hypersensitive to Cry5B PFT (e.g., 21% reduction in viability for nhr-57 RNAi at 20 µg/mL Cry5B PFT versus vector-only RNAi control, P = 0.02; n = 90; see Materials and Methods) and therefore defective in PFT INCED. More impressively, we find that knock down of nhr-57 completely suppresses the resistance to Cry21A PFT associated with loss of EGL-9 (Figure 6). Taken together, these results indicate that the nuclear receptor nhr-57 is a second functional downstream effector of the hypoxia PFT defense pathway.
Although the above data demonstrate the hypoxia pathway is important for PFT INCED, they do not directly address whether the defense against PFTs is related to a low oxygen response or to some other function of the HIF-1 pathway. We therefore examined whether the hypoxia pathway itself is activated by PFTs using nhr-57 expression, the canonical marker for HIF-1 pathway activation by low oxygen in C. elegans (see above). We find that 4 and 8 hours of treatment with PFT significantly induces nhr-57 expression 5.3 and 3.6 fold respectively (Figure 7A). Shorter treatments with PFT do not. Thus, PFT induces the hypoxia pathway.
If a low oxygen response is involved in responding to PFTs, then one might predict that exposure to low oxygen might confer protection against PFT attack since the low oxygen environment might strongly and rapidly induce the correct protective response. We therefore exposed C. elegans hermaphrodites to low (2%) oxygen levels minus or plus the presence of E. coli-expressing Cry5B PFT. We find that low oxygen is indeed protective against PFT intoxication in that animals exposed to PFT in a low oxygen environment for 24 hours are significantly healthier than animals exposed to PFT in normoxia (Figure 7B). Similar results were obtained for animals exposed to a low oxygen environment for three days (Figure S4). In contrast and as expected, hif-1(ia04) mutant animals exposed to Cry5B PFT do not get any protection when placed in a hypoxic environment (Figure 7B), confirming that the protective effect of hypoxia against PFT is due to activation of the HIF-1 pathway.
Our results demonstrate that the hypoxia pathway protects C. elegans against PFTs, whether Bt Cry protein PFTs or a PFT used by a mammalian pathogen, V. cholerae VCC. We find that activation of HIF-1 pathway by removal of any of EGL-9/PHD, VHL-1, or RHY-1, makes C. elegans more resistant to PFTs than they normally are. This resistance is completely abrogated upon loss of HIF-1, which can additionally result in animals hypersensitive in PFTs. Resistance to PFTs functions in the cells directly targeted by PFTs and is not associated with other hypoxia-mediated stress resistance phenotypes. Furthermore, exposure to PFT induces transcriptional activation of the HIF-1 low oxygen pathway, and exposure of animals to low oxygen protects animals against PFT intoxication, through a HIF-1-dependent mechanism. A schematic summarizing our findings here is in Figure 7C. Consistent with our finding that activation of the HIF-1 pathway is protective against PFTs, it has been shown that expression of the HIF-1α protein is increased in human airway cells by S. aureus supernatants, of which α-toxin is a major constituent [44].
The simplest interpretation of our data is that PFT intoxication is associated with low oxygen in cells, and that the hypoxia pathway is therefore needed to protect the cells against this condition. Alternatively, although less parsimoniously, it is possible that both hypoxia and PFTs trigger the same set of HIF-1 downstream mediators that are protective against both assaults but that are not otherwise linked by the presence of low oxygen. Two downstream effectors of the hypoxia PFT INCED pathway are the UPR and nhr-57. The fact that nhr-57 is involved in hypoxia PFT INCED suggests that multiple transcriptional responses are key to mounting an effective defense against PFTs. The link between the XBP-1 UPR, hypoxia, and PFTs is intriguing. It has already been shown in mammalian cells that hypoxia induces activation of the XBP-1 UPR as detected by an up-regulation in xbp-1 mRNA splicing by low oxygen [45]. Furthermore, it has been shown that XBP-1 protects cells against hypoxia-induced apoptosis [45]. Therefore, we speculate that one role of the hypoxia pathway in PFT INCED is to induce an XBP-1-linked protective response against PFT/low oxygen-mediated apoptosis. Given that the p38 MAPK pathway is also linked to PFT defenses and the UPR, it will be interesting to explore further links between hypoxia, the UPR, p38 and apoptotic pathways in response to PFTs.
Although this report is not the first of hypoxia pathway involvement in immunity, it is the first showing a link between hypoxia and protection of cells that are being directly attacked by a virulence factor. Control of the metabolic shift to glycolysis by HIF-1α has been shown to play an important role in myeloid cell-mediated inflammatory response [46]. Furthermore, it has been shown that bacteria increase HIF-1α protein expression and stimulate HIF-1α transcriptional activity in macrophages, regulating the expression of immune effectors molecules, including antimicrobial peptides, nitric oxide and tumor necrosis factor-α [47]. Our results point to a new and different role of the hypoxia pathway, namely in providing autonomous protection of epithelial cells against PFTs.
To our knowledge, these results are the first to demonstrate that an intracellular pathway can be altered to promote general resistance to PFTs. Although a few receptor mutants that confer resistance to PFTs have been previously identified, these do not confer general resistance. A logical extension of our findings is that significant therapeutic benefit against a wide range of bacterial pathogens such as S. aureus, Streptococci, Clostrida, V. cholerae (all of which use PFTs as virulence factors) could be achieved by up-regulation of HIF-1 and/or by hypoxia. The identification of the hypoxia pathway as an important PFT INCED pathway thus unexpectedly provides a novel and potentially powerful means of protecting against the single most common mode by which bacterial pathogens attack us.
Strains were maintained at 20°C under standard conditions [48]. The wild-type strain for this study is N2 Bristol [48]. The strains egl-9(sa307), hif-1(ia04), egl-9(sa307) hif-1(ia04), vhl-1(ok161), rhy-1(ok1402), the Hawaiian strain CB4856 and HT1593 [unc-119 (ed3)] were obtained from the Caenorhabditis Genetic Center (CGC). All strains were either previously outcrossed or outcrossed here at least six times (e.g., egl-9(ye49), rhy-1(ok1402)). egl-9(sa307) is a null allele of egl-9 that carries an internal 243-bp deletion removing part of exons 5 and 6 [23]. hif-1(ia04) allele removes exons 2, 3 and 4 of hif-1, including the DNA binding domain, and is believed to be a null allele [49]. The vhl-1(ok161) allele removes exons 1 and 2 of vhl-1 and is believed to be a null allele [25]. The rhy-1(ok1402) allele deletes exons 2, 3 and 4 of rhy-1 and is also believed to be null [26]. Images were acquired using an Olympus SZ60 dissecting microscope and a Canon PowerShot A620 digital camera.
For production in Bt, the cry21A gene was cloned under 700 bp of the cry6A promoter region and subcloned into the Bt/E. coli shuttle vector pHT3101. The plasmid was transformed into a nontoxic host Bt strain (4D22). Cry21A SCLs were prepared using standard procedures [50] and the concentration was measured relative to BSA standards on protein gels.
Mutagenesis and selection of Cry21A resistance mutants was carried out as described for Cry5B [20] except Cry21A SCLs were used to a final concentration of 0.25 µg/mL Cry21A. The 68,000 F2 animals were taken from a larger population of 1,300,000 F2 animals that came from 240,000 mutagenized F1 animals. Animals were incubated for 72 hours at 20°C and scored for overall health, including color, size, movement and brood size. Clonal lines were established from candidates and retested.
Complementation tests were performed by testing F1 progeny from the cross between egl-9(sa307) males and dpy-17(e164);ye49 hermaphrodites. As a control, cross-progeny from egl-9(sa307) males into dpy-17(e164) and from N2 males into dpy-17(e164);ye49 were also tested. ye49 was mapped between dpy-11(e224) and unc-76(e911) using standard three-factor mapping. A dpy-11(e224) ye49 unc-76(e911) triple mutant was then made in order to perform single nucleotide polymorphism mapping with the Hawaiian strain (CB4856) [51].
Genomic DNA and cDNA prepared from egl-9(ye49) animals were used to sequence the egl-9 gene. For transformation rescue, a 13.4kb-PCR fragment covering from 3kb upstream to 2kb downstream of egl-9 transcript was amplified with primers GAGCAACTCGTGGGTTTGTT and CTTCCAGAGGCGTTGTCTTC using the LongAmp Taq (Biolabs) from N2 genomic DNA and injected in egl-9(ye49) worms as described [52].
For tissue-specific rescue, egl-9 rescuing plasmids were constructed by PCR amplification of unc-31 and cpr-1 promoters and then fused to egl-9 and gfp open reading frames using the Multisite Gateway cloning system (Invitrogen) and pCG150 (containing unc-119 rescuing fragment) [53]. The constructs were verified by sequencing and integrated into HY0843 [unc-119(ed3);egl-9(sa307)] by ballistic bombardment [54] with a PDS-1000/He Biolistic Particle Delivery System (Bio-Rad, Hercules, CA). Two independent lines of each transgenic strain were examined.
For Cry21A E. coli toxin assays, we used E. coli JM103 with pQE9 empty vector or a cry21A gene insert under control of the lacZ promoter [11]. Since Cry21A is expressed at very high levels by E. coli [11] and too potent for scoring for resistance, we diluted the toxin-expressing bacteria with non-toxin-expressing bacteria at a ratio of 1∶40 for all tests in this study, similar to that previously described for Cry5B studies [13],[15].
Dose-dependent mortality assays with purified Cry5B were performed as described [52]; hermaphrodite viability was scored after 6 days at 25°C. Cry21A SCLs were used for quantitative mortality assays as described above except hermaphrodite viability was scored after 72 hours at 20°C. Each assay was set up with triplicate wells for each concentration of Cry toxin, and each experiment was performed in at least three independent trials. Typically 180 worms were scored for each concentration.
V. cholerae lifespan assay was performed as described [55] except the overnight culture was spread on Brain Heart Infusion (BHI) plates. CVD109 Δ(ctxAB zot ace) and CVD110 Δ(ctxAB zot ace) hlyA::(ctxB mer) Hgr strains, derived from V. cholerae El Tor E7946, were used [56]. The experiment was performed three times with approximately 50 worms per strain at room temperature (22°C). P. aeruginosa lifespan assays were performed on slow-killing plates as described [15]. Heat shock assays were performed as described [57]. For the oxidative stress analysis, synchronized young adults were exposed to 7.5 mM t-butyl hydrogen peroxide as described [58] and were observed after 6 hours. Life-span assays were initiated by allowing adult hermaphrodites to lay eggs on NG plates spread with OP50. When these eggs hatched and the nematodes reached the L4 stage they were transferred to fresh NG plates with OP50 supplemented with 25 µM 5-fluorodeoxyuridine (FUDR) to prevent eggs from hatching. The nematodes were scored for live/dead every 48 hours by tapping the nose at least three times (no movement for all taps was scored as dead).
For RNAi tests, adult hermaphrodites were allowed to lay eggs on NG plates containing 100 µM Isopropyl β-D-1-thiogalactopyranoside (IPTG) and 50 µg/mL ampicillin spread with E. coli strain HT115 expressing double-stranded (ds) RNA (from the Ahringer library [59]) for 8 hours and then removed. The eggs were allowed to develop into L4 larvae on RNAi plates at 20°C. L4 hermaphrodites (ten per genotype or line) were picked onto toxin plates spread with 100 µl of a mixture of E. coli strain HT115 expressing the same dsRNA and HT115 harboring cry21A-expressing vector at the ratio 40∶1. For no toxin control plates, 100 µl of HT115 with dsRNA was spread. nhr-57(RNAi) testing on Cry5B PFT was performed slightly differently (Kao et al., manuscript in preparation). Briefly, rrf-3(pk1426) animals were fed E. coli-expressing dsRNA in liquid media with 1mM IPTG at 25°C for ∼30 h. 20 µg/ml of Cry5B or 20 mM HEPES control were then added, as well as 200 µM FUdR. Hermaphrodites viability was scored after 6 days at 25°C. (As this assay is set up differently, direct comparison with dose-dependent mortality assays presented in Table 1 and associated Figures is not possible).
To test worms under hypoxia, L4 wild type and hif-1(ia04) mutant animals were pipetted onto toxin plates spread with 30 µl of a mixture of E. coli OP50 strains expressing or not Cry5B at the ratio 1∶1. Plates were placed immediately in a 2% O2 chamber for 24 hours, while control plates were placed in room air. Images were taken with an Olympus BX60 microscope as described [15].
Real time PCR was performed as described [15]. To determine the levels of spliced xbp-1 mRNA, we used primers xbp-1_sqf2 GCATGCATCTACCAGAACGTC and xbp-1_sqr2 GTTCCCACTGCTGATTCAAAG to amplify cDNA from wild-type and egl-9(sa307) animals. The forward primer xbp-1_sqf2 anneals to exon 1 and the reverse primer xbp-1_sqr2 anneals to the exon1-exon 2 junction sequence produced when intron 1 is spliced out. The experiment was carried out using two independent sets of cDNA and two repeats within each set. Primers TTATCGAGTTTCTCGCATTGG and AAGTCTGCAATCACGCTCTGT were used to quantify expression of nhr-57. Induction of expression of nhr-57 by Cry5B was tested in glp-4(bn2) animals treated for 1, 2, 4 and 8 hours on E. coli OP50 strains expressing Cry5B or not. The experiment was carried out using three independent sets of cDNA. Normalization in all cases was to eft-2 transcript levels.
LC50 values were determined by PROBIT analysis [60]. Mortality assays were plotted using GraphPad Prism 5.0 (San Diego). Statistical analysis between two values was compared with a paired t-test. Statistical analysis among three or more values of one independent variable was done with matched one-way ANOVA with Tukey's method and of more than two independent variables by two-way ANOVA with the Bonferroni post test. For lifespan analysis, survival fractions were calculated using the Kaplan-Meier method and survival curves compared using the logrank test. Statistical significance was set at P<0.05.
|
10.1371/journal.pbio.1001422 | Recombination Modulates How Selection Affects Linked Sites in Drosophila | One of the most influential observations in molecular evolution has been a strong association between local recombination rate and nucleotide polymorphisms across the genome. This is interpreted as evidence for ubiquitous natural selection. The alternative explanation, that recombination is mutagenic, has been rejected by the absence of a similar association between local recombination rate and nucleotide divergence between species. However, many recent studies show that recombination rates are often very different even in closely related species, questioning whether an association between recombination rate and divergence between species has been tested satisfactorily. To circumvent this problem, we directly surveyed recombination across approximately 43% of the D. pseudoobscura physical genome in two separate recombination maps and 31% of the D. miranda physical genome, and we identified both global and local differences in recombination rate between these two closely related species. Using only regions with conserved recombination rates between and within species and accounting for multiple covariates, our data support the conclusion that recombination is positively related to diversity because recombination modulates Hill–Robertson effects in the genome and not because recombination is predominately mutagenic. Finally, we find evidence for dips in diversity around nonsynonymous substitutions. We infer that at least some of this reduction in diversity resulted from selective sweeps and examine these dips in the context of recombination rate.
| Individuals within a species differ in the DNA sequences of their genes. This sequence variation affects how well individuals survive or reproduce and is transmitted to their offspring. Genes near each other on individual chromosomes tend to be passed to offspring together—neighboring genes are unlikely to be separated by exchanges of genetic material derived from different parents during meiotic recombination. When genes are inherited together, however, the evolutionary forces acting on one gene can interfere with variation at its neighbors. Thus, variation at multiple genes can be lost if natural selection acts on one gene in close proximity. Recombination can prevent or reduce this loss of variation, but previous tests of this phenomenon failed to account for recombination rate differences between species. In this study, we show that some parts of the genome differ in recombination rate between two species of fruit fly, Drosophila pseudoobscura and D. miranda. Avoiding an assumption made in previous studies, we then examine sequence variation within and between fly species in those parts of the genome that have conserved recombination rates. Based on the results, we conclude that recombination indeed preserves variation within species that would otherwise have been eliminated by natural selection.
| Homologous meiotic recombination has an important role in molecular evolution. Sufficient recombination uncouples the evolution of different sites on the same chromosome allowing positive or negative selection at one site to act independently from selection at another site. If there is less than effectively free recombination between two selected sites, then linkage results in selection at one site interfering with selection at another site. This has been termed “Hill–Roberson interference” [1]–[6]. Hill–Robertson interference increases the probability of fixation of deleterious mutations, decreases the probability of fixation of advantageous mutations, and reduces overall DNA sequence diversity. Thus, the breakdown of linkage disequilibrium between loci experiencing Hill–Robertson interference allows selection to act more efficiently, purging deleterious mutations and accelerating adaptation [1]–[6].
Such indirect effects of recombination on the genome [7] result in a positive association between the rate of recombination and adaptive evolution [8]–[10]. For example, recombination rate is positively associated with codon usage bias, whereby those codons coded by the most abundant tRNAs are “preferred” and used more often [11],[12]. Recombination has direct effects on a genome sequence as well, because recombination influences base composition through biased gene conversion and the distribution of repetitive elements, hotspot sequences, and indels [7],[13]–[17]. Understanding the magnitude of indirect effects in light of these direct effects has proved challenging [7].
One striking association is a positive relationship of local recombination rate and nucleotide diversity [13],[18],[19]. Originally described in Drosophila melanogaster [13], the positive relationship between recombination rate and nucleotide diversity has been demonstrated in a wide range of taxa, including humans, mice, yeast, maize, and tomatoes (reviewed in [20]). It is not fully understood how much of this relationship results from recombination's indirect versus direct effects on the genome. For instance, mutations created during crossing over or double-strand break repair may generate new polymorphisms and hence increase diversity [21]–[27]. Alternatively, recombination may indirectly influence genetic diversity by mitigating the genomic footprint of selective sweeps and background selection [28]–[30].
One way to distinguish between these general explanations is to evaluate the relationship of between-species nucleotide divergence at neutral sites and local recombination rate, because truly neutral mutations are substituted at the same average rate between species as they appear between generations, even if linked to sites under selection [31],[32]. This allows us to predict that both within-species nucleotide diversity and between-species nucleotide divergence would have a positive relationship with local recombination rate [13], if the recombination–diversity association was purely caused by mutation. In contrast, selective sweeps and background selection will cause an association between recombination and within-species nucleotide diversity, but not a relationship between recombination and between-species nucleotide divergence [30],[32]. The absence of an association of between-species nucleotide divergence and local recombination rate suggests that variation in recombination rate translates to variation in the efficiency of selection [13]. Past work relating nucleotide divergence to recombination rate found no relationship between these two variables in several species of Drosophila, mouse, beet, yeast, and other species [13],[20],[33]–[37]. Furthermore, in several species, evidence indicates that segregating ancestral polymorphisms may be responsible for correlations between divergence and recombination rate ([38]–[40], also suggested by [25],[41]).
The test above, however, implicitly assumes that local recombination rates are conserved between the two species used to generate the nucleotide divergence measure. If recombination rate has diverged between the two species, no relationship between local recombination rate and nucleotide divergence may be detected even when recombination is mutagenic (see Figure S1). Recombination rates, especially at fine scales, are often not conserved among closely related species, as is the case between humans and chimpanzees [42]–[44]; thus, the assumption of conservation of recombination rates may be violated in previous studies, and a more definitive understanding of the diversity–recombination association awaits estimates that are free from this assumption.
Though there are theoretical expectations concerning how recombination rate should affect selection efficiency [45],[46], it is unclear empirically whether variation in local recombination rates translates into significant variation in the efficiency of selection [7]. Several empirical studies have tackled this problem [12],[38],[47]–[52], and many findings suggest that recombination rate influences the efficiency of positive or negative selection in regions of moderate or high recombination. Still, various confounding factors (e.g., biased gene conversion, gene density) may produce spurious correlations between both recombination and substitution rate, and some authors suggest that there is no strong empirical evidence for recombination affecting the efficiency of selection (apart from reduced selection in regions with essentially no recombination [7]).
The Drosophila pseudoobscura system is ideal for pursuing questions about recombination rate variation and its molecular evolutionary consequences. The average crossover rate of D. pseudoobscura (about 7 cM/Mb in females) is over twice that of D. melanogaster [53]. There is also considerable fine-scale (<200 kb windows) variation in the local recombination rate within the genome of D. pseudoobscura and within the genome of its sister species, D. persimilis [25],[33],[54]. While some recombination data are available for D. pseudoobscura and D. persimilis, these sister taxa interbreed in the wild [55]–[57] and are, therefore, not ideal for examining the divergence–recombination association. For example, shared polymorphism due to hybridization and recent speciation may be responsible for the positive divergence–recombination association found in a previous study [25] (see also [38],[39]). Fortunately, a third species exists (D. miranda) that is phylogenetically close to D. pseudoobscura but does not interbreed with D. pseudoobscura. Since there is still some residual shared ancestral polymorphism [58], we also obtained the genome sequence for a slightly more distantly related outgroup species, D. lowei (Figure S2). Sequence from D. lowei is useful for generating a proxy for neutral mutation rate across the genome.
In this work, we generate and compare two fine-scale recombination maps for D. pseudoobscura, which each cover approximately 43% of the D. pseudoobscura physical genome and one fine-scale recombination map that covers approximately 31% of the D. miranda physical genome. In order to circumvent the assumption of classic studies, we analyze the relationship of local recombination rate to nucleotide diversity and divergence in regions with very similar recombination rates between the two species. By employing a linear model framework to account for multiple covariates, we conclude that the contribution of recombination to diversity is significant and positive, but recombination contributes little to divergence. This indicates that recombination is likely to modulate the footprint of selection in the genome. Next, we tested the impact of recombination rate on the efficiency of selection. We examined whether recombination rate (1) affects the distribution of nonsynonymous substitutions across the genome and (2) affects the pattern of diversity around nonsynonymous and synonymous substitutions. In particular, we use a generalized linear model to test how recombination modulates the magnitude and physical extent of the loss of diversity surrounding substitutions. Our analysis of these putative selective sweeps should be less sensitive to common confounding factors such as gene expression and GC content than previous measures. In total, this work allowed us to determine that recombination rate has an important impact on how selection shapes diversity across the genome of Drosophila pseudoobscura and its close relatives.
We first discuss general features of the recombination landscapes we observed in Drosophila pseudoobscura and D. miranda before we address the implications of these observations for understanding diversity, divergence, and the nature of selection in the genomes we sequenced.
We generated linkage maps for chromosome 2 and parts of the X chromosome for D. pseudoobscura and D. miranda. Using a backcross design and inbred lines, we developed two replicate recombination maps (referred to here as “Flagstaff” and “Pikes Peak”) for D. pseudoobscura and one recombination map for D. miranda using the Illumina BeadArray platform to distinguish heterozygotes from homozygotes of the inbred lines used in the backcross design. These maps (Table S1) measure recombination rate across <200 kb windows, and we refer to these as “fine-scale” maps.
Recombination was surveyed across approximately 43% of the D. pseudoobscura physical genome and about 31% of the D. miranda physical genome (Tables S1 and S2). For each of the three maps, nearly the entire assembled region of chromosome 2 (97.8%–99.4%), the majority of the XR chromosome arm (70.8%–89.4%), and part of the XL chromosome arm (∼22%–23%) were surveyed (Table S2). After removal of likely erroneous putative double recombinants, ambiguous genotypes, and markers that did not work or gave inconsistent genotypes, recombination was measured for three different crosses for 1,158–1,404 individuals per map (Table S1). Excluding larger intervals at the telomeres and centromeres, intervals between markers had a median size across the three maps of 141–148 kb for chromosome 2 and 146–160 kb for the XR chromosome arm (Table S1).
For chromosome 2, recombination rates ranged from 0–30.8 cM/Mb in D. pseudoobscura and 0–24.0 cM/Mb in D. miranda (Table S2). The number of individuals surveyed is often slightly different per interval; therefore, for all intervals where no recombination was detected, we report 0 cM/Mb. The recombination rate for those intervals with “0 cM” should be interpreted as <1 recombination event per total number of individuals surveyed for each interval (Dataset S1). Recombination near the telomere and centromere was measured at a broader scale than the remainder of chromosome 2 because we expected these regions to have lower crossover rates than the center of the chromosome (chromosome 2 is telocentric). Because of this limitation, comparisons of recombination rates between the ends of the chromosome and the center are more tentative. Nonetheless, examining recombination across roughly 3 Mb of sequence at the telomeric end and 3 Mb at the centromeric end, we found up to an 8.9-fold difference between the recombination rates at the middle of chromosome 2 relative to the centromeric end. The Pikes Peak D. pseudoobscura map exhibited the largest reduction of recombination at the telomeric or centromeric ends relative to the center of the chromosome for all three maps, though in the Flagstaff D. pseudoobscura map and the D. miranda map, recombination rates were reduced by at least 2.6-fold in the centromere and telomere relative to the center of the chromosome (Table S3).
For the XR chromosome arm, recombination rates ranged from 0–25.2 cM/Mb in D. pseudoobscura and 0–32.3 cM/Mb in D. miranda (Figure S3 presented with 95% confidence intervals; see also Dataset S1, Table S2). The number of crossovers per individual for both chromosome 2 and the XR arm was close to 1 (1.01–1.06) for D. pseudoobscura and was 1.40–1.54 for D. miranda, illustrating that a greater overall recombination rate in D. miranda relative to D. pseudoobscura is observed in both an autosome and a sex chromosome.
The XL chromosome arm was not surveyed as intensively (∼22%–23% of the XL arm in Pikes Peak and D. miranda and ∼60% of the XL arm in Flagstaff; Figure S4 presented with 95% confidence intervals; Dataset S1). The number of crossovers per individual appears consistent with ∼1 crossover per chromosome arm, as in D. pseudoobscura XR and chromosome 2, but the average number of crossovers per individual on the XL reflects how much of the arm was surveyed. For example, when ∼22%–23% of the arm was surveyed, crossovers per individual ranged from 0.23–0.26 (Table S2).
A binomial Generalized Linear Model (GLM) with size of the interval as a covariate and interval identity as a factor in the model indicated significant heterogeneity in recombination rate among intervals for chromosome 2, XR, and XL (each tested separately) for each of the three maps (each tested separately, interval identity p<0.00001, χ2≥64.67, df≥3, in all cases). Furthermore, 95% confidence intervals (generated via the same method in [54]) do not overlap in many cases between different intervals (shown in Figures 1, S3, S4; Dataset S1). Overall, we observe heterogeneity in fine-scale recombination rates within each of the three maps (see Figures 1, S3, and S4 with 95% confidence intervals plotted; Dataset S1; statistical quantification between maps given in section below), and we note a reduction in recombination rate around the telomeric and centromeric ends consistent with other studies in Drosophila [33].
Our three fine-scale crossover maps utilized markers on average 141–160 kb apart (median interval size for each of the three maps, with the exception of XL where the median distance between markers was 200–1,775 kb for the three crosses). We additionally examined three regions on chromosome 2 in more detail. Each of these regions spanned a total of 99–125 kb, and we placed markers every ∼20 kb within the region (16 total intervals; Tables S4 and S5). These regions were originally picked because previous data [25],[33] indicated that recombination rates for each of these regions differed (regions are referred to as 6 Mb, 17 Mb, and 21 Mb, which indicate approximate location on chromosome 2). We refer to these as “ultrafine-scale” maps. For these ultrafine maps, we followed the same backcross scheme as above, and we scored approximately 10,000 individuals for each marker (Table S5). For the 16 ultrafine intervals (Tables S4 and S5), each interval was on average 20.61 kb long (range 12.6–27.4 kb). Recombination rates range from 1.6–21.2 cM/Mb for these ∼20 kb intervals (Figure 2; see Table S5 for 95% CI). The ultrafine-scale map uncovered variation in recombination rates that was not apparent with the fine-scale maps. For example, for the 17 Mb ultrafine-scale region on chromosome 2, the recombination rates for the two fine-scale intervals spanning this region (17.5–17.7 Mb) are 5.6 and 4.4 cM/Mb. The ultrafine-scale recombination rates, in contrast, ranged from 3.5–21.2 cM/Mb (markers spanning 17.5–17.7 Mb). This heterogeneity in recombination rates within the ultrafine regions was statistically significant (binomial GLM similar to that described in fine-scale section above: p = 0.0011, df = 14, χ2 = 35.91; 95% confidence intervals given in Table S5) and highlights the fact that “broader” scale measures of recombination rates (such as the fine-scale measures here) are averages of true variation in recombination rate.
For comparisons of recombination rates between fine-scale maps, we restricted our analysis to intervals that were condensed to have nearly identical physical marker placement between the three fine-scale maps (Figures S5 and S6; Table S6). Recombination was estimated as detailed above, using the number of crossovers spanning the newly defined physical intervals. After condensing across all three maps, 97 intervals remained for chromosome 2 and 44 intervals for XR (see Tables 1 and S6 fornumber of individuals, size, range of these condensed intervals,and base pairs between markers on each map). The XL chromosome arm was not included in the analysis that used condensed intervals across maps because too few intervals overlapped between all three maps. When comparing two maps, intervals were condensed between those two maps only (see Datasets S2 and S3 for rare events logistic regressions for all two-map and three-map comparisons).
Recombination rates did not differ significantly between the two D. pseudoobscura maps for either the XR or chromosome 2 for the two-map comparisons (each chromosome analyzed separately, rare events logistic regression, absolute value of z>0.3901, p>0.236, in both cases; Dataset S2). For chromosome 2, one interval was significantly different in recombination rate after correcting for multiple tests [59]. For the XR, no intervals between the two D. pseudoobscura maps were significantly different in recombination rate after correcting for multiple tests. The 95% confidence intervals for the odds ratio of the difference between maps were narrow and located around zero, indicating that the maps are likely very similar (chromosome 2, 0.87–1.10; XR, 0.94, 1.28; within-species two map comparison). It is unlikely that the single significant difference observed within the same species is because of slight differences in marker placement between the two maps. The marker placement for this interval was nearly identical between the two maps (left marker, 102 nucleotides different between maps; right marker, 17 nucleotides).
For both chromosome 2 and the XR chromosome arm, Drosophila miranda had significantly higher recombination rates than both D. pseudoobscura maps (Figure S5, Table 1, Datasets S2 and S3). A rare events logistic regression of two-map comparisons indicated that the recombination rate of the D. pseudoobscura crosses we surveyed is about 76%–78% of the D. miranda recombination rate we observed on chromosome 2 (absolute z value>4.5374, p<0.001 for D. miranda relative to either D. pseudoobscura map, Table 1). The recombination rate of D. pseudoobscura is about 68%–71% of the D. miranda recombination rate on the XR chromosome arm (rare events logistic regression absolute z value>5.101, p<0.001 for D. miranda relative to either D. pseudoobscura map, Table 1).
After the global difference between D. miranda and D. pseudoobscura is accounted for by the rare events logistic regression, recombination rates within and between species appear very similar for chromosome 2 (Figure S5; Datasets S2 and S3). None of the intervals for the two-map comparison between D. miranda and D. pseudoobscura–Flagstaff were significantly different after correction for multiple tests, though power to detect significant differences on a per interval basis was likely weak (see confidence intervals in Datasets S2 and S3). For example, 15 of the 115 intervals on chromosome 2 showed at least a 3-fold difference in recombination rate between maps (Datasets S2 and S3), though this magnitude of difference was not significant in our rare events logistic regression after correcting for multiple tests. Likewise, only one of the intervals for the two-map comparison between D. miranda and D. pseudoobscura–Pikes Peak was significantly different after correction for multiple tests, but 19 of the 123 intervals exhibited at least a 3-fold difference in recombination rate between maps for chromosome 2.
The XR chromosome exhibited a qualitatively larger difference in recombination rate between species than chromosome 2. After the global difference between D. miranda and D. pseudoobscura is accounted for by a rare events logistic regression, two of the intervals between D. miranda and D. pseudoobscura–Flagstaff for the two-map comparison and seven of the intervals between the D. miranda and D. pseudoobscura–Pikes Peak two-map comparison were significantly different after correction for multiple tests. Six of the 72 intervals between D. miranda and D. pseudoobscura–Flagstaff two-map comparison exhibited at least a 3-fold difference, and 12 of 102 intervals between D. miranda and D. pseudoobscura–Pikes Peak exhibited at least a 3-fold difference (Dataset S2).
Twenty-seven of 97 condensed intervals (three-map comparison, condensed between all three maps) for chromosome 2 were considered to be “conserved” within and between species. This means that they displayed a nonsignificant difference across all three maps when analyzed with a rare events logistic regression and had an odds ratio between 0.62 and 1.615 after the effect of map identity was taken into account. These “conserved” intervals were used for further downstream analyses (see “Diversity, Divergence, and Recombination”; Table S7). For the XR, seven of 44 intervals condensed between all three maps were conserved within and between species according to the criteria outlined above.
In sum, we observe strong conservation in recombination rates within a single species, while between species, we see globally elevated recombination rates in D. miranda. Once the global difference is accounted for, there are few intervals with significant differences in recombination rate within and between species. Thus, it is possible and parsimonious that recombination rate is generally conserved at the scale examined here (∼180 kb) over moderate evolutionary timescales (2–2.5 my).
We used various Illumina platforms to resequence genomic DNA from 10 D. pseudoobscura lines using virgin females from lines that were inbred for five or more generations with full-sibling single-pair mating (Table S8). Drosophila pseudoobscura populations across North America display very little differentiation, as indicated by low FST values (always<0.10, often<0.05 for loci located outside of the inversion polymorphisms of the third chromosome) [60],[61]. Therefore, the choice of strains sequenced for estimating diversity covered much of the species range but was fairly random. We also sequenced two lines of D. persimilis (one of these was provided by S. Nuzhdin), two lines of D. pseudoobscura bogotana (one of these was provided by S. Nuzhdin), one line of D. lowei, and three lines of D. miranda (two provided by D. Bachtrog, Table S8; Short Read Archive accession numbers SRA044960.1, SRA044955.2, and SRA044956.1; see also http://pseudobase.biology.duke.edu/). The divergence between D. persimilis and D. lowei was used to generate measures of a proxy for neutral mutation rate across the genome. In all diversity and divergence calculations, the reference sequences for the D. pseudoobscura and D. persimilis genomes were both included [62],[63]. Details of diversity and divergence calculations are discussed in Text S1 (see section titled “Fine-Scale Recombination Maps: Computational Methods for Diversity and Divergence Measures”). Briefly, average pairwise diversity and divergence was calculated for 4-fold degenerate sites, focusing exclusively on unpreferred codons [64], though we obtained very similar results when using all 4-fold degenerate sites. Overall, recombination is significantly and positively associated with average pairwise diversity but not average pairwise divergence at 4-fold degenerate sites of unpreferred codons. We examined this relationship in several ways.
We analyzed each chromosome for each uncondensed recombination map independently using a generalized linear model for diversity and a separate model for divergence (Tables S9, S10, and S11). After accounting for multiple covariates, diversity at 4-fold degenerate sites of unpreferred codons shows a significant, positive relationship with recombination, while divergence at 4-fold degenerate sites of unpreferred codons does not (Tables S9 and S10). This result is consistent for each of the three recombination maps (D. pseudoobscura–Flagstaff, D. pseudoobscura–Pikes Peak, and D. miranda) for both chromosome 2 and the XR chromosome arm (Tables S9 and S10). The XL chromosome arm contained too few intervals for analysis for D. pseudoobscura–Flagstaff. For D. pseudoobscura–Pikes Peak and D. miranda, diversity showed a significant, or nearly significant, positive relationship with recombination, while divergence did not (Table S11).
The analysis above suggests that the recombination–diversity relationship is probably the result of the effect of recombination on selection at linked sites (sensu [13],[18]); however, inadvertently including regions with discordant recombination rates between species in the analysis above could result in a pattern that supports this hypothesis—even when recombination is predominantly mutagenic (Figure S1). To resolve this potential bias, we restricted analysis to only regions that exhibited conserved recombination rates between all three chromosome 2 maps (N = 27 intervals; described above) and examined recombination in association with average pairwise D. pseudoobscura diversity at 4-fold degenerate sites of unpreferred codons (Table 2; Figures S7 and S8) and average pairwise D. pseudoobscura–D. miranda divergence at 4-fold degenerate sites of unpreferred codons (Table 3; Figures S7 and S8). The effect of recombination on diversity was significant when the analysis was restricted to only those regions with the most conserved recombination rates (quasibinomial GLM, F = 6.123, p value = 0.024), and the effect of recombination on divergence remained nonsignificant (quasibinomial GLM, F = 0.138, p value = 0.714). These regions contained only one interval within 4 Mb of the telomeric end and no intervals within 4 Mb of the centromeric end of the chromosome; thus, these results are not a function of broad-scale regional recombination rate differences across the chromosome. These results support the hypothesis that recombination affects diversity through the effect of selection on linked sites. We did not perform an analysis on conserved windows for the X chromosome, as only seven intervals were conserved within and between species.
To determine the impact of recombination rate on selection at linked sites in the genome, we used two generalized linear models to analyze the relationship of recombination rate and several measures that may be indicative of the efficiency of selection: (1) abundance of nonsynonymous substitutions and (2) average pairwise nucleotide diversity at 4-fold degenerate sites around nonsynonymous substitutions. We analyzed the association of recombination rate with these two measures in a generalized linear model framework to account for covariates such as gene density, GC content, and a proxy for neutral mutation rate. Biased gene conversion may influence substitution rates; thus, we controlled for GC content in all of the analyses below [7],[16],[65],[66]. We did not consider gene expression as a covariate, though some studies point to a negative relationship with recombination rate [67].
The relationship of recombination rate to nonsynonymous substitution abundance was examined with the D. pseudoobscura Flagstaff fine-scale recombination maps. Nonsynonymous substitution abundance was measured as the nonsynonymous substitutions on the branch leading to D. pseudoobscura+D. persimilis as identified with PAML. The response variable was the number of nonsynonymous substitutions in each gene, and the covariates of the linear model included (1) the number of synonymous substitutions in the gene in question allowing for inclusion of genes where Ks = 0 (sensu [50]), (2), GC content of the gene, (3) gene density of 50 kb on either side of the midpoint of the gene, and (4) average pairwise divergence at 4-fold degenerate sites of unpreferred codons between D. persimilis and D. lowei as a proxy for neutral mutation rate within the gene. We found no relationship (Table 4) between recombination and nonsynonymous substitution abundance with the fine-scale data (generalized linear model with Poisson distribution, z = −0.614, p = 0.539).
In response to selective sweeps, a trough in diversity should be visible around selected variants [30],[68]–[72]. We analyzed diversity surrounding the nonsynonymous substitutions along the lineage leading to D. pseudoobscura+D. persimilis identified by PAML. We compared the average pairwise diversity patterns at 4-fold degenerate sites surrounding these substitutions in relation to the Flagstaff recombination rate and distance in basepairs from the substitution (Text S1). In regions with high recombination rates, the footprints of selection are thought to be narrower than in regions with low recombination rates, where strong linkage between sites will create a stronger signature of sweeps [39],[69],[71],[73]. As a control, similar analyses were performed using synonymous substitutions along the D. pseudoobscura+D. persimilis lineage following [68]. Synonymous substitutions, in many cases, evolve in a more neutral fashion than nonsynonymous substitutions ([68], but see [74],[75]). In a recent genome-scale analysis conducted with data similar to what are presented here, little reduction in diversity was seen around synonymous substitutions [68]; this study instead saw an increase in diversity, which disappeared after correction for local mutation rates.
We considered 60 kb on either side of the substitution along the D. pseudoobscura lineage divided into 1,000 bp nonoverlapping windows (sensu [68]). For each 1,000 bp window, the response variable was the number of polymorphic 4-fold degenerate sites. The generalized linear model included the following covariates: (1) total 4-fold degenerate sites, (2) GC content, (3) proportion of coding bases, (4) divergence of D. lowei–D. persimilis at 4-fold degenerate sites as a proxy for neutral mutation rate, and (5) proportion of bases that were nonsynonymous substitutions. The identities of each nonsynonymous substitution were included as random effects. This generalized linear mixed model with Poisson distribution included the following factors: absolute physical distance from the substitution, fine-scale-derived estimates of recombination rate, and the interaction between these two factors. A negative interaction term means that short distances from a substitution and high recombination rates have similar effects on diversity as large distances and low recombination rates. We expect the interaction term for distance and recombination rate to be much reduced in magnitude for synonymous substitutions in comparison to the nonsynonymous analysis.
We found a small but significant negative interaction term of physical distance from the nonsynonymous site and recombination rate on nucleotide diversity around nonsynonymous substitutions (Poisson GLMM, z = −7.52, p<0.001; Table 5, Figures 3 and S9). In other words, higher rates of recombination allow for recovery of diversity at shorter physical distances from the nonsynonymous site than lower recombination rates (Figure S9). In contrast, a weaker interaction was detected for the interaction of distance and recombination rate on diversity around synonymous substitutions along the D. pseudoobscura lineage (Poisson GLMM, z = −2.43, p = 0.015; Table 6, Figures 3 and S9). GLM plots for the very low recombination rates of <0.5 cM/Mb show wider dips in diversity (and more associated noise; Figure S9) than plots for recombination rates of >0.5 cM/Mb (Figure S9).
Distance from a substitution had a positive, significant effect on diversity as expected if linked selection of substitutions generates a dip in diversity (Tables 5, 6, and S12). Recombination rate also had a positive, significant effect on diversity as expected, if either recombination was mutagenic or if positive/negative selection was operating on the chromosome (Tables 5, 6, and S12). The proportion of nonsynonymous substitutions around a substitution had a negative significant effect on diversity surrounding a nonsynonymous site as expected if many of these substitutions combine forces to generate stronger selective sweeps (Tables 5, 6, and S12). The interaction term pointing to deeper dips in diversity for lower recombination rates is no longer significant when examining only 5 kb or 15 kb on either side of the focal substitution (it is negative for nonsynonymous substitutions and positive for synonymous substitutions), but it is conceivable that this lack of significance represents an issue with window size or sampling.
Overall, our study identified both global and local differences in recombination rate between two closely related species of Drosophila. Aside from regions with exceptionally low recombination rates [12],[76], variation in local recombination rates between species must be accounted for prior to concluding that the association between recombination rate and diversity is probably caused by recombination modulating the effects of selection at linked sites [77]. By restricting our analysis in the Drosophila pseudoobscura system to only those regions with conserved recombination rates within and between species, we rejected the hypothesis that recombination rate (at the scale tested) significantly affects divergence at 4-fold degenerate sites for unpreferred codons. These results support the conclusion that recombination has a substantial impact on how selection affects diversity in the genome. Furthermore, additional analyses suggest that recombination rate variation affects the impact of Hill–Robertson effects like selective sweeps and background selection in this system.
Here and in other recent work [54], we demonstrate that ultrafine-scale patterns of crossover rate (intervals spanning 20 kb) are also significantly heterogeneous in D. pseudoobscura. In each ultrafine region on chromosome 2, recombination rates varied by up to 6-fold (17 Mb region) over only approximately 120 kb (6 Mb region variation is 3.6-fold, and 21 Mb region variation is 5.1-fold), and ultrafine-scale maps reveal variation not detected in the fine-scale maps. This was especially apparent for the 17 Mb region, where ultrafine-scale recombination rates ranged from 3.5 to 21.2 cM/Mb, and fine-scale recombination rates in the same area ranged only from 4.4 to 5.6 cM/Mb. This heterogeneity suggests that our fine-scale measures (intervals spanning <200 kb) are averages of actual variation in recombination rate.
In humans, broad-scale variation averages over the density and intensity of ∼2 kb hotspots that occur in clusters every 60–90 kb [78],[79]. The majority of recombination occurs at these hotspots, and the majority of recombination is governed by the DNA binding protein PRDM9 and its recognition motifs in humans [17],[80]–[84]. Interestingly, several studies in different regions of the D. melanogaster genome indicate that linkage disequilibrium decays rapidly [37],[85]–[87], suggesting that the heterogeneity we observed in ultrafine-scale maps may not be governed by clustered hotspots similar to those in humans, or at least that a nontrivial amount of recombination may occur outside such “hotspots.”
To assess whether “hotspots” of some sort exist in D. pseudoobscura, genome-wide patterns of linkage decay need to be investigated or incredibly fine-scale maps (interval size <5 kb) need to be made. Such a line of inquiry would help address basic questions about the requirements for functional recombination across various taxa. For example, there are several notable differences regarding the formation and function of the synaptonemal complex and the role of double-strand breaks across taxa [88]–[93]. Furthermore, the Drosophila lineage completely lacks several proteins essential for generating crossovers and double-strand break repair in other organisms [89],[94]. It is likely that understanding particular sequence features associated with recombination on a kilobase scale in Drosophila will uncover more details about the mechanistic underpinnings of meiosis that differentiate these species and the distribution of crossovers across the genome.
Recombination rates at broad scales are conserved between populations and species [33],[95]–[100] (see also review in [20]). Our fine-scale data are generally consistent with these findings except that D. pseudoobscura has about three-fourths the rate of recombination, on average, as D. miranda for chromosome 2 and about three-fifths the rate of recombination of D. miranda on the XR chromosome arm. Notably, D. melanogaster has one of the lowest recombination rates in the genus, as evidence indicates that D. mauritiana, D. simulans, D. virilis, D. pseudoobscura, D. miranda, and D. persimilis all exhibit higher rates of recombination [33],[53],[99]; this should be considered when interpreting hitchhiking and linkage data from D. melanogaster to patterns of recombination in Drosophila in general.
Our results indicate that recombination affects diversity through mediating selection in the genome. While accounting for multiple covariates, we found no association between recombination and average pairwise divergence at 4-fold degenerate sites of unpreferred codons, and a significant, positive association of recombination with average pairwise diversity at 4-fold degenerate sites of unpreferred codons. Using data from our fine-scale maps, we ensured that recombination rates are nearly identical between the species used to generate divergence estimates; thus, we absolved a key assumption made in previous studies (see Figure S1). Data from Drosophila suggest both positive and negative selection are markedly less efficient in nearly nonrecombining regions of the genome [12],[47],[76],[101],[102], and a relationship of diversity but not divergence to recombination is apparent for other species of Drosophila [13],[33],[40],[49], mouse [36], beet [35], tomato [103],[104], Caenorhabditis [38], and yeast [105]. This last example is especially interesting because recombination is known to be mutagenic in yeast [24],[27], but there is a negative or absent divergence–recombination correlation [34],[105]; thus, it may be that recombination is somewhat mutagenic in many organisms, but the power of recombination to modulate the diversity eroding effects of selection likely has a much greater impact on the genome.
In other systems, the divergence–recombination association is positive, which may be interpreted as evidence that recombination is predominately mutagenic. A positive divergence–recombination association is apparent for humans [106],[107], maize [108], and in an inverted region between D. pseudoobscura and D. persimilis [25]. This association may be attributable to mutation [21], but unmeasured variables or segregating ancestral polymorphism could predispose a system to exhibiting a positive divergence–recombination relationship [34],[38]–[41]. For instance, in C. briggsae, segregating ancestral polymorphism leads to the signature of recombination-associated mutation (i.e., a positive divergence–recombination association), but further examination shows the majority of polymorphism heterogeneity is caused by recombination affecting the impact of selection at linked sites [38].
Since recombination probably mediates the effects of hitchhiking in our system, we sought to understand whether this hitchhiking is primarily positive or negative (background, purifying) selection and if recombination rate variation has a significant impact on the potential efficacy of selection. Evidence is emerging that in many organisms, especially those with large population sizes, selection may play a substantial role in shaping the genome [109]. For partial selfers, it seems that background selection substantially affects the genome [110]–[113], while in outcrossing species Drosophila, mice, and Capsella grandiflora a large fraction of the genome may be influenced by positive selection [40],[114]–[116]. The majority of studies find strong support that recombination can shape adaptive evolution when comparing regions of no recombination to regions with some or abundant recombination (reviewed in [7]). However, after accounting for multiple covariates in regions with detectable recombination rates, there is often very little relationship between recombination rate and the efficacy of selection [7],[12],[65].
Across chromosome 2, we found no relationship between the number of nonsynonymous substitutions and the recombination rate as measured with our fine-scale Flagstaff map. Reanalysis of the fine-scale data after removal of the first and last 3 Mb of the chromosome did not change the relationship of fine-scale recombination rate to nonsynonymous substitutions.
Our observation of a reduction of average pairwise diversity at 4-fold degenerate sites around nonsynonymous substitutions (Figure S9) is consistent with the idea that positive selection may have fixed many nonsynonymous substitutions along the ancestral lineage leading to D. pseudoobscura+D. persimilis, as has been argued elsewhere for other Drosophila species [68],[117]. While potentially less common, dips in diversity could also be caused by deleterious mutations that can get fixed by chance if deleterious selection coefficients are small enough—a situation we call “loser's luck” (Figure S10; but see [117],[118]), and theoretical investigations of entirely neutral substitutions showed that their quick fixation can also lead to dips in diversity [119]. Thus, while many of the dips in diversity we see may be caused by positive selection, both loser's luck and fixation of neutral substitutions may also contribute.
Diversity may be recovered slightly farther from a nonsynonymous substitution in areas of low recombination than in areas of high recombination, and such a relationship is not as pronounced for synonymous substitutions fixed along the lineage leading from the common ancestor of D. pseudoobscura and D. persimilis (Tables 5 and 6; Figure S9). Similarly, in Arabidopsis, haplotype blocks around nonsynonymous SNPs are larger than around synonymous SNPs [120]. Our data agree with theoretical expectations [69],[71] and past studies that show negative correlations of polymorphisms and nonsynonymous substitutions in Drosophila ([40],[68],[121],[122]; indeed, our data also show a significant negative relationship for nonsynonymous substitutions and within-species polymorphisms, generally (Tables 5 and 6). Yet the negative interaction term between recombination rate and distance from focal substitutions we observed is dependent on window size and distance from the substitution examined.
Our study documented global and local differences in recombination rate between two closely related species, and these data indicate that recombination probably modulates Hill–Robertson effects in the genome, causing a positive association of diversity with recombination. While we found no overall association of recombination rate with the number of nonsynonymous substitutions at the fine scale, we found evidence for dips in diversity around nonsynonymous substitutions that are dependent on the distance from the substitution, local recombination rate, and a number of other factors. In total, our study adds to the growing literature that indicates that selection must be a ubiquitously important factor for shaping diversity across much of the genome [30],[69],[71].
Using a backcross design, we developed two recombination maps for D. pseudoobscura (Flagstaff and Pikes Peak) and one recombination map for D. miranda (Text S1). For each cross, Duke's Genomic Analysis Facility genotyped 1,440 individual backcrossed flies for 384 line-specific SNP markers (see “SNP Development” section in Text S1) using the Illumina BeadArray platform (Illumina, San Diego, CA) [123].
Recombination events were scored when an individual fly's genotype changed from heterozygous to homozygous (for the parent in the backcross) or vice versa for autosomes and when the fly's genotype changed between the possible allele combinations for the sex chromosome arms XL and XR. Double crossovers were defined as adjacent intervals with different genotypes on both sides (for instance, a single homozygote genotype call nested in a tract of heterozygote genotype calls). We deemed these as genotyping errors as crossover interference is high within 2 Mb [124] and removed the single inconsistent genotype, scoring it as missing data. CentiMorgans were defined as the number of recombination events over the total number of individuals examined for each recombination interval, and we scaled this raw measure with a correction for recombination interference [125]. Throughout the article, recombination rates are given in Kosambi centiMorgans [125] per Megabase (cM/Mb).
Approximately 1,400 backcross progeny were scored for the Pikes Peak D. pseudoobscura map, approximately 1,250 backcross progeny were scored for the Flagstaff D. pseudoobscura map, and approximately 1,170 backcross progeny were scored for the D. miranda map (see Table S1 for the final number of individuals, number of intervals, and size of intervals over which recombination was measured).
Physical genomic distances used to calculate centiMorgans per Megabase (cM/Mb) per interval were based on the D. pseudoobscura reference genome v2.6 (Flagstaff) and v2.9 (Pikes Peak, D. miranda). Marker order was confirmed by the R (The R Foundation for Statistical Computing 2010) package OneMap [126] using the algorithms Recombination Counting and Ordering [127] and Unidirectional Growth [128]. Onemap does not accommodate backcrossed designs for sex chromosomes; therefore, we specified an F2 intercross design in these cases. We found one small inversion in D. miranda relative to D. pseudoobscura on chromosome 2. We estimated the left breakpoint was between the markers at 10,491,527 and 10,660,216 bp, and the right breakpoint was between the markers at 13,318,705 bp and 14,068,383 bp from the telomeric end of chromosome 2. This inversion corresponds to one previously documented between D. miranda and D. pseudoobscura between markers rosy and nop56 [129]. Figure S6 illustrates that recombination rate differences are probably not due to differences in gene order; thus, we used the D. pseudoobscura orientation for this inversion when comparing recombination between maps and excluded intervals that included the breakpoints. Confidence intervals (95%) for cM/Mb for each recombination interval were calculated by permutation [33],[54]. Confidence intervals for those intervals where we did not find a single recombinant individual were estimated from a binomial distribution—simply, we solved the equation (1−x)N = 0.05, where x is the 95% upper bound of recombination frequency, and N is the number of individuals surveyed.
The rationale for regressing out the effect of species (when identifying conserved intervals) was to account for the globally higher recombination rate in D. miranda relative to D. pseudoobscura and to identify regions where the recombination profile overlapped (e.g., where peaks and troughs can be overlaid). To delimit conserved regions using data that have not been corrected for elevated recombination rate of D. miranda, one might identify a region with very similar recombination rates between D. miranda and D. pseudoobscura, but this region may be a trough in recombination rate for D. miranda and a peak in recombination rate for D. pseudoobscura. Not correcting for the global elevation of D. miranda may lead to falsely concluding that a region has a conserved recombination profile between two maps. Thus, we used a rare events logistic regression (Zelig package in R) between each set of condensed fine-scale recombination maps to identify regions of conserved recombination after accounting for map identity (Flagstaff–Pikes Peak, Flagstaff–D. miranda, Pikes Peak–D. miranda). The package Zelig uses the same model as a logistic regression, but it corrects for a bias that is introduced when the sample contains many more of one of the dichotomous outcomes than the other. Recombination events conditioned on the total number of observations was the response variable, and species, interval, and species-by-interval were included as factors in the model. We defined “divergent” intervals as those where tests in each interval between the species from the rare events logistic regression had a q-value of <0.05 after correction for multiple tests [59]. “Conserved intervals” were those intervals that displayed a nonsignificant difference across all three maps when analyzed with a rare events logistic regression and had an odds ratio between 0.62 and 1.615, after accounting for a species effect. We did not correct for multiple tests in defining conserved intervals. The effect size, the confidence intervals for the effect size, p values, and multiple-test corrected q-values are available in Datasets S1, S2, and S3.
In this way, only intervals that were conserved within and between species were delineated as conserved intervals. The final dataset used to differentiate between the mutagenic and selection hypotheses contained 27 conserved intervals on chromosome 2. We did not use the XR to differentiate between the mutagenic and selection hypotheses—of the 44 intervals condensed across three XR maps, only seven were conserved within and between species. We chose not to combine data from chromosome 2 and XR, as there is some evidence for different evolutionary patterns between autosomal and sex chromosomes in Drosophila [130].
Details of how diversity and divergence were measured from the next generation sequencing data are given in Text S1. We analyzed the effect of recombination on diversity and divergence by applying a quasibinomial GLM as the data were overdispersed, which has several statistical properties favorable to analyzing proportions such as pairwise diversity [131],[132]. Diversity or divergence was used as a response variable by binding the number of SNP bases to the number of non-SNP, eligible bases with cbind in R. We included recombination rate, proportion of G or C bases within the recombination interval, gene density (measured as a proportion of nucleotides within the recombination interval that are coding), a proxy for neutral mutation rate (see Text S1), and interaction terms as factors in the model. See Text S1 for filtering steps that were required for a nucleotide to be considered an eligible base.
For these models, the analysis presented is restricted to those conserved, condensed intervals with highly similar recombination rates between all three maps, unless otherwise noted. This restriction removes a classic bias by requiring that the intervals have similar recombination rates between the two species compared for the divergence measures (Figure S1). Similar linear models were also analyzed using the uncondensed intervals for each of the three maps individually (Tables S9, S10, and S11). All statistics were performed in R version 2.12.1 (The R Foundation for Statistical Computing 2010) unless otherwise noted.
Using Flagstaff 16 and Flagstaff 14, we followed the same backcross scheme described in the section “Fine-Crossover Maps: Crosses and Technical Details.” Over 10,000 progeny from this backcross were stored in 96-well plates, frozen at −20°C and amplified for markers over these three regions. PCR products were visualized on a polyacrylamide gel using LICOR 4300 (see the section “Ultrafine Crossover Maps” in Text S1).
The number of nonsynonymous substitutions, specific to the D. pseudoobscura+D. persimilis lineage, were calculated for each gene using PAML using the resequenced genomic and reference genomic data described in Table S8 (one D. lowei, three D. miranda, three D. persimilis, two D. pseudoobscura bogotana, and 11 D. pseudoobscura genomes, filtered for quality as described above). We used a tree rooted with D. lowei and considered the branches leading to [D. persimilis (D. pseudoobscura, D. pseudoobscura bogotana)] to be the foreground branches (additional details in Text S1). We included D. persimilis a part of the foreground branch because relatively extensive interbreeding occurs between D. pseudoobscura and D. persimilis across much of the genome, aside from a few inverted regions [133]–[135].
Following [50], we used a GLMM with Poisson distribution to examine the potential for recombination rate to shape the distribution of nonsynonymous substitutions along the D. pseudoobscura+D. persimilis lineage. The model contained the following main effects: the number of silent segregating sites in each gene, GC content in each gene within Flagstaff 16, the proportion of coding bases 50 kb on either side of the gene's midpoint, weakly selected average pairwise divergence within the gene between D. persimilis and D. lowei at 4-fold degenerate sites of unpreferred codons (a proxy for neutral mutation rate), recombination rate observed for the interval containing the gene, and a random variable included to account for pseudoreplication of multiple genes per interval. The response variable was the number of nonsynonymous substitutions observed in each gene. This model construction allowed the inclusion of genes whose synonymous substitution count was zero (sensu [50]). The GC content from Flagstaff16 was used as this was the line used for backcrossing in the crossing scheme, and the Flagstaff map (D. pseudoobscura) was used in this analysis.
We used 4-fold degenerate sites of unpreferred codons to measure the average levels of diversity as a function of distance from amino acid substitutions along the D. pseudoobscura+D. persimilis lineage (as identified by PAML, see above).
Generalized linear mixed models with a Poisson distribution were used to compare the diversity around nonsynonymous substitutions along the D. pseudoobscura+D. persimilis lineage in relation to distance from the site and recombination rates measured in the Flagstaff cross. Measures of diversity at 4-fold degenerate sites were taken 60 kb (sensu [68]) from the site in either direction (120 kb total) with nonoverlapping bins of 1,000 bp. The random effects of identities of each substitution were estimated. We included as covariates (1) divergence between D. persimilis and D. lowei at 4-fold degenerate sites (a proxy for neutral mutation rate), (2) proportion of bases that were either G or C in Flagstaff 16 within the 1,000 bp window, (3) proportion of codons that were nonsynonymous substitutions within the 1,000 bp window, and (4) proportion of bases that were coding over each 1,000 bp window. The absolute value of the distance from the site and local recombination rate (at the particular nonsynonymous substitution) were included in the model as well as the interaction between distance and recombination rate. All effects in the model were standardized to mean zero and unit standard deviation. As a control, similar analyses were performed using synonymous substitutions along the D. pseudoobscura+D. persimilis lineage. Synonymous substitutions should evolve in a more neutral fashion; thus, less of an interaction between distance and recombination rate is expected. Any 1,000 bp window with less than 75 eligible, 4-fold degenerate sites was excluded from the analysis. Any nonsynonymous or nonsynonymous changes with less than 10 windows were excluded from the analysis. For the 60 kb analysis, after all filtering steps, our data consisted of 4,338 nonsynonymous and 8,670 synonymous substitutions along the D. pseudoobscura+D. persimilis lineage on chromosome 2. Four-fold degenerate sites were used here, rather than 4-fold degenerate sites at unpreferred codons, because too little data were available in each 1,000 bp nonoverlapping window.
|
10.1371/journal.pgen.1003911 | Human Intellectual Disability Genes Form Conserved Functional Modules in Drosophila | Intellectual Disability (ID) disorders, defined by an IQ below 70, are genetically and phenotypically highly heterogeneous. Identification of common molecular pathways underlying these disorders is crucial for understanding the molecular basis of cognition and for the development of therapeutic intervention strategies. To systematically establish their functional connectivity, we used transgenic RNAi to target 270 ID gene orthologs in the Drosophila eye. Assessment of neuronal function in behavioral and electrophysiological assays and multiparametric morphological analysis identified phenotypes associated with knockdown of 180 ID gene orthologs. Most of these genotype-phenotype associations were novel. For example, we uncovered 16 genes that are required for basal neurotransmission and have not previously been implicated in this process in any system or organism. ID gene orthologs with morphological eye phenotypes, in contrast to genes without phenotypes, are relatively highly expressed in the human nervous system and are enriched for neuronal functions, suggesting that eye phenotyping can distinguish different classes of ID genes. Indeed, grouping genes by Drosophila phenotype uncovered 26 connected functional modules. Novel links between ID genes successfully predicted that MYCN, PIGV and UPF3B regulate synapse development. Drosophila phenotype groups show, in addition to ID, significant phenotypic similarity also in humans, indicating that functional modules are conserved. The combined data indicate that ID disorders, despite their extreme genetic diversity, are caused by disruption of a limited number of highly connected functional modules.
| Intellectual Disability (ID) affects 2% of our population and is associated with many different disorders. Although more than 400 causative genes (‘ID genes’) have been identified, their function remains poorly understood and the degree to which these disorders share a common molecular basis is unknown. Here, we systematically characterized behavioral and morphological phenotypes associated with 270 conserved ID genes, using the Drosophila eye and photoreceptor neurons as a model. These and follow up approaches generated previously undescribed genotype-phenotype associations for the majority (180) of ID gene orthologs, and identified, among others, 16 novel regulators of basal neurotransmission. Importantly, groups of genes that show the same phenotype in Drosophila are highly enriched in known connectivity, also share increased phenotypic similarity in humans and successfully predicted novel gene functions. In total, we mapped 26 conserved functional modules that together comprise 100 ID gene orthologs. Our findings provide unbiased evidence for the long suspected but never experimentally demonstrated functional coherence among ID disorders. The identified conserved functional modules may aid to develop therapeutic strategies that target genetically heterogeneous ID patients with a common treatment.
| Intellectual Disability (ID) is defined by an IQ below 70, deficits in adaptive behavior and an onset before the age of 18. ID disorders are among the most common and important unmet challenges in health care due to their tremendous phenotypic and genetic heterogeneity [1], [2]. Many ID disorders are monogenic, and disease gene identification over the past decade has been very successful. More than 400 causative genes (referred to as ID genes) have been identified, providing unique stepping stones for understanding the molecular basis of cognition in health and disease. Some ID genes appear to work together in specific pathways and processes, such as Rho GTPase pathways, MAP kinase signalling and synaptic plasticity [3], [4]. This has led to the suggestion that ID genes highlight key molecular networks that regulate human cognition [1], [2], [5]–[7]. Such networks are of wide interest for both fundamental neuroscience and translational medicine, and can pave the way for developing treatment strategies [2]. However, their identification is limited by the paucity of available information on the function of most ID genes. Model organisms such as the mouse have effectively been used as experimental systems to gain insights into ID gene function and neuropathology [8]. Because such studies are time and cost intensive, ID research, whether in vitro or in vivo, has so far not moved beyond studying individual or small groups of genes. Novel approaches are required to allow functional studies to catch up with disease gene identification. We used Drosophila melanogaster as the model organism for this study. Genes, pathways, and regulatory networks are well-conserved between flies and humans [9]. Drosophila provides numerous approaches to investigate defects in neuronal function and behavior. Furthermore, fly models of selected ID disorders have already provided major insights into ID pathologies and have triggered the first therapeutic approaches [10], [11]. The efficiency of this organism and its available genome-wide toolboxes [12], [13] make Drosophila a powerful model to generate comparative phenotype datasets that can provide global insights into ID gene function and connectivity.
Here, we present a large-scale in vivo assessment of ID gene function and an in silico analysis of their Drosophila phenotypes and phenotype classes. We investigated the role of 270 evolutionarily conserved ID gene orthologs (referred to from here on as ‘Drosophila ID genes’) in the Drosophila compound eye, a highly organized array of ommatidia and photoreceptor neurons that allows for simultaneous assessment of neuronal function and physiology, and for multiparametric morphological analysis.
This comparative survey revealed a large number of novel functions for Drosophila ID genes including previously unappreciated regulatory roles in basal neurotransmission. It identified novel phenogroups in Drosophila that show phenotypic coherence in humans and molecular modules that can predict novel gene functions. Our study demonstrates that ID disorders converge on a limited number of highly connected functional modules.
To generate novel insights into the neuronal and molecular basis of cognitive (dys)function, we set out to manipulate established monogenic causes of ID in humans using Drosophila as a model. At the start of this project we conducted a systematic, manually curated disease gene survey. Of the identified 390 ID genes, 285 were conserved in Drosophila (for curation criteria and orthology see Materials and Methods). 95% of these genes, 270 Drosophila ID genes, can be targeted with Drosophila transgenic conditional RNA interference (RNAi) lines from an established validated toolbox [12], [14], [15]. This approach is a suitable approximation to the human disease conditions since (partial) loss of gene function is thought to be the causative mechanism for more than 250 of the 270 ID genes investigated (see Materials and Methods and Table S1A). We used a total of 498 RNAi lines, including two independent RNAi constructs per gene whenever available (Table S1A). To maximize the reliability in our primary screen, we selected lines which exceed previously determined quality criteria that guaranteed high reproducibility (see Materials and Methods, discussion, and Neumüller et al. [15]). Our strategy to ablate Drosophila ID gene expression primarily in the developing eye, including the photoreceptor neurons, was directed at identifying i) Drosophila ID genes that, if perturbed, cause defects in neuronal function, ii) Drosophila ID genes that affect viability, and iii) Drosophila ID genes that control different aspects of eye morphology (Figure 1A). We reasoned that these three classes and their subcategories might break down the large number of Drosophila ID genes into phenogroups, containing genes with a coherent function. Systematic targeting of a defined, larger group of genes in the eye and phenotypic characterization of various phenotypes has to our knowledge not previously been reported. Thus the degree to which phenotypes would be obtained was unknown.
The fast phototaxis assay is an efficient and robust test for neuronal function. It is based on the fly's innate behavior to move towards a light source [16], critically depends on proper performance of photoreceptor neurons, and can be quantified using the Phototaxis Index (PI) (Figure S1A). We optimized the assay using known vision mutants and their corresponding RNAi lines (Figure S1B,C). Under the chosen screening conditions (GMR-Gal4; UAS-dicer2 driver line, 28°C) all proof of principle RNAi lines showed strong defects, phenocopying their mutant phenotypes (Figure 1B, Figure S1B,C), which validated the efficiency of our approach.
In parallel to phototaxis, Drosophila ID gene knockdown progeny were examined for morphological eye phenotypes. As proof of principle for this additional approach, we tested RNAi lines against two Drosophila ID genes with reported eye phenotypes: ubiquitin protein ligase 3a (ube3a), the Drosophila ortholog of UBE3A implicated in Angelman syndrome, and daughterless (da), the ortholog of TCF4 implicated in Pitt-Hopkins syndrome. RNAi lines against both genes resulted in the expected defects, rough eyes [17] and complete loss of interommatidial bristles [18], respectively (Figure 1C). Progeny of the GMR-Gal4; UAS-dicer2 driver crossed to the genetic background line of the RNAi lines served as controls in all experiments of our study. Controls showed no considerable eye phenotypes (see Materials and Methods) and wildtype-like performance in the phototaxis assay.
In our screen, RNAi against the majority of all Drosophila ID genes (180 genes, 67%) resulted in lethal, phototactic or morphologic phenotypes (Figure 1D, Table S1B,C). Knockdown of the remaining 90 Drosophila ID genes (33%) did not yield functional or morphological eye phenotypes. The identified phenotype groups are described below.
Eighteen Drosophila ID genes (7%) gave rise to (partial) lethality and are thus essential in the targeted tissues (Table S1B,C). The eye driver GMR-Gal4 has recently been reported to show some expression outside the eye, which likely accounts for the lethality that was already reported by others [12], [19], [20]. Expression of these 18 genes was subsequently knocked down specifically in neurons, using the pan-neuronal driver elav-Gal4 (Figure 1A, grey asterisk). Only ERCC2 (human gene symbol)/Xpd (Drosophila gene symbol) and TPI/Tpi did not show lethality when ablated in neurons. Sixteen of the 18 GMR-Gal4-induced lethal genes also showed 100% lethality before adult stages upon selective neuronal knockdown (Table S1B). Thus, 16 Drosophila ID genes that are essential in neurons were identified using this strategy.
Ablating ID gene orthologs in the Drosophila eye and quantitatively assessing phototaxis yielded PIs between 1.1 and 5.9. Using a stringent cut-off of <4.0 to define phototaxis defects, we identified 25 phototaxis defective Drosophila ID genes (Figure 2A, Table S1B). Among these is the ortholog of ATP6V0A2, the vacuolar proton pumping ATPase subunit Vha100-1, mutations in which have been previously identified in an unbiased large scale phototaxis screen [21].
Electroretinograms (ERGs) were performed as a secondary screen to confirm that defects in phototaxis behavior are indeed caused by defective photoreceptor function and to further dissect the cause of defective vision in these ID models. ERGs are extracellular field recordings that measure the potential difference between the photoreceptor layer and the remainder of the fly body during light stimulation, revealing photoreceptor receptor transients (de- and repolarization) and synaptic communication (‘on’ and ‘off’ transients) [22]. Of the 24 Drosophila ID genes tested, we confirmed that 21 exhibited defective neuronal physiology. Of these, ATP6V0A2/Vha100-1 and SNAP29/usnp showed isolated synaptic defects characterized by normal receptor potentials but complete absence of ‘on’ and ‘off’ transients (Figure 2B). Two further Drosophila ID genes, DARS2 and GCH1, exhibited decreased amplitudes of receptor transients and reduced synaptic signalling, whereas the majority (17 of 21) of phototaxis hits were characterized by nearly absent depolarization and only residual synaptic communication (Figure 2B). In summary, we identified 21 Drosophila ID genes that are required either specifically for synaptic transmission or more broadly for basal neurotransmission and physiology. Only Vha100-1 has been previously demonstrated to be required for synaptic transmission in Drosophila photoreceptors. The majority of genes (16 of 21) had not been previously implicated in basal neurotransmission in any system or organism (Figure 2B, Table S2).
Internal eye architecture and the state of photoreceptors were monitored in order to obtain further insights into the cellular basis of the identified neurophysiological defects. Each wild-type ommatidium contains eight photoreceptors, organized in a stereotypical pattern (Figure 3A,B). Histological sections of ERG-defective Drosophila ID conditions detected a number of phenotypes (Figure 3, Table S1B). For example, knockdown of TBCE/tbce, implicated in hypoparathyroidism-retardation-dysmorphism syndrome, showed structural defects of developmental origin. R8 photoreceptors, normally located underneath photoreceptor 7, failed to be maintained in their appropriate proximal position and thus appeared in distal sections (Figure 3C). Moreover, rhabdomere extension towards the retina base, a process taking place during pupal development, failed in the majority of ommatidia (Figure 3C′) leading to distally accumulated “bulky” rhabdomeres (Figure 3C). This defect has recently been associated with regulators of the actin cytoskeleton that are linked to ID [23], [24]. In contrast, RNAi against several ERG defective Drosophila ID genes, including PEX7, ARFGEF2 and PAFAH1B1 caused neuronal degeneration of variable degrees, identifying a role for the encoded proteins in neuronal maintenance (Figure 3D–F). Thirteen of 21 ERG defective Drosophila ID conditions, including NKX2-1, PRPS1 and ATP6V0A2 knockdown animals, showed intact and properly organized photoreceptors (Figure 3G–I). Some of these conditions showed darker photoreceptor cytoplasm or pigment cell abnormalities (Figure 3G–I and Table S1B).
In summary, we identified genes required for neuronal development or maintenance among the ID orthologs that cause neurophysiological defects. In 20% of these cases the data confirm or extend previous findings. In the majority of instances (80%) these functions are novel (Figure 3, Table S2).
External eye morphology was systematically assessed in the primary screen to determine whether multiparametric phenotyping could identify which Drosophila ID genes work together in common developmental processes or molecular pathways. Thirteen phenotypic categories were identified: mildly rough, rough, partially fused ommatidia, fused ommatidia, fewer bristles, no bristles, stubble bristles, long bristles, necrosis, loss of pigmentation, small eye, wrinkled surface and dented surface (Figure 4A–M and Table S1B). 163 Drosophila ID genes showed at least one of these morphological phenotypes, which were classed as eye morphology defective. Mildly rough and rough phenotypes were the most numerous. Other defects occurred frequently in combination with these and/or with other phenotypes (Figure 4N). In all, RNAi-mediated knockdown of Drosophila ID genes in the eye generated a series of specific phenotype categories and identified a large number of genes with a role in the development of this tissue.
Interestingly, the frequency of morphological phenotypes among the phototaxis defective genes was very similar to their overall frequency in our screen. Thus, these phenotype classes do not significantly correlate (p = 0.13, hypergeometric test), which is also illustrated by the random distribution of morphologic phenotypes along the entire spectrum of phototactic performance (Figure 2A). We conclude that vision and external eye morphology do not depend on the same genetic/molecular machineries and provide a largely independent assessment of gene function.
We next sought to determine whether Drosophila eye morphology defects could provide insights into conserved functional networks that underlie human ID disorders. To our knowledge, such a correlation has not previously been evaluated. Therefore, we first examined the expression, annotated functions and protein interactions, comparing EMD (Eye Morphology Defective)- and NED (No Eye Defect)- ID genes (classes indicated in Figure 1D; the terms EMD- and NED-ID genes refer to Drosophila genes throughout the text).
Based on EST data from 45 human tissues [25], the human orthologs of both EMD-ID and NED-ID genes were widely expressed. For each gene we determined the tissue in which its normalized expression is highest (normalized for overall expression per tissue; see Materials and Methods). We found that the largest fraction among EMD-ID orthologs (9.8%, 16 genes) had their highest normalized expression in human ‘nerve’ tissue. This was also, among all tissues, the tissue where EMD- and NED-ID gene orthologs differ the most, as only 2.2% (2 genes) of NED-ID orthologs had their highest expression in ‘nerve’ (4.4 fold enrichment EMD-ID over NED-ID, P = 0.046). In contrast, the tissue in which most NED-ID orthologs had their highest expression was parathyroid (11.1%, 10 genes) (Figure S2A).
EMD-ID genes were also enriched for nervous system-related phenotypes in FlyBase, such as neuroanatomy, neurophysiology and photoreceptor defects (Figure S2B) as well as for Gene Ontology (GO) terms and KEGG pathways related to neuronal processes in humans. In contrast, NED-ID genes were enriched for GO terms related to metabolic processes (Figure S2C,D). The frequencies of human postsynaptic density proteins (hPSD; 1458 proteins, ∼7% of human genes [25]) among human orthologs of EMD- versus NED-ID proteins were also compared. In general hPSD proteins were significantly enriched among all ID genes (3 fold, χ2, P = 3.65e-18, ID genes (58) vs. human genome (1458)) but to a different extent among the two eye phenotype-based classes of ID genes: 25% of human orthologs of EMD-ID genes encoded hPSD proteins (3.4 fold enriched vs. genome, 41 proteins, Table S3), compared to 13% of human orthologs of NED-ID genes (1.8 fold enriched vs. genome, 12 proteins, Table S3). hPSD proteins are thus enriched by ∼2 fold among human orthologs of EMD-ID genes relative to NED-ID genes (χ2, P = 0.04).
In summary, human orthologs of EMD-ID genes tend to be more specific for the nervous system than the NED-ID gene orthologs with respect to their expression at the RNA and protein levels and with respect to the pathways they are involved in. The above determined fly phenotypes, human gene expression and annotated functions were plotted in a circos diagram to provide a global view of ID gene properties and to illustrate the consistent asymmetry in this composite landscape of ID (Figure 5, segments 2–8; a zoomable electronic version of the circos is provided as Figure S3). Annotated genetic interactions (DroID) and protein-protein interactions (PPI; from HPRD) between ID genes were also retrieved and integrated (Figure 5, segments 1 and 9). Interestingly, ID gene-encoded proteins have more than three times as many PPIs with each other as random proteins (PIE = 3.1; p<0.0001; taking into account the systematic biases in PPI networks for intensely studied genes that are caused by their high number of measured interactions [26]). These data substantiate that ID genes operate in common pathways. Restricting the analysis to human EMD-ID gene orthologs increased this connectivity, not just relative to the PPI database (PIE = 5.8; p<0.0001), but also relative to all screened ID genes (PIE = 1.7; p = 0.003). NED-ID gene orthologs also showed increased connectivity (PIE = 8; p<0.0001) relative to random proteins from the PPI database. The different biology of EMD-ID versus NED-ID orthologs that we observed at the pathway level is therewith supported by an enrichment of protein interactions within each class. The finding that ID genes show a high connectivity is, given their heterogeneity, not trivial.
To shed light on the functional connectivity of ID, we further examined Drosophila genetic interactions, comprehensive protein interaction data (HPRD and human interologs) and co-purified protein complexes (DPIM) and integrated these connections with the phenotypes we obtained. Strikingly, connections among mildly rough and among rough ID genes were each 6 fold enriched over randomly chosen Drosophila ID genes (p<0.0001). Connections between long bristles genes showed 20 fold (p<0.002), and connections between other bristles phenotype categories 24 fold (p<0.001) enrichment relative to randomly chosen Drosophila ID genes. This modularity extends beyond the eye morphological phenotypes. Lethal genes showed an 18 fold enrichment (p<0.001), and the most enriched phenotype class, the ERG defective genes, reached 47 fold enrichment in homotypic interactions (p<0.002) (i.e. interactions between genes that fall into the same Drosophila phenotype category). Connections within the categories fused ommatidia, necrosis, loss of pigmentation, and small eye, wrinkled or dented surface have not yet been reported in any of the utilized databases. The identified enrichments in known connectivity validate the approach to map molecular modules in ID through Drosophila phenotyping.
We next mapped the phenotype-based homotypic ID modules that are underlying the determined enrichments in connectivity among our phenotype categories (see Materials and Methods). In total, we identified 26 functionally coherent ID modules composed of 100 Drosophila ID genes and 200 homotypic connections (Figure 6A and its high resolution image provided as Figure S4). For the remaining 170 ID genes (63%), no homotypic connections were annotated.
Since Drosophila phenogroups showed high enrichments in known connectivity, they should be able to accurately predict novel gene functions and phenotypically relevant connections. To test this hypothesis, we further investigated the previously undocumented phenotype of abnormally long bristles, which identified a group of eight Drosophila ID genes. Five of these genes, PTEN, TSC2, RPS6KA3, MYCN and Myo5A, form a connected module (Figure 6A,B, module 9) associated with cancer biology [27]–[29]. In addition, PTEN, TSC2, RPS6KA3 and Myo5A also play a role in synapse development and plasticity in post-mitotic neurons [4], [30]. Therefore our data suggested an unappreciated role for MYCN, the fifth protein in the module, in this process. To address this prediction, synapse development at the Drosophila larval Neuromuscular junction (NMJ) was quantified. The NMJ is a well-established model synapse that has already provided a number of fundamental insights into ID gene function and pathways [10], [24]. Pan-neuronal knockdown of MYCN in larvae caused abnormally small synapses (Figure 6C). We also predicted a role in synapse development for the remaining three long bristles genes PIGV, UPF3B and DMD (encoding dystrophin). Indeed, not only does loss of dystrophin affect synaptic transmission [31] and has recently been found to cause susceptibility to malignant tumors in mice [32], it also affects activity of Akt [33], a kinase that directly regulates TSC2. DMD may thus connect to the long bristles module and act upstream of Akt-TSC2 signalling in tumor and synapse biology. PIGV catalyzes a step in the GPI-anchor biosynthesis pathway, and UPF3B functions in nonsense-mediated mRNA decay (NMD). Both have not yet been implicated in synaptic development or cancer although other members of the PIG family and NMD factors have [34], [35]. Knockdown of PIGV and UPF3B, like knockdown of MYCN, caused a significant reduction in synaptic size (Figure 6C), consistently observed among RNAi lines. To address whether smaller synapses represent a phenotype that is common among Drosophila ID genes or whether these characterize the long bristles module more specifically, three further Drosophila ID gene sets of equal size were randomly selected from the modules and screened for synaptic growth defects. Of the three gene sets targeted by a total of 16 RNAi lines, only a single RNAi line caused a smaller synapse (6% vs. 100% of RNAi lines targeting long bristles genes; p<0.001, χ2) (Figure S5). A further single RNAi line in another gene set caused an increase in synaptic size (13% vs. 100% that cause any defect in synapse growth; p<0.01, χ2). No phenotypes were present in the third dataset, see Figure S5. Thus, Drosophila eye phenogroups can predict novel functions of Drosophila ID genes and connections between them. In addition to this experimental validation, a number of our predictions are further supported by targeted literature search (Figure 6B dashed lines, Table 1, 2 and S4, discussion). Further conclusions from our phenotype data and their suggested implications are indicated in Table 1 and 2. We conclude that our data add considerable information on ID gene functional connectivity, and provide a comprehensive, integrated picture of modular genotype-phenotype networks in our disease model.
Are the identified Drosophila phenotype groups relevant to humans? To test this, we asked whether the corresponding genes showed, in addition to ID, also other similar disease phenotypes. Using the Human Phenotype Ontology (HPO) database [36], we first determined that, relative to human orthologs of NED-ID genes, EMD-ID gene orthologs were enriched for morphological features of the head/neck (∼3 fold, 64 vs. 22 of top 200 features, p<10−6, χ2). In contrast, NED-ID gene phenotypes were enriched for disorders of metabolism and homeostasis (17 fold, 17 vs. 1 of top 200 features, p<10−3, χ2), which is consistent with the associated GO terms discussed above. We further inspected individual fly eye phenotype groups and determined their associated human mean phenotypic similarity scores [37]. This score reflects the degree of overlap between human disease features associated with each gene. To address the phenotypic similarity beyond ID, we excluded ID and all terms residing below it in the HPO hierarchy as features from the calculation of the similarity scores. Comparison of similarity scores in each phenotype group against the background expectation for all genes in the HPO database revealed that the phenotypic classes fused ommatidia, bristle phenotypes other than long bristles and necrosis phenotype classes showed no significant human phenotypic cohesion. In contrast, the remaining phenotype groups, mildly rough, rough, long bristles, loss of pigmentation, small eye and wrinkled or dented surface, lethal and ERG defective were each associated with significantly increased human phenotype similarity (Figure 6D). Moreover, NED-ID genes also showed highly significant coherence in their associated human phenotypes. This is consistent with their enrichment for disorders of metabolism/homeostasis and with the high connectivity among NED-ID genes, together validating them as an independent phenotype category and illustrating that in comparative functional studies also the absence of phenotypes can be informative.
Altogether, our findings demonstrate that Drosophila phenotype groups identify coherent disease phenotypes and highly connected functional modules among the large group of genetically heterogeneous ID disorders.
The number of genes that are known to cause Intellectual Disability is growing rapidly. Some phenotypic overlap can be observed among ID disorders and a number of ID genes have been proposed to operate in joint molecular pathways. Despite these interesting observations, to date neither a comparative phenotype annotation for ID genes nor a systematic integration of the genotype-phenotype network spaces [38] has been attempted. Here we have combined large-scale phenotyping and bioinformatics to systematically generate and analyze phenotypes that are associated with 270 human ID gene orthologs in Drosophila.
A previously validated transgenic RNAi library [12] was used as discovery toolbox in this study. Because our past work determined significant differences in knockdown levels induced by RNAi using this toolbox (20–60% of wt mRNA levels [39]–[42]) and because we consistently found morphological eye phenotypes with two independent RNAi constructs only for 54% of the investigated ID genes, it seems likely that a number of RNAi lines are not efficient enough to evoke phenotypes. To limit the impact of such false-negatives on our analyses, we included phenotypes caused by single RNAi lines. This strategy has been applied in previous RNAi screens using the same toolbox [14], [15], [43]. Although we cannot exclude the occurrence of false-positive and -negative findings on the single gene level, phototaxis and eye morphology proof of principle experiments were successful and reliably recapitulated previously reported mutant phenotypes (Figures 1D and S1). Twelve percent of Drosophila ID genes (33 genes) have annotated anatomical eye defects in Flybase. Most of these genes were reliably picked up in our screen (29 genes, 88%), indicating that the degree of false-negative hits is low (Table S5). High reproducibility of phenotypes was previously reported for RNAi lines with a high s19 specificity score of >0.85 [15]. In our screen, we were able to use lines with an s19 value of 0.98–1 in 97% of all cases (Table S1B), exceeding this standard.
There is evidence from the literature for (partial or complete) loss-of-function as the underlying disease mechanism in 93% of the ID genes/disorders investigated in our screen (see Table S1A). Therefore, knockdown by RNAi appears to represent a suitable approach to model most of the studied ID genes. For 6% of the investigated ID genes we found support for gain-of-function mechanisms. Most of these (affecting 9 of 15 genes) are activating mutations in the Ras-MAPK pathway. This may limit the conclusions that can be drawn for these genes from our phenotypes. Nonetheless, we note that loss of Ras-MAPK signalling also compromises cognitive functions in mouse and humans [4]. Our phenoclustering approach successfully grouped these nine Ras-MAPK components into a single phenotype module.
Close inspection of the determined homotypic modules (Figure 6A) showed that in few cases, genes that act in established common pathways or processes are divided over different modules due to their distinct Drosophila eye phenotypes. This is the case for NF1, a direct negative regulator of Ras proteins that does not group together with HRAS and KRAS genes (module 1), as well as for mitochondrial NDUF and peroxisomal PEX genes that are divided over different modules (5, 10 and 11, 20, respectively). Since the NED phenotype is involved, it is possible that some of these ‘splits’ are due to inefficiency of RNAi lines leading to false-negatives, as discussed above. However, others appear to reflect the biology of the genes/gene groups. For example, NF1, in contrast to the above discussed nine Ras-MAPK genes, is a negative regulator of Ras-MAPK signalling. It is therefore conceivable that its knockdown causes another phenotype (NED) than knockdown of the positive Ras-MAPK regulators (rough eye). A second negative regulator of this pathway, SPRED1, which has recently been found to directly interact with NF1 [44], is a NED gene as well. For the PEX genes, we would a priori have expected these to cluster together in our screen. It is worth noting though that the distribution of different PEX genes into phenotypic modules matches the molecular architecture of the peroxisomal machinery [45]. PEX1 and PEX6 (module 20) represent the two cytosolic AAA proteins that directly interact to form the peroxisomal export complex. In contrast, PEX10 and PEX12 (module 11) are both ring-finger proteins that directly interact with each other to form the ubiquitin ligase complex. This complex is required for matrix protein import and subsequent release of the cytosolic matrix protein receptor encoded by PEX5, the third PEX protein in module 11 [45]. In summary, the determined homotypic modules are unlikely to give an error-free and complete picture of biologically meaningful relations between the studied ID genes. However, the consistent properties of EMD- versus NED-ID genes, the high degree of known connectivity among our phenogroups, their increased phenotypic similarity in humans and the demonstrated validation of the predicted synapse phenotypes argue that false (negative and positive) discovery rates in this study are limited.
In our screen, we identified more than 160 Drosophila ID genes that give rise to aberrant eye morphology, of which only 17% have been described previously on Flybase (Table S5). Furthermore, we identified 16 Drosophila ID genes that were required in the eye and in neurons for fly viability. Nearly half of these act in transcription or glycosylation-related processes. A further 21 Drosophila ID genes were required specifically for synaptic transmission or, more broadly, for basal neurotransmission. Histological analyses revealed that seven of these genes were essential for neuronal maintenance, whereas the majority was associated with functional defects despite structurally intact photoreceptors, implying that they impact neuronal transmission directly. CG7830, for example, is orthologous to two human non-syndromic ID genes, TUSC3 and MAGT1. These two genes encode subunits of oligosaccharyltransferase complexes required for N-glycosylation [46], which have recently been found to possess Mg2+ transport activity [47]. In neurons, defects in TUSC3 and MAGT1-mediated Mg2+ homeostasis might thus directly impact Mg2+-dependent ion channels. All defects in basal neurotransmission that we identified in our study (Figure 2B) provide a cellular mechanism that can directly underlie cognitive deficits in patients.
Phenomics, the phenotype correlate of genomics, is an emerging discipline in biomedical research [38], [48], [49]. Despite recently established adequate data depositories such as the HPO database, human phenomics lags behind genomics [48], limiting the recognition of genetic networks based on human phenotype data. Furthermore, the often small number of patients per genetic condition and the impact of environmental factors limit progress in human phenomics and are likely to remain bottlenecks in disease research. Comparative phenomic analyses in model organisms can contribute to the identification of evolutionarily conserved genotype-phenotype correlations in the human disease landscape.
Which animal phenotypes are relevant to ID disorders? Apart from defects of the nervous system such as the synapse, learning and memory defects [50], [51], we here show that also less complex phenotypes can be informative. Phenologs are defined as phenotypes enriched among orthologous genes in two organisms [52]. They can be used to unbiasedly identify and predict human disease models, even when the relationship between the phenotypes is not immediately obvious. This is illustrated by the predictive value of a specific yeast growth phenotype as model for mouse angiogenesis defects [52]. In Flybase, the available information on eye phenotypes is limited. However, the total fraction of annotated morphological eye phenotypes is three times higher among Drosophila ID genes than genome-wide (12.2% of Drosophila ID genes with annotated eye defects (Table S5) vs. 3.9% genome-wide, p = 1.01e-09, hypergeometric test). Thus, eye phenotypes are more likely to associate with Drosophila ID genes than with random genes, suggesting that to a certain degree they can serve as phenologs of human cognitive dysfunction. Furthermore, genes associated in fly with the same phenotype group show significant phenotypic similarity also in humans, validating Drosophila as a model for human disease phenomics of genetically highly heterogeneous disorders.
Using the genotype-phenotype associations generated in this study, we found strong homotypic connectivity among ID genes. Integrating public interaction data with the generated Drosophila eye phenotypes led to novel insights in gene function and functional connectivity. In total, we detected more than two dozen homotypic modules. About half of these (14 of 26) are pairs. Thus, while informative, these clusters likely represent only a minority of all biologically relevant interactions. Some of the connections within modules are well established, such as the PPIs that delineate the Ras-MAP kinase signalling pathway at the core of the largest phenotype module (Figure 6A). Our phenotypes imply novel gene functions and functional connections within each of the established phenotype categories. The long bristles cluster successfully predicted that MYCN, PIGV and UPF3B are critical for synapse development. Other predictions remain to be tested experimentally, but a number of them are already supported by other studies (Table 1, 2 and S4). For example, despite lack of data in the utilized databases, the microtubule and neuronal migration-disorder related rough eye module two can be linked to other rough eye genes such as CC2D2A, TMEM67 and SMC3, and potentially to other rough eye genes such as Rab3GAP1, Rab3GAP2, ARFGEF2, FKRP, VLDLR and ARX as supported by shared human neuronal migration phenotypes (Figure 6B, dotted lines). CC2D2A- and TMEM67-associated ID disorders are ciliopathies, and apart from its established role in chromosome cohesion, SMC3 has been recently shown to be required for Planar Cell Polarity, a process underlying cilium formation [53], [54]. These data therefore point to an intimate connection between neuronal migration disorders and ciliopathies. Indeed, a recent paper reported that migrating interneurons display dynamic primary cilia that carry receptors for guidance cues, the dynamics of which are disturbed in a ciliopathy [55].
Another example is the fused ommatidia phenotype (Figure 3J′), which resembles a phenotype previously reported in the literature as “glossy”. This phenotype has been proposed to identify genes with mitochondrial function [56], which is required for synaptic energy supply, receptor trafficking and calcium buffering. Indeed, among the twelve Drosophila ID genes in this phenotype category are the fly orthologs of PPOX, SURF1 and DBT, three further genes with established mitochondrial function. Also ASL, a cytosolic enzyme of the urea cycle that partly takes place in mitochondria, gives rise to this phenotype. Four other fused ommatidia Drosophila ID genes encode regulators of transcription including MED12, a subunit of the mediator complex that in yeast has been shown to regulate transcription of genes with mitochondrial function [57]. In this context, it is important to note that functional connectivity between transcription factors and their target genes remains undetected in many databases, whereas this phenotype-based approach can identify or increase confidence in such relations. The “no bristles” category contains the Drosophila orthologs of FGFR2, FGFR3, PAFAH1B1 (encoding Lis1) and the transcription factor TCF4, and comprises only a single annotated connection (FGFR2, FGFR3, Figure 6A). However, ModENCODE data show that the TCF4 ortholog da targets the two Drosophila FGF receptor genes htl and btl [58](Figure 6B), supporting further functional connections within this mini-cluster. Given the number of ID genes that encode transcription regulators, disruption of gene regulatory networks that comprise several ID genes are likely to contribute to the aetiology of ID.
In the era of Next Generation Sequencing in human genomic research and diagnostics, the necessity to provide functional evidence of identified candidate disease genes is increasing exponentially. Here we have demonstrated that human disease phenomics in Drosophila is feasible, despite 1300 million years of evolutionary distance between the two species [59]. The identified genotype-phenotype modules, in combination with efficient fly phenotyping, should be applicable to facilitate identification of causative mutations among multiple DNA variants. Moreover, mapping molecular modules in ID provides a step towards network-based strategies that can target genetically heterogeneous patients with a common treatment. Recent research has demonstrated that cognitive defect in several animal models of ID are reversible in adulthood [60], [61]. Two of these genes, PTEN and TSC2, are part of the long bristles cluster, making other partners in this module attractive targets for genetic and pharmacologic rescue experiments and future clinical trials.
ID genes were identified in the literature, in public and in-house databases, and manually curated by clinical specialists. Also conditions that might not be primarily regarded as ID syndromes (due to other prominent features or partial penetrance) were considered if independent genetic as well as independent clinical evidence for ID was found. Conditions with clinically or genetically low evidence or treatable metabolic conditions were not considered. To enrich for genes that act in neurodevelopmental processes underlying cognition, also genes associated with neurodegenerative manifestation (late onset), severe neurologic defects and early lethality were excluded. The orthologs of 390 ID genes (as of beginning of 2011) were determined using ENSEMBL's orthology classes (www.ensembl.org) and treefam annotations, including manual curation. One-to-one and one (fly)-to-many (human) orthologs were considered, identifying 285 fly orthologs. RNAi lines were available for 95% of these, which are subject of this study. In eight cases, two human paralogs are implicated in ID and have a common ancestor in Drosophila. Drosophila phenotypes and data associated with these were assigned to both human genes.
Of the 270 investigated human ID disorders/genes, 200 are recessive (OMIM, the Online Mendelian Inheritance in Men database), and 28 further ID genes are reported to be haploinsufficient [62]. For 24 of the remaining 42 ID genes, evidence for (partial) loss-of-function as the underlying mechanisms exist (Pubmed, summarized on OMIM), illustrating that for >93% of ID disorders the pathomechanism is (partial) loss-of-function. In a very few cases (4/270) no data are available that would allow conclusions about loss versus gain-of-function as ID underlying mechanism. Support for gain-of-function mechanisms accounts for 5% (14/270) of the investigated ID genes.
Conditional knockdown of Drosophila ID genes was achieved with the UAS-GAL4 system [63], using a w; GMR-Gal4; UAS-dicer2 driver [12], [19] and UAS-RNAi lines [12]. UAS-RNAi lines, their genetic background controls (60000, 60100) and UAS-dicer2 (60009) were obtained from the Vienna Drosophila RNAi Centre (VDRC). GMR-Gal4 (1104), elav-Gal4; UAS-dicer2 (25750), nonA4b18 (125), norpA45 (9051), w*; sr1 ninaE17 es (5701) and w*; ort1 ninaE1 (1946) were obtained from the Bloomington Drosophila stock center (Indiana University). Crosses were cultured according to standard procedures and raised at 28°C unless indicated otherwise.
Information collected in previous RNAi screens [14], [15], [43] was utilized to select genetic tools (GB and KK collections, see www.vdrc.at). ID lines from the site-integrated KK library were included in the primary screen. These lines bear no risk for gene disruption at the integration locus, ensure high expression and represent independent constructs that do not overlap with those of the GB collection. They are also characterized by minimized off-targets, reflected in high s19 values (Table S1B). Including the potent KK library in our screen allowed us to use lines with highly specific s19 scores of 0.98–1 in 97% of all cases.
A modified countercurrent apparatus was used to fractionate genotypes among six tubes, according to their visual activity (see Figure S1). The phototaxis index (PI) is calculated as ∑i*Ni)/N, where N is the number of flies, i is the tube number, and Ni is the number of flies in the ith tube. Average PI and standard deviation were calculated from three independent experiments on different test days. Assays were performed under standardized conditions, and progenies from control crosses served as internal controls. Populations of 40–70 flies, mixed sex, at the age of day 3–4 after eclosion and a walking time of 15 seconds were used. Based on the average PI of the control (PI = 5.2), and a maximal standard deviation of 1.2 per RNAi line, we defined a stringent cut-off of PI<4 to define a phototaxis hit.
Eye morphology defects were scored by two independent experimentators. Despite a reported effect of GMR-Gal4 driver constructs on eye development [64], our driver controls showed merely mildly rough phenotypes in a maximum of 10% of eyes. A mildly rough phenotype was therefore only scored if present in the majority (>90%) of knockdown eyes. No other eye phenotypes were observed in controls.
Three to four days old females of the appropriate genotype were fixed in 1% glutaraldehyde, dehydrated by an ethanol series (25, 50 and 75%), critically-point dried and mounted on aluminum stubs. Samples were coated in gold by sputter coating and afterwards examined with a JEOL 6310 SEM.
Heads from 3–4 days old female progenies raised at 25°C were prefixed for 30 min in 2% glutaraldehyde buffered with 0.1 M Sodium cacodylate pH 7.4, bisected and fixed for another 24 hours. Bisected heads were postfixed for 1 hour in 1% Osmium teroxide in Paladebuffer pH 7.4 with 1% Kaliumhexacyanoferrat (III)-Trihydrat, dehydrated in ethanol and propyleenoxide and embedded in a single drop of Epon. Semi thin, 1 µm thick transverse and longitudinal sections were stained with 1% Toluidine Blue.
ERGs were performed as previously described [65]. Flies were tested at day one after eclosion. Per genotype eight to ten flies were recorded and the average of five representative recordings is shown.
Segment 2, 3 and 4 muscle 4 Type 1b neuromuscular junctions (NMJs) of wandering L3 panneuronal knockdown larvae were analyzed after dissection, a 30 min fixation in 3.7% PFA and immunolabelling with an anti-discs large 1 antibody (anti-dlg1, supernatant, 1∶25) (Developmental Studies Hybridoma Bank, University of Iowa). NMJ pictures were obtained using a Leica automated brightfield multi-color epifluorescence microscope. Images were automatically processed and the synapse area was measured by an advanced in house-developed Fiji/ImageJ macro. Mutant synapses were compared to their proper genetic background controls. For the X-linked UPF3B RNAi line 31444 and its control, exclusive female knockdown animals were selected. UPF3B RNAi line 31445 was not available at the stock centre for retesting. In contrast, for AP1S2, NDUFS8 and CHD7 independent RNAi lines were available at the time of synapse evaluation and have been utilized. At least 16 synapses were analyzed per genotype. Random sets of Drosophila ID genes subjected to NMJ analysis were determined from homotypic modules using a PHP script-based random number generator. Constraints were set on the min and max values and previously generated numbers were excluded to avoid duplicates. Independent sets of specified size were generated for subsequent analysis.
Drosophila ID genes were assigned to all phenotype categories that describe (an aspect of) the observed associated defects. Since RNAi induces variable knockdown that will in some cases not be sufficiently strong to evoke a loss-of-function phenotype, “single hit” genes were included in the further data analysis, as in previous Drosophila RNAi screens [14], [15], [43]. In any other scenario, one inefficient RNAi line would disqualify the efficient one, which would likely result in a large amount of false-negatives. For annotations of already known defects associated with EMD- and NED-ID or all Drosophila ID genes, the Drosophila genes annotated with defective phenotypic classes behavior, neuroanatomy, neurophysiology, behavior, photoreceptor, cell cycle and stress response phenotypes as well as with anatomy defective classes retina and photoreceptor cell were fetched from FlyBase (version march 2012) (www.flybase.org) [66]. A hypergeometric distribution test was carried out to check the enrichment of these phenotypes within EMD-ID and NED-ID genes against the background of (fly) phenotypes associated with all Drosophila genes that have orthologs in human.
EST profiles from cDNA libraries of 45 normal human tissues were retrieved from the NCBI UniGene database [67] (ftp://ftp.ncbi.nih.gov/repository/UniGene/Homo_sapiens/Hs.profiles.gz) and expression abundance for each gene across the tissues was calculated. Since average expression between tissues varied significantly, we ranked genes in each tissue according to their expression levels. Subsequently we determined for each gene the tissue of its highest normalized expression as the one in which the gene had its highest rank.
Overrepresentation of GO biological process and pathway terms for human EMD- and NED-ID gene orthologs against the human genome background data sets were identified using the Database for Annotation, Visualization and Integrated Discovery (DAVID) v6.7, web based program [68], [69].
Direct physical protein-protein interaction data sets (HPRD_Release9_041310.tar.gz) from the Human Protein Reference Database (HPRD [70]) were downloaded and used as the standard protein interaction data for our study. Human interologs [71] (containing interactions from HPRD, BioGRID, IntAct, MINT, and Reactome; version 2012_04), DPIM-coAP complex data (protein interactions determined in large-scale co-affinity purification screens, Drosophila Protein Interaction Mapping project [72] (DPIM; version 2012_04), and Drosophila Genetic interaction data (version 2012_04) were downloaded from DroID (http://www.droidb.org/) [73], [74]. Physical interaction enrichment (PIE) scores of human orthologs of EMD- and NED-ID genes were calculated against HPRD, using the PIE algorithm with a minor modification in the normalization factor [26] to account for biases in the number of reported interactions for disease genes. Interaction enrichment scores for the specific phenotype categories within EMD, for lethal and for ERG ID gene products represent the number of unique connections determined from the combined interaction data sets per phenotype (HPRD, human interologs, DPIM-coAP complex and genetic interactions) divided by the number of connections for randomly (10,000 times) chosen ID genes from the combined interaction data sets.
Circos-0.56, a freely available software package [75] was downloaded and used for the depiction of most phenotypes and significantly enriched features, determined as described above.
The combined interaction data sets (see ‘Interaction network datasets and analyses’ above) were loaded into and visualized with the Cytoscape v2.8.1 tool [76]. Different phenotypes were colored using the MultiColored Nodes plug-in v2.4.0 [77]. Homotypic phenotype modules were identified among the entire ID interactome using Cytoscape's v2.8.1 ‘create new network from attribute’ algorithm. The phenotype-based homotypic ID modules are defined as connected genes with shared phenotype. Thus, genes with a non-overlapping phenotype cannot be part of the same phenotype-based module.
The Human Phenotype Ontology (HPO) [36] genes-to-phenotype mapping file, build 694, was downloaded from the HPO website (www.human-phenotype-ontology.org). This file maps genes to lists of standardized phenotypic features organized in a hierarchical structure (ontology). Phenotype similarity was determined based on these feature lists, using an adapted version [37] of a previously published algorithm [78] that takes the hierarchical structure into account. Basically, the human phenotypic similarity per gene pair was determined by calculating the correlation coefficient of the HPO feature vectors associated with each gene. The seven HPO features in the “Intellectual Disability” subtree were excluded from the feature vectors as the analyzed genes were selected based on this feature. Features were weighted according to their rarity and the number of features present in the vector. Before the feature vectors were compared, they were first supplemented with indirectly annotated features based on the feature hierarchy. This was accomplished by recursively adding parent features with progressively lower weights until the root of the feature hierarchy was reached. For each fly phenotype category, the mean pair-wise phenotypic similarity score was determined for all human genes associated with it. As a control, each set's score was compared with those of 1000 equal-sized sets of genes randomly sampled from the full list of HPO genes. For comparing the over-represented individual features of EMD-ID and NED-ID genes, we first identified the top 200 most significantly over-represented human phenotypic features for each gene set. This number was chosen to ensure that all considered features were over-represented at a corrected p-value threshold of 0.05 (Hypergeometric distribution; 206 and 563 features associated with NED-ID and EMD-ID genes respectively meet this threshold). Subsequently we determined what percentage of these specific features fall into the various top level HPO phenotypic categories, and compared these between EMD- and NED-ID genes.
|
10.1371/journal.pbio.1002577 | Prediction Errors but Not Sharpened Signals Simulate Multivoxel fMRI Patterns during Speech Perception | Successful perception depends on combining sensory input with prior knowledge. However, the underlying mechanism by which these two sources of information are combined is unknown. In speech perception, as in other domains, two functionally distinct coding schemes have been proposed for how expectations influence representation of sensory evidence. Traditional models suggest that expected features of the speech input are enhanced or sharpened via interactive activation (Sharpened Signals). Conversely, Predictive Coding suggests that expected features are suppressed so that unexpected features of the speech input (Prediction Errors) are processed further. The present work is aimed at distinguishing between these two accounts of how prior knowledge influences speech perception. By combining behavioural, univariate, and multivariate fMRI measures of how sensory detail and prior expectations influence speech perception with computational modelling, we provide evidence in favour of Prediction Error computations. Increased sensory detail and informative expectations have additive behavioural and univariate neural effects because they both improve the accuracy of word report and reduce the BOLD signal in lateral temporal lobe regions. However, sensory detail and informative expectations have interacting effects on speech representations shown by multivariate fMRI in the posterior superior temporal sulcus. When prior knowledge was absent, increased sensory detail enhanced the amount of speech information measured in superior temporal multivoxel patterns, but with informative expectations, increased sensory detail reduced the amount of measured information. Computational simulations of Sharpened Signals and Prediction Errors during speech perception could both explain these behavioural and univariate fMRI observations. However, the multivariate fMRI observations were uniquely simulated by a Prediction Error and not a Sharpened Signal model. The interaction between prior expectation and sensory detail provides evidence for a Predictive Coding account of speech perception. Our work establishes methods that can be used to distinguish representations of Prediction Error and Sharpened Signals in other perceptual domains.
| Perception inevitably depends on combining sensory input with prior expectations. This is particularly critical for identifying degraded input. However, the underlying neural mechanism by which expectations influence sensory processing is unclear. Predictive Coding theories suggest that the brain passes forward the unexpected part of the sensory input while expected properties are suppressed (i.e., Prediction Error). However, evidence to rule out the opposite mechanism in which the expected part of the sensory input is enhanced or sharpened (i.e., Sharpening) has been lacking. In this study, we investigate the neural mechanisms by which sensory clarity and prior knowledge influence the perception of degraded speech. A univariate measure of brain activity obtained from functional magnetic resonance imaging (fMRI) is in line with both neural mechanisms (Prediction Error and Sharpening). However, combining multivariate fMRI measures with computational simulations allows us to determine the underlying mechanism. Our key finding was an interaction between sensory input and prior expectations: for unexpected speech, increasing speech clarity increases the amount of information represented in sensory brain areas. In contrast, for speech that matches prior expectations, increasing speech clarity reduces the amount of this information. Our observations are uniquely simulated by a model of speech perception that includes Prediction Errors.
| The observation that our perception of the world not only depends on sensory input but also on our prior knowledge has been of longstanding interest in psychology [1] and neuroscience [2–5]. There is widespread agreement that sensory input and prior knowledge are combined in neural representations; by which we mean the specific patterns of neural activity that are associated with the content of our sensory experiences. However, despite extensive experimental work in many sensory modalities [6–16], the neural and computational mechanisms by which prior knowledge guides perception are unclear [17,18].
One proposal is that neural representations of expected sensory signals are enhanced or tuned [19,20]. Critically, in this account, perceptual representations are sharpened by relevant prior expectations in much the same way as if the quality of the sensory input was increased [17,18]. Alternatively, Predictive Coding schemes suggest that expected sensory input is explained away and unexpected information is represented in the form of prediction errors (cf. in engineering [21,22] and neuroscience [3,23,24]). One intuitively attractive aspect of Predictive Coding, both for engineering and neuroscience, is its assumption that minimal effort should be invested in further processing of sensory information that is already known or expected.
Our goal in this work is to distinguish these two fundamental coding schemes for how prior expectations influence perception. Do neural representations of sensory signals contain only the unexpected parts of the sensory evidence (from now on we will refer to these as “Prediction Errors”)? Or do they contain an enhanced version of the expected sensory evidence (from now on “Sharpened Signals”)? Our approach allows us to test each of these coding schemes against behavioural and fMRI data to determine how expected sensory signals are neurally coded.
Sharpening and Predictive Coding schemes have proved hard to distinguish in neuroscience [2,5,25]. Predictive Coding theories have proposed that each level of a cortical hierarchy contains two functionally distinct subpopulations (i.e., prediction and prediction error units [3,20,24,26]). In these accounts, the signals that are passed forward from one level of the hierarchy to the next (i.e., the feedforward signals) represent Prediction Error. This Prediction Error signal is also used to update prediction units within the same level of the cortical hierarchy (through lateral interactions), such that prediction units represent a sharpened version of the sensory signal [3]. Therefore, evidence for Sharpened Signal representations has been used to support both Predictive Coding theories [20] as well as pure Sharpening theories without computation of Prediction Errors [27]. However, evidence for Prediction Error representations would be uniquely consistent with Predictive Coding and challenge pure Sharpening accounts.
Speech perception provides a biologically significant domain in which prior knowledge has been clearly shown to guide perception (for review, see [28]). Behavioural experiments show that numerous sources of proximal and distal prior knowledge (including subtitles, lip-reading, lexical constraint, or semantic predictability) can enhance subjective and objective perceptual outcomes for degraded speech [29–33]. The dominant computational theories of speech perception have included interactive-activation mechanisms that lead to enhanced representations of expected signals (i.e., Sharpened Signals), most notably in the TRACE model [34] but also in other influential models of speech perception [35–38]. More recent work has proposed Predictive Coding schemes, which use Prediction Error signals [4,7,39] to explain how prior expectations improve sensory processing. However, evidence to overturn Sharpening accounts has been lacking.
One challenge for existing research is that both suggested computational schemes predict reduced neural activity during perception of expected speech signals, either due to suppression of unexpected noise (in Sharpened Signals) or suppression of expected signals (in Prediction Errors). Brain regions in and around the left posterior superior temporal sulcus (STS) are proposed to support perceptual processing of speech [40,41] and integrate expectations from different modalities with speech input [8,39,42–46], and activity in this region is proposed to show effects of prior training on speech responses [47–49]. While these studies provide abundant evidence that prior knowledge can influence the magnitude of activity in the posterior STS during speech perception, they do not determine the computational mechanism by which relevant prior knowledge enhances perception of speech.
However, multivariate analyses of the representational content of brain responses can differentiate these two accounts by testing whether representations of speech signals are enhanced (in line with Sharpened Signals) or suppressed (Prediction Errors) when they match prior expectations. Therefore, we used representational similarity analysis [50] on multivoxel response patterns in the posterior STS. This approach is “information based” because it measures how much information about the phonetic form of speech is contained in spatial fMRI activation patterns in each of the experimental conditions that we tested [51,52]. We focus on the posterior STS because this is both a region in which effects of prior knowledge on speech processing have been repeatedly shown and also a region in which syllable identity can be decoded from multivariate BOLD signals [53–57].
To guide our interpretation of this data, we constructed two computational simulations based on either Sharpened Signals or Prediction Errors. Both these simulations can explain observations of perceptual enhancement and reduced fMRI responses in the left posterior STS for degraded speech that matches prior expectations. Crucially, however, these simulations make distinct predictions for the results of multivariate representational similarity analysis. In our Sharpened Signal model, simulated neural representations are enhanced for degraded speech that matches prior expectations in the same way as for speech that is presented with more sensory detail (Fig 1A). However, in our Prediction Error model (Fig 1B), these two manipulations have an interactive effect on simulated neural representations: the effect of increasing sensory detail depends on whether or not speech matches prior expectations. Increased sensory detail for expected speech leads to reduced information about the phonetic form of speech in simulated Prediction Errors. In contrast, increased sensory detail for unexpected speech leads to more Prediction Error and, hence, more information in simulated neural representations. In our experimental work, we test both these proposals using representational similarity analysis (RSA) fMRI applied to BOLD responses time-locked to the onset of a degraded spoken word.
To obtain experimental evidence to differentiate these two computational accounts, we therefore simultaneously manipulated (1) prior knowledge of speech content by having participants read matching/mismatching written words or neutral text (“XXXX”) before spoken words [8,33,58] and (2) sensory detail in speech by presenting vocoded spoken words at one of two different levels of acoustic degradation (Fig 2) [59,60]. In this way, we could test whether representations of the phonetic form of speech in the posterior STS [55,57,61] are enhanced similarly by changes in prior knowledge as by changes to sensory detail (in line with Sharpened Signals) or whether these two factors interact (in line with Prediction Errors).
First, we confirmed that, consistent with both Predictive Coding and Sharpening, providing informative prior expectations improves perception of degraded speech. Participants’ report of the degraded spoken words was improved by both increased sensory detail and matching prior information from a preceding written word (Fig 3). A two-way repeated measures ANOVA with the factors sensory detail (4- versus 12-channel) and prior knowledge (Match versus Neutral) revealed significant main effects of sensory detail on word report (12-channel: 85.39% > 4-channel: 57.83% correct; F(1, 20) = 133.419, p < 0.001, eta squared = 86.96) and prior knowledge (Match: 84.42% > Neutral: 63.49% correct; F(1, 20) = 89.582, p < 0.001, eta squared = 81.75), and a significant interaction (F(1, 20) = 74.997, p < 0.001, Fig 3A).
These effects of sensory detail and prior knowledge combined such that 4-channel vocoded speech in the Match condition was reported with equivalent accuracy as 12-channel vocoded speech in the Neutral condition (79.17% versus 83.53% correct, t(20) = -1.427, p = 0.169). Nonetheless, word report was further enhanced in the Match 12-channel condition compared to the Neutral 12-channel condition (89.68% versus 83.53% correct, t(20) = 3.267, p = 0.004) and the Match 4-channel condition (89.68% versus 79.17% correct, t(20) = -4.460, p < 0.001). Word report in the Match 12-channel condition was also more accurate than in a condition in which the spoken word was omitted and participants were prompted to report the preceding written word (89.68 versus 82.14% correct in the written only condition, t(20) = 2.348, p = 0.029). These findings confirmed that participants used prior knowledge to enhance perception of degraded speech even when relatively clear 12-channel speech was presented. Behavioural responses in the Mismatch conditions resemble the pattern of results in the Neutral condition (see S1 Fig).
Second, we sought to localise the univariate BOLD activity decrease for degraded spoken words that follow matching written words relative to words following neutral cues. These observations replicate previous findings but do not distinguish between accounts in which this effect is due to suppression of unexpected noise (Sharpened Signals) or suppression of expected signals (Prediction Errors). Univariate BOLD responses were influenced by both increased sensory detail and matching written text. A two-way repeated measures ANOVA with the factors sensory detail (4- versus 12-channel) and prior knowledge (Match versus Neutral) revealed a main effect of matching versus neutral prior knowledge on responses in the left posterior STS, as predicted, and in other regions of the speech processing network (Fig 3B and 3C, S1 Table: main effect of Match/Neutral, p < 0.05 FWE voxel correction). Mean beta values extracted from the left posterior STS showed a reduction during Match in contrast to Neutral conditions (Fig 3; inspection of contrast estimates from all other clusters also revealed less activity for Match than Neutral). In addition, there was a main effect of sensory detail in bilateral insula, SMA, left premotor, and orbitofrontal cortex (S2 Table; main effect of 4/12-channel, p < 0.05 FWE). Inspection of contrast estimates revealed less activity for 12- than 4-channel in most clusters; the reverse pattern was only observed in the right middle orbitofrontal gyrus). The interaction of prior knowledge and sensory detail did not reach corrected significance (S3 Table).
Increased BOLD activity for Mismatch > Match resembles the difference in BOLD activity found for Neutral > Match (see S1 Fig and S4 Table). This confirms that our observed effects are not due to differences in attention, anticipation of more difficult trials, or baseline differences between the Match and Neutral conditions (see S1 Text), but rather due to the influence of matching prior knowledge on speech perception.
The behavioural and univariate results appear to be in line with both Sharpening and Predictive Coding theories. Although the underlying coding schemes differ, both accounts suggest that increased sensory detail and matching prior information should improve recognition performance and that prior matching knowledge should reduce univariate fMRI responses. To confirm this, we constructed two computational models of spoken word recognition, which only differed by using representations of Sharpened Signals or Prediction Errors to simulate how sensory information and prior knowledge are combined (see S2 Fig for details). In both these models, behavioural performance (i.e., word recognition) was simulated by the model’s ability to identify the correct word presented in degraded speech, and univariate fMRI results (i.e., the magnitude of hemodynamic activity in the left posterior STS) were simulated by the number of processing iterations required for the model to settle. By simulating the univariate fMRI signal with the number of model iterations, we assume that the hemodynamic signal as measured by fMRI integrates over several seconds of neural activity and that a longer duration of neural processing should result in an increased amplitude of the fMRI signal [63]. Six parameters were optimised for each model: the amount of sensory degradation used to simulate 4- and 12-channel vocoded speech (which influences word report and processing time), variability and confidence in behavioural responses (which influences word report), and the rate and duration of model updating (which primarily influences processing time; see S3 Fig for sensitivity analysis of the optimized parameters).
We used Akaike weights to compare goodness of fit to word report and univariate hemodynamic responses in the left posterior STS (see Materials and Methods for details). Based on 1,000 replications using the best-fitting set of parameters, a probability density function for the predicted outcome of behavioural and univariate results was generated for both model simulations. We then used the evidence ratio of Akaike weights to compare the relative likelihood of the two models given the observed data. The ratio of the Akaike weights revealed a slightly higher likelihood of Sharpened Signal model than of the Prediction Error model for both the behavioural results (wPE/wSharp = 0.9307) and the univariate results (wPE/wSharp = 0.8149). Both of these values are close to 1, indicating that there is a negligible difference between the two models [64]. The good fit observed between these models and behavioural and univariate hemodynamic data from the current experiment suggests that computation of Sharpened Signals and Prediction Errors can explain the effect of increased sensory detail and matching prior information during perception of degraded words (model simulations and experimental results shown in Fig 3).
Although both models can accurately simulate behavioural and univariate fMRI results, they perform different underlying computations and make different assumptions about the effect of matching prior knowledge on neural representations of speech signals. The Sharpened Signal model predicts that degraded speech is better represented in the STS when it matches prior knowledge, because expected sensory features of the speech input are enhanced and unexpected sensory features are suppressed. In contrast, the Prediction Error model assumes that the expected part of the speech input is explained away (i.e., reduced) and only Prediction Errors (i.e., the difference between heard and expected speech) are represented in the STS. To test these two simulations, we assessed the neural representation of speech information by means of RSA [50]. This approach allowed us to quantify the amount of information about the phonetic form of speech that is carried by the spatial pattern of fMRI activity in each of our four critical conditions.
We designed our experiment to test for categorical representations of syllable similarity, because previous studies (in fMRI [55,57] and intracranial recordings [61]) showed that categorical representations of speech, such as vowels and syllables rather than acoustic cues, are decodable from the STS. Neural representational similarity was first measured by computing a representational dissimilarity matrix (RDM) for multivoxel fMRI responses for each item and condition (see Materials and Methods for details). To quantify the amount of speech information, we computed the Fisher-z-transformed Spearman correlation between the observed RDM and a hypothesised RDM of interest that tested for increased similarity between pairs of syllables that shared the same vowel and had other segments in common (e.g., “sing” and “thing”) compared to pairs of unrelated words (e.g., “sing” and “bath”, see Fig 4A). This similarity measure was computed separately for each condition. This analysis targets speech representations in the posterior STS by testing for similarity of words that have similar phonetic forms but different lexical or semantic representations. We did not compare identical words presented in different scanning sessions.
To test our two computational simulations of spoken word recognition, we applied the same multivariate analysis to representations of the sensory input in the Sharpened Signal and Prediction Error models for each of our four conditions (for details, see Materials and Methods). As for the multivoxel fMRI RSA, we quantified the difference in pattern similarity between pairs of similar and dissimilar syllables (e.g., “sing” and “thing” versus “sing” and “bath;” see Fig 4A). The simulation for the Sharpened Signal model showed increased similarity for both increased sensory detail and matching prior information (Fig 4C). In contrast, the simulation for the Prediction Error model showed an interaction between sensory detail and prior information (Fig 4D). Specifically, there was greater pattern similarity for similar syllable pairs in the Neutral 12-channel than in the Neutral 4-channel condition, whereas in the Match 12-channel there was less pattern similarity than in the Match 4-channel condition. This outcome resembles the interaction of sensory detail and prior knowledge shown for multivariate fMRI results in the posterior STS ROI (Fig 4B). In addition, we repeated the cross-subject consistency analysis on representations generated by individual simulated participants. For the Prediction Error but not for the Sharpened Signal model, this showed the same crossover interaction of sensory detail and prior knowledge as in the equivalent fMRI analysis, suggesting a common underlying explanation (see S1 Text and S6A–S6C Fig).
Again, we used the evidence ratio of Akaike weights to compare the evidence for both models given the pattern similarity results in the left posterior STS (see Materials and Methods). Importantly, both models used parameters optimised to simulate the behavioural and univariate fMRI results, and no modifications or parameter optimisation were performed when simulating similarity in spatial patterns of fMRI activity. For the multivariate fMRI results, the evidence ratio of the Akaike weights revealed that the multivariate fMRI patterns very strongly supported the Prediction Error model over the Sharpened Signal model (wPE/wSharp = 1.898 x 1011, tested based on the independent ROI in the posterior STS [57]). Hence, computational simulations provided compelling evidence that multivariate fMRI results are more consistent with computation of Prediction Errors than of Sharpened Signals in the posterior STS during the perception of degraded speech.
We used multiple approaches (behavioural, computational, univariate, and multivariate fMRI) to investigate how prior expectations improve perception of degraded speech in order to distinguish Sharpened Signal and Prediction Error computations. Our experimental findings, first of all, replicate the existing literature [31–33,65] by showing that behavioural report of degraded words was improved both by matching expectations and by increased sensory detail (Fig 3A). Second, we show that matching expectations reduced BOLD activity during speech processing in left posterior STS (Fig 3B and 3C). Like other previous observations in the literature [8,39,43,45,46,66], these findings are in line with either Sharpened Signal or Prediction Error computations for combining prior knowledge and sensory input. This is confirmed by our computational simulations, which show that a good fit to behavioural and univariate fMRI data is achieved by models that include either of these two coding schemes (Fig 3D–3G). These model simulations are also consistent with the proposal that BOLD responses in the Match condition are lower because word identification is easier (as suggested by the behavioural improvements we observed). More informative results come from fMRI multivoxel pattern similarity, which revealed an interaction between prior knowledge and sensory detail in the posterior STS (Fig 4B). Specifically, for degraded speech that follows neutral expectations, increased sensory detail improved the amount of sensory information contained in fMRI multivoxel patterns. However, for speech that matched expectations, increased sensory detail led to a reduction in the amount of information represented in the posterior STS as measured by similarity analysis. This interaction is uniquely consistent with a Prediction Error model in which expected sensory input is explained away, and deviations from expectation are represented as Prediction Errors (Fig 1B). Our results, therefore, provide evidence for computation of Prediction Errors but not of Sharpened Signals (see simulations in Fig 4C and 4D).
Why is this interaction between sensory detail and prior knowledge shown in multivariate representations of speech so diagnostic of Prediction Error computations? In explaining this interaction, we will first consider the situation in which listeners have uninformative prior expectations. In the absence of specific expectations (as in the Neutral condition), both Sharpening and Prediction Error accounts propose that the amount of sensory information represented in neural patterns should increase with the amount of sensory detail in the input. In Prediction Error schemes, the brain does not pass forward the sensory input directly, but rather the discrepancy between expectations and sensory input. These Prediction Errors will provide an informative representation of the sensory input if these expectations are uninformative and the sensory input is sufficiently clear. Thus, our observation of enhanced coding for Neutral 12-channel compared to Neutral 4-channel stimuli is equally consistent with Prediction Error as with the traditional view that the brain directly represents the sensory input. The true test of Prediction Error schemes is provided by conditions in which specific and accurate expectations guide perceptual processing.
The hallmark of Prediction Error in our data is that for speech that matches prior expectations increasing the sensory detail reduces the informativeness of multivariate representations (Fig 1B). This is a counterintuitive finding, because clear speech that matches a previously presented written word (our Match 12-channel condition) is most accurately perceived, whereas multivariate representations are more informative in the less intelligible Match 4-channel condition. This is to be expected because Prediction Errors will be substantially reduced for conditions in which sensory input matches prior knowledge. Hence, increases in sensory detail lead to a better correspondence between sensory input and listeners’ prior expectations of clear speech. Our observation of reduced representation of speech content for Match 12-channel compared to Match 4-channel stimuli is entirely consistent with Prediction Errors but stands in marked contrast to the outcome expected for Sharpened Signals—or, indeed, any account in which sensory representations directly encode perceptual outcomes. Low pattern similarity for the condition with the clearest perceptual outcome (Match 12-channel) might appear surprising given previous findings that perceptual representations can be decoded from low-level response patterns [67–69]. However, these findings can be reconciled with Prediction Error schemes by recalling that these previous experiments used presentation conditions similar to the Neutral condition in our experiment (i.e., an uninformative prior).
Prediction Error can also explain the apparent increase in the informativeness of speech representations in the Match 4-channel condition compared to the Neutral 4-channel condition. Our simulations reveal that when sensory signals are severely degraded (such as for 4-channel vocoded speech), informative Prediction Errors are derived from the residual of matching expectations (in the Match condition). A specific expectation, as provided by our written word cue, when combined with a less informative stimulus, remains “unfulfilled” and is therefore represented as a negative but informative Prediction Error. Informative Prediction Errors (either positive or negative) are absent when prior expectations are uninformative (in the Neutral condition). Hence, both Prediction Error and Sharpened Signal models can explain our observation of increased representation of 4-channel speech that matches prior expectations. Other similar studies in the literature have explored whether visual representations of expected stimuli are sharpened or reduced [20,26] but have yielded contradictory findings. While Prediction Error was supported by univariate hemodynamic responses to unexpected classes of visual stimuli (faces versus houses, [26]), multivariate responses supported sharpening of expected visual gratings [20]. Two other differences between our work and this previous multivariate study are noteworthy. First, we separated the neural response to the cue (written word) and stimulus (spoken word). Second, we tested the interaction between sensory information and prior knowledge. Only Prediction Error can explain the full interaction of sensory detail and prior knowledge described above.
By the Prediction Error scheme, there should be a negative correlation between neural representations in the Neutral 12-channel condition (a positive Prediction Error) and the Match 4-channel condition (a negative Prediction Error, apparent in positive and negative Prediction Error in Fig 1B). However, an additional analysis of the present data showed that there was neither a negative nor a positive correlation between these conditions (we tested for a positive correlation because a negative Prediction Error could evoke a positive hemodynamic response due to metabolic costs of neural inhibition). These null findings cannot rule out the possibility that both conditions do indeed contain complementary information based on positive and negative Prediction Errors. Direct neural data (e.g., from intracranial recordings) might provide a more sensitive test of this proposal. Taken to an extreme (i.e., without any sensory input), computation of negative Prediction Errors could also explain previous results showing that the omission of an expected stimulus causes an increased signal [70–72] from which stimulus identity can be decoded [72–74].
Current Predictive Coding theories suggest that cortical regions involved in sensory processing contain two subpopulations of neurons: (1) prediction error units that represent the unexpected part of the incoming sensory information and (2) prediction units that represent the expected part of the incoming sensory information (and can be sharpened by matching prior expectations) [3,24,75]. These models have thus drawn support from empirical evidence showing either Prediction Errors [26,39,49,76] or Sharpened Signals [20] by attributing neural responses to prediction error and prediction units, respectively. Our goal in this study was to test two functionally distinct coding schemes in isolation by building computational models in which a simulated cortical area passes only one type of information forward (only Prediction Errors or Sharpened Signals). In the context of these simulations, our results provide clear evidence for representations of Prediction Errors. However, our multivariate fMRI findings do not oppose theories of Predictive Coding that propose Sharpened Signals coded by prediction units in addition to Prediction Errors in prediction error units [3,23,24]. The absence of evidence for Sharpened Signals in our data from the STS could be explained by previous proposals that fMRI measurements are dominated by responses from prediction error units (as [26,77] have argued for visual cortex). It could be that other neural measures, such as neurophysiological recordings with depth electrodes [78] or laminar-specific ultra-high field strength fMRI [79,80] are better able to detect responses from prediction units and could provide evidence of laminar-specific representations of Prediction Errors and Sharpened Signals.
Nonetheless, the interaction observed in the present study favours Predictive Coding theories (with representations of Prediction Error) over the traditional view that the brain directly passes forward the sensory input, as hypothesised in a Sharpening scheme without representations of Prediction Error. Our simulations show that in Sharpening schemes, the Match 12-channel condition should contain the clearest representation of speech content. This was not observed in the present data (compare Fig 4B and 4C). Our work not only provides evidence to support the hypothesis that integration of prior expectation and perceptual input for speech is achieved through computation of Prediction Errors or Sharpened Signals, but also introduces a new and critical diagnostic finding for Prediction Error responses: For unexpected stimuli, increased sensory detail should improve the amount of sensory information contained in neural patterns. However, for stimuli that match expectations, increased sensory detail should lead to a reduction in the amount of information represented. Future studies in other sensory modalities and domains might benefit from adopting similar methods.
Our work joins a number of recent fMRI and MEG/EEG studies in proposing an important role for Prediction Error computations in speech perception [4,7,8,39,81]. In these earlier studies, the observation of decreased activation for expected stimuli in the STG has been interpreted as a neural correlate of reduced Prediction Error and, hence, as evidence for Predictive Coding theories. However, almost all established computational theories of speech perception can also explain this observation. For example, TRACE [34] implements a form of neural sharpening in which prior knowledge enhances the representation of expected sensory signals and suppresses sensory noise, producing a reduced neural response overall. Similar, interactive activation models [35–38] might predict exactly the same decrease in STG activity for expected stimuli, as observed in these previous neuroimaging studies. Thus, existing empirical evidence proposed for Predictive Coding is also largely consistent with Sharpening theories. Even our previous comparison of Predictive Coding and Lexical Competition accounts of spoken word recognition [39] challenged the competitive lexical selection mechanism implemented in TRACE, but did not test the Sharpening mechanism traditionally described as Interactive Activation.
In this context, then, the results of our study have important implications for understanding speech perception, a domain in which the presence and function of top-down processes has been much debated [82,83]. By directly quantifying the information represented in multivariate signals during perception of correctly expected and unexpected speech, we provided evidence that the neural mechanisms underlying speech perception are in line with Prediction Error simulations. Prior knowledge of speech content is used to explain away sensory evidence such that speech representations encode Prediction Error.
The present multivariate interaction of sensory detail and prior information supports a Predictive Coding theory for how matching expectations improve perception of degraded speech. In contrast, enhanced representation of attended compared to unattended speech supports Sharpening mechanisms [84–87]. These findings could be reconciled by theories proposing that expectation (Prediction Error) and attention (Sharpening) operate in parallel, as suggested in some Predictive Coding theories [3,88]. However, more detailed computational specification of attentional mechanisms will be required to test these theories with experimental data. Comparing neural representations of attended and unattended speech signals at varying levels of expectation and degradation may be informative.
There are three reasons why our results are of general interest for the study of speech and other domains of perception. One key aspect of our approach is that we assessed the perception of speech presented at varying levels of signal degradation. As in accounts proposing Bayesian perceptual inference [89], this provides the best opportunity to observe influences of prior knowledge on perception. In doing so, we also test the perception of speech in listening conditions similar to the way that speech is most often heard in the real world [90]. A second form of generality is that prior expectations for speech were derived from written text. Our results may therefore also inform other situations in which prior knowledge and sensory information are combined across different modalities for speech [91–93] and other cross-modal stimuli [94–96]. Third and perhaps most important, however, is that the representations of Prediction Error that we have observed during speech perception might apply to many other sensory domains in which prior knowledge has been shown to influence perception (such as audition [6,7,76,97], vision [9–12,20,98,99], touch [13], gustation [14,100], olfaction [15], and pain [16]). The interactive effect of prior knowledge and sensory input on neural representation of degraded stimuli provides a stronger test of Predictive Coding theories of perception than has been provided by existing methods, as it offers the potential to challenge alternative views based purely on Sharpening mechanisms.
In summary, the present results show that both increased sensory detail and matching prior expectations improved accuracy of word report for degraded speech but had opposite effects on speech coding in the posterior STS. Following neutral text, increased sensory detail enhanced the amount of speech information, whereas matching prior expectations reduced the amount of measured information during presentation of clearer speech. These findings support the view that the brain reduces the expected and, therefore, redundant part of the sensory input during perception, in line with representations of Prediction Error proposed in Predictive Coding theories.
Ethical approval was provided by Cambridge Psychology Research Ethics committee (CPREC) under approval number 2009.46. All participants provided their written informed consent.
Twenty-five healthy native-English speakers (aged 18–40, with self-reported normal hearing and language function) participated in the experiment. Three participants had to be excluded because they were insufficiently attentive to the written text during the scanning runs (they reported less than 50% of the written words correctly when prompted). One additional participant had to be excluded due to technical problems. The reported analyses are therefore based on 21 participants (mean age 25 y [range 19 to 38 y], 9 females).
Word stimuli consisted of 24 different monosyllabic words, each with a consonant-vowel-consonant structure. The words were selected as eight triples of three similar words, each sharing the same vowel and with offset and onset changes between items (eight triples: thing/sing/sit, bath/path/pass, deep/peep/peak, pork/fork/fort, doom/tomb/tooth, take/shake/shape, kite/tight/type, zone/moan/mode). These stimuli were recorded by a male native speaker of Southern British English and noise-vocoded (4- and 12-channel) using custom scripts written in Matlab [59]. The syllables were filtered into 4 or 12 approximately logarithmically spaced frequency bands from 70 to 5,000 Hz [101], with each pass band 3 dB down with a 16 dB/octave roll off. In each band, envelopes were extracted using half wave rectification, and pitch synchronous oscillations above 30 Hz were removed with a second-order Butterworth filter. The resulting envelopes were multiplied with a broadband noise and then band pass filtered in the same frequency ranges as the source and recombined. To ensure that acoustic intensity was matched across all stimuli, the RMS amplitude of each sound file was equalised. Finally, we applied an additional filter to ensure a flat frequency response when the spoken words were presented via Sensimetrics insert headphones in the scanner (http://www.sens.com).
Participants read written words and listened to subsequently presented degraded spoken words (see Fig 2). There were four conditions containing different pairings of written and spoken words: (1) matching written text + spoken words (“SING” + sing); (2) neutral written text (“XXXX”) + spoken words (sing); (3) partially mismatching written text + spoken words (“SIT” + sing); (4) totally mismatching written text + spoken words (“SING” + doom). In addition, we included a fifth condition in which only written text (“SING”) was presented to test whether participants attended to the written words. Only the match and neutral conditions (condition 1 and 2) were repeated sufficiently (six presentations per item per condition) to permit multivariate RSA (see below for details). In occasional catch trials, a response cue, which consisted of a visual display of a question mark, was presented 1,000 ms after trial onset. This cued participants to say aloud the written or spoken word that they saw or heard previously. This design does not allow the analysis of response times, because participants were cued to respond after a delay. A previous behavioural study in our lab showed that response times for reporting vocoded spoken words are uninformative even when collected in such a way as to permit response time analyses [102]. The partial and total mismatch conditions (condition 3 and 4) were included to make sure that participants paid attention to both the written and the spoken word; these conditions ensured that they could not simply report the preceding written word. Due to the small number of trials, RSA analysis was not possible for neural responses measured in the Mismatch condition. We can, however, report behavioural and univariate fMRI results for the Mismatch condition; this confirms that behavioural and neural enhancement following matching written text is not due to prestimulus attention or anticipation (because prestimulus processes will be identical following mismatching text but enhanced perception is not typically observed) [8,33].
Trials commenced with presentation of a fixation cross (1,000 ms), followed by presentation of a written word (500 ms), again followed by a fixation cross (500 ms), and finally the presentation of a spoken word. Written cues (i.e., written words, neutral “XXXX”, and fixation cross) were presented in grey in the centre of the black screen. Trials were 3 to 9 s long, depending on the number of inserted null events to decorrelate the events within each run (76 trials of 3 s without null event, 45 trials of 6 s with a null event of 3 s, and 15 trials of 9 s with a null event of 6 sec, resulting in 211 TRs per run with null events).
Spoken words were presented after 4- or 12-channel noise-vocoding to produce two different levels of sensory detail in the speech input. Altogether, this resulted in 816 trials, including 1/6 catch trials (136 trials) in which participants had to give their verbal response (24 neutral and 24 match words x 6 repetitions x 2 levels of sensory detail = 576 trials, 24 written-only words x 6 repetitions = 144 trials, 24 partial mismatch and 24 total mismatch words x 2 levels of sensory detail without repetition on the word level = 96 trials; i.e., 11.8% of the trials contained mismatching information). These trials were split into 6 runs of 136 trials each, ensuring that each word in each condition occurred once in each scanning run. With additional catch trials, each run took 11.7 min, and the overall experiment lasted approximately 70 min for all 6 runs. Stimulus delivery was controlled and behavioural responses were recorded with E-Prime 2.0 software (Psychology Software Tools, Inc.).
Verbal responses recorded in the scanner were transcribed by two independent raters (the first author and a native English speaker with a PhD in phonetics who was naïve to the stimulus set) and disagreements adjudicated by a third rater (the senior author). All raters were blind to which word and stimulus condition was presented in each trial. Responses were scored for whole-word accuracy and analysed using Matlab. Because the percent correct performance scores were bound to [0;1], we applied an arcsine transformation [103] before we computed a two-way repeated measures ANOVA and the corresponding post-hoc pared t tests.
Data were analysed using SPM8 (http://www.fil.ion.ucl.ac.uk/spm) applying automatic analysis (aa) pipelines [104]. The first three volumes of each run were removed to allow for T1 equilibrium effects. Scans were realigned to the first EPI image. The structural image was coregistered to the mean functional image and the parameters from the segmentation of the structural image were used to normalise the functional images, which were resampled to 2 mm isotropic voxels. The realigned normalised images were then smoothed with a Gaussian kernel of 8 mm full width half maximum. Data were analysed using the general linear model with a 128 s high pass filter. We included the onset of 11 event types in the GLM, each convolved with the canonical SPM haemodynamic response: eight conditions come from specifying the onset of spoken words paired with four types of written text (matching, neutral, partially mismatching, and totally mismatching) crossed with two types of vocoding (4- and 12-channel). We also specified onsets for written words and neutral strings (“XXXX”) as well as the onset of the visual task cue that instructed participants to say the spoken word. Following parameter estimation of the first level model, we conducted a repeated measures ANOVA with two factors: prior knowledge (matching versus neutral text) and level of sensory detail (4- versus 12-channel) to assess the main effects and interaction of these factors.
We were interested in the effect of hearing speech that matches prior expectations on BOLD responses in the left posterior STS. To locate these ROIs for the multivoxel RSA (see below), we tested for a main effect of prior knowledge (F-contrast “Neutral versus Match”) and identified a cluster at p < 0.05 FWE voxel-corrected in the left posterior STS.
Multivariate analyses were conducted on realigned data within each participant’s native space without normalisation or spatial smoothing. An additional first-level model was constructed for each participant that contained the same set of regressors as the first level model used for the univariate analysis, except that regressors for individual spoken words were used in each of the four conditions for which there were sufficient numbers of repetitions for item-specific modelling (4- and 12-channel vocoded words following neutral or matching text). This resulted in 103 conditions per participant per run: 24 words for each of these four conditions and the remaining seven conditions from the univariate model. For each of the 96 item-specific regressors in these four conditions, we estimated single-subject T-statistic images for the contrast of speech onset compared to the unmodelled resting period, averaged over the six scanning runs.
We used the resulting single condition and item T-images for RSA [50] using the RSA toolbox [52]. We used T-images so that effect sizes were weighted by their error variance, which reduces the influence of large but variable response estimates for multivariate analyses [105]. RSA involves testing whether the observed similarity of brain responses in specific conditions (a neural RDM) corresponds to a hypothetical pattern of similarity between these conditions (hypothesis RDM). We constructed four hypothesis RDMs to test for greater similarity between syllable pairs within the same stimulus triple (i.e., syllables that shared the same vowel and had similar onset or offset segments like “sing” and “thing,” as compared to dissimilar syllables like “sing” and “bath”) within each of four critical conditions: Match 4-channel, Neutral 4-channel, Match 12-channel, and Neutral 12-channel. The design of our experiment was motivated by previous work that showed that STS encodes vowel and syllable similarity [55,61], rather than spectrotemporal acoustic cues [61]. The comparisons used in our ROI analysis test for global similarity in representations of the phonetic form of similar-sounding spoken words because multiple consonantal features as well as the vowel are preserved within each syllable triple (e.g., bath/path/pass). We chose to analyse similarity of neural representations for phonetically similar but non-identical words for two reasons: (1) this approach allowed us to merge all six runs into a single analysis, which reduced the noise in the estimation of the T-images relative to a split-half method, and (2) comparing similar but non-identical word pairs makes our method insensitive to other forms of lexical or semantic similarity that could lead to similar neural representations for identical word pairs (e.g., in regions that code for word meaning [106]). Similarity between items in different conditions and between identical items (i.e., the main diagonal) was therefore not included in our hypothesis RDMs (see Fig 4A).
We measured multivoxel RDMs by computing the dissimilarity (1–Pearson correlation across voxels) of T-statistics for a specific item and condition. In a searchlight analysis, the sets of voxels were extracted by specifying grey-matter voxels (voxels with a value > 0.33 in a probabilistic grey-matter map) within an 8-mm radius sphere of each grey matter voxel (with a voxel size of 3 x 3 x 3.75 mm, i.e., a maximum of 65 voxels per sphere). This was repeated for all searchlight locations in the brain. The similarity between the observed RDM and each of the hypothetical RDMs was computed using a Spearman correlation for each searchlight location, and the resulting correlation coefficient returned to the voxel at the centre of the searchlight. This resulted in a Spearman correlation map for each participant in each grey matter voxel. To assess searchlight similarity values across participants at the second level, the Spearman correlation maps for each participant were Fisher-z-transformed to conform to Gaussian assumptions, normalized to MNI space, and spatially smoothed with a 10-mm FWHM Gaussian kernel for group analysis. These second-level analyses used a within-subject analysis of variance similar to those used for the univariate fMRI analysis.
We used two computational implementations of Sharpened Signal and Prediction Error models of spoken word recognition (using update mechanisms based on [75]), to simulate observed behavioural performance (i.e., word recognition), univariate fMRI results (the magnitude of hemodynamic activity in the STS), and RSA fMRI results (the similarity of representations for word pairs in the left posterior STS) in each of our four experimental conditions. The sensory representations supplied at the input, the output lexical representations, and the specification of matching or neutral prior knowledge was identical for both simulations. We used a localist lexical representation (i.e., a set of 24 units, each of which was activated to represent a single word), as in previous models of spoken word recognition such as TRACE [34] or Shortlist [107]. The input to the model was provided as a distributed set of phonetic features (derived from [108]). These are similar to the acoustic/phonetic features supplied as the input to TRACE or in recurrent network simulations such as the Distributed Cohort Model [109]. However, to avoid the complexity of representing temporal information (and given the slow haemodynamic responses measured by fMRI), we assumed that speech information is provided in parallel over three groups of units for the initial consonant, medial vowel, and final consonant of our CVC words.
The key difference between the Sharpened Signal and Prediction Error models concerns the computations by which prior knowledge is combined with degraded sensory representations of expected spoken words. In the Sharpened Signal simulation, expected sensory features receive additional activation through increased sensory gain [19,20], whereas in the Prediction Error model, prior expectations contribute to perception by subtracting expected input from sensory representations (i.e., computation of Prediction Error [3,23,24]). In both simulations, an iterative settling procedure was used such that feature representations of the input are combined with prior knowledge to generate feature representations that convey Sharpened Signals or Prediction Errors respectively (hereafter “sharpened features” and “prediction error features”). These representations were used to update lexical activations, and updated lexical activations in turn led to modified top-down expectations. This settling procedure continued until a settling criterion was reached or a maximum number of iterations had been performed.
|
10.1371/journal.pbio.0050004 | Biosynthesis of Selenocysteine on Its tRNA in Eukaryotes | Selenocysteine (Sec) is cotranslationally inserted into protein in response to UGA codons and is the 21st amino acid in the genetic code. However, the means by which Sec is synthesized in eukaryotes is not known. Herein, comparative genomics and experimental analyses revealed that the mammalian Sec synthase (SecS) is the previously identified pyridoxal phosphate-containing protein known as the soluble liver antigen. SecS required selenophosphate and O-phosphoseryl-tRNA[Ser]Sec as substrates to generate selenocysteyl-tRNA[Ser]Sec. Moreover, it was found that Sec was synthesized on the tRNA scaffold from selenide, ATP, and serine using tRNA[Ser]Sec, seryl-tRNA synthetase, O-phosphoseryl-tRNA[Ser]Sec kinase, selenophosphate synthetase, and SecS. By identifying the pathway of Sec biosynthesis in mammals, this study not only functionally characterized SecS but also assigned the function of the O-phosphoseryl-tRNA[Ser]Sec kinase. In addition, we found that selenophosphate synthetase 2 could synthesize monoselenophosphate in vitro but selenophosphate synthetase 1 could not. Conservation of the overall pathway of Sec biosynthesis suggests that this pathway is also active in other eukaryotes and archaea that synthesize selenoproteins.
| Biosynthesis of the 20 canonical amino acids is well established in eukaryotes. However, many eukaryotes also have a rare selenium-containing amino acid, selenocysteine, which is the 21st amino acid in the genetic code. Selenium is essential for human health, and its health benefits, including preventing cancer and heart disease and delaying aging, have been attributed to the presence of selenocysteine in protein. How selenocysteine is made in eukaryotes has not been established. To gain insight into its biosynthesis, we used computational analyses to search completely sequenced genomes for proteins that occur exclusively in organisms that utilize selenocysteine. This approach revealed a putative selenocysteine synthase, which had been previously identified as a pyridoxal phosphate–containing protein dubbed soluble liver antigen. We were able to characterize the activity of this synthase using selenophosphate and a tRNA aminoacylated with phosphoserine as substrates to generate selenocysteine. Moreover, identification of selenocysteine synthase allowed us to delineate the entire pathway of selenocysteine biosynthesis in mammals. Interestingly, selenocysteine synthase is present only in those archaea and eukaryotes that make selenoproteins, indicating that the newly defined pathway of selenocysteine biosynthesis is active in these domains of life.
| Selenocysteine (Sec) is a selenium-containing amino acid that is cotranslationally inserted into protein and is recognized as the 21st amino acid in the genetic code [1–3]. Sec is incorporated into protein in all three lines of descent, eukaryota, archaea, and eubacteria, but unlike other amino acids, Sec synthesis occurs on its transfer RNA (tRNA), designated tRNA[Ser]Sec [4,5]. tRNA[Ser]Sec is initially aminoacylated with serine by seryl-tRNA synthetase and the seryl moiety provides the backbone for Sec synthesis. The biosynthesis of Sec was established in Escherichia coli in the early 1990s [6–8]. Bacterial Sec synthase (SecS) (E. coli selenocysteine synthase [SelA]) is a pyridoxal phosphate (PLP)-dependent protein that converts the serine attached to tRNA[Ser]Sec to Sec by initially removing the hydroxyl group from serine to form an aminoacrylyl intermediate. This intermediate serves as the acceptor for activated selenium, and when selenium is donated, selenocysteyl-tRNA[Ser]Sec is formed. The active selenium donor in bacteria is synthesized from selenide and ATP by E. coli selenophosphate synthetase (SelD), and the product of the reaction has been identified as monoselenophosphate (SeP) [9].
A distant homolog of bacterial SelA (SelA-like) is present in some archaea but is not active as SecS [10], and it does not always co-occur in archaea with Sec insertion systems. In addition, no SelA sequences could be detected in eukaryotes. Although Sec insertion systems are different in bacteria from those in archaea and eukaryotes [11–13], several factors have been characterized in mammals that most certainly have a role in Sec biosynthesis. For example, the soluble liver antigen (SLA) was initially identified as a 48-kDa protein bound to Sec tRNA[Ser]Sec that was targeted by antibodies in patients with an autoimmune chronic hepatitis [14]. SLA was subsequently reported to exist as a separate family within a larger superfamily of diverse PLP-dependent transferases [15], and this protein has been proposed to function as the mammalian SecS (e.g., see [3,15–17]). Further evidence that SLA is involved in selenium metabolism is that it was found to occur in a protein complex with other factors involved in the biosynthesis of Sec and/or its insertion into protein [17,18]. In addition, a kinase that phosphorylated a minor seryl-tRNA was reported in 1970 [19] that was subsequently isolated, characterized, and found to specifically phosphorylate the seryl moiety on seryl-tRNA[Ser]Sec [20]. The resulting phosphoseryl-tRNA[Ser]Sec was proposed either as a candidate substrate for SecS (see [3,20] and references therein) or it served as a storage form [21] . Furthermore, two genes initially thought to have a role in selenophosphate synthesis, sps1 and sps2, have been reported in mammals [22–25], and the product of sps2 is a selenoprotein, SPS2 [22,24]. The Sec-to-Cys mutant form of SPS2 has low enzyme activity [22,24,26] and can complement SelD in Escherichia coli cells transfected with the mammalian mutant form [26]. Complementation of SelD− E. coli cells with SPS1 or SPS2 has suggested that SPS1 may have a role in recycling Sec via a selenium salvage system and SPS2 may be involved in the de novo synthesis of selenophosphate from selenide [27]. However, it should be noted that, to our knowledge, selenophosphate has never been shown to serve directly as the active selenium donor in Sec biosynthesis in eukaryotes.
Herein, we used a comparative genomics search and experimental analyses to show that SLA is the mammalian SecS. This protein belongs to a different family of PLP-containing enzymes and uses O-phosphoseryl-tRNA[Ser]Sec rather than seryl-tRNA[Ser]Sec as substrate. SecS dephosphorylates O-phosphoseryl-tRNA[Ser]Sec and accepts the active selenium donor to yield selenocysteyl-tRNA[Ser]Sec. We also demonstrated unequivocally that the selenium donor in eukaryotes is SeP by using this compound as a substrate in a reaction with SecS and phosphoseryl-tRNA[Ser]Sec. Selenophosphate is indeed synthesized in mammals by SPS2, whereas the distant homolog of SelD in mammals, SPS1, did not synthesize the active selenium donor. Conservation of the overall pathway of Sec biosynthesis suggests that it is also active in other eukaryotes and archaea.
We analyzed completely sequenced genomes of eukaryotes and archaea for the occurrence of selenoproteins. Twenty-six eukaryotes and three archaea that had these proteins and 24 eukaryotes and 24 archaea that did not were identified. Comparative genomics studies were then carried out to identify genes that co-occur with selenoproteins in (1) eukaryotes (Table 1) and (2) archaea (Table 2). Each of the searches had known components of Sec insertion machinery as top candidates as well as an additional protein, herein designated as SecS. In mammals, SecS is also known as SLA. SLA was first detected as an autoimmune factor that coimmunoprecipitated tRNA[Ser]Sec from cell extracts in patients with autoimmune chronic hepatitis [14], and it also bound other Sec insertion components [17,18]. SecS formed a separate family within a larger superfamily of diverse PLP-dependent proteins and was previously suggested to convert a tRNA-bound serine to Sec [15]. We found that it occurred exclusively in both eukaryotes and archaea which had selenoproteins but was lacking in the other organisms examined (Figure 1 and Tables 1 and 2). These observations strongly suggested that SecS may be the missing SecS in eukaryotes and archaea. Based on the multiple sequence alignment and phylogenetic analysis of SecS and other PLP-dependent proteins, including SelA, SelA-like, and SepCysS, it was clear that bacterial SelA and archaeal SelA-like proteins [10], on one hand, and SecS, on the other, belonged to completely different families of PLP-containing proteins (Figures 1 and 2), suggesting that their similar functions arose by convergent evolution. SecS was also distantly homologous to SepCysS, a protein recently found to synthesize cysteine from phosphoserine in some archaea [28] (Figures 1 and 2). After identifying a likely SecS candidate by comparative genomics analysis, we experimentally verified its function as described below.
To elucidate Sec biosynthesis in mammals, we initially examined the ability of tRNA[Ser]Sec, seryl-tRNA[Ser]Sec, and O-phosphoseryl-tRNA[Ser]Sec to bind to the recombinant mouse SecS (mSecS). The coprecipitated product was detected by Northern blotting (Figure 3A) and the amount of binding was quantitated (Figure 3B). O-Phosphoseryl-tRNA[Ser]Sec bound more efficiently to mSecS than the other tRNA[Ser]Sec forms, while seryl-tRNA[Ser]Sec bound least efficiently, suggesting that O-phosphoseryl-tRNA[Ser]Sec may be a substrate for mSecS. It is not clear why tRNA[Ser]Sec binds to mSecS (see also [18]), albeit less efficiently than O-phosphoseryl-tRNA[Ser]Sec. Seryl-tRNASer and tRNASer, however, did not manifest any binding to mSecS (unpublished data).
To assess whether the phosphate moiety on O-phosphoseryl-tRNA may be removed by mSecS to generate an intermediate that serves as an acceptor for the active selenium donor, the 32P-labeled form of O-phosphoseryl-tRNA[Ser]Sec was incubated with mSecS (Figure 3C). mSecS removed the phosphoryl moiety from O-phosphoseryl-tRNA[Ser]Sec (see lane 2). Interestingly, SelA was also capable of dephosphorylating O-phosphoseryl-tRNA[Ser]Sec (lane 3). Neither mSPS2-Cys [mouse selenophosphate synthetase 2 containing an Sec (UGA)-to-Cys (UGC) mutation] nor SelD appeared to have any effect on O-phosphoseryl-tRNA[Ser]Sec (lanes 4 and 5). The dephosphorylation of O-phosphoseryl-tRNA[Ser]Sec by mSecS and SelA is further considered below. However, it should be noted that the data in Figure 3 strongly suggest that the dephosphorylated product is not seryl-tRNA[Ser]Sec as the product binds efficiently to mSecS but seryl-tRNA does not (Figure 3B).
We next identified the active selenium donor by assessing whether mSPS1 and mSPS2-Cys synthesized SeP. 31P NMR spectroscopic analysis of the products of the mSPS2-Cys–catalyzed reaction manifested a signal at +23.2 ppm, albeit weakly (Figure 4A1), that corresponded to SeP [9,29]. Since the mSPS2 used in this experiment was a Sec-to-Cys mutant and might not be expected to generate SeP efficiently, we cloned SPS2 from Caenorhabditis elegans, which naturally contains Cys in place of Sec at the presumed active site of SPS2 [13]. C. elegans SPS2 clearly generated a signal at +23.2 ppm (Figure 4A2). As expected, SeP was also formed in the presence of E. coli SelD, selenide, and ATP (Figure 4A3) [9,29]. However, no signal at +23.2 ppm was observed when mSPS1 replaced mSPS2-Cys or SelD in the reaction, indicating that mSPS1 did not synthesize SeP (Figure 4A4).
As the peak at +23.2 ppm was relatively weak in the product analysis of mSPS2-Cys (Figure 4A1), the ordinate and abscissa of the area between 15 and 30 ppm were expanded as shown in Figure 4B. Clearly, there was a peak at +23.2 ppm corresponding to SeP, demonstrating that mSPS2-Cys produced SeP. The signal for SeP was also evident with C. elegans SPS2 and with SelD, but mSPS1 did not produce this signal.
To further examine the hydrolysis of ATP by mSPS2-Cys, C. elegans SPS2, SelD, and mSPS1, each component was incubated with [α-32P]ATP with and without selenide (Figure 4C). Hydrolysis of ATP to AMP was largely dependent on the presence of selenide with the three enzymes, mSPS2-Cys, C. elegans SPS2, and SelD, that produced SeP (see above), and all three hydrolyzed ATP to ADP independently of selenide. Although mSPS1 hydrolyzed ATP to ADP and apparently only slightly to AMP, this degradation was independent of selenium. These data provide further evidence that mSPS1 cannot synthesize SeP from selenide.
Previous studies analyzing Sec biosynthesis did not utilize SeP to assess whether this compound served directly as the active selenium donor. We therefore examined the ability of SeP to donate selenium directly in Sec biosynthesis. Sec was indeed synthesized when SeP [9,29] was added in the reaction with O-phosphoseryl-tRNA[Ser]Sec and mSecS (Figure 5A). This observation confirms unequivocally that SeP is the active donor of selenium in Sec biosynthesis and that SecS is the missing SecS. Control assays demonstrated that Sec was not formed when SeP was omitted from the reaction, when seryl-tRNA was used in place of O-phosphoseryl-tRNA[Ser]Sec, or when another protein, thioredoxin (Trx), was substituted for mSecS in the reaction. As expected, a reaction consisting of SelA, seryl-tRNA[Ser]Sec, and SeP also synthesized Sec (Figure 5A). O-Phosphoseryl-tRNA[Ser]Sec could also serve as a substrate and replace seryl-tRNA[Ser]Sec in reactions with SelA and SeP, thus using the dephosphorylated product as an acceptor for activated selenium to synthesize selenocysteyl-tRNA[Ser]Sec.
Sec was also synthesized on tRNA[Ser]Sec when O-phosphoseryl-tRNA[Ser]Sec was incubated with mSecS, mSPS2-Cys, ATP, and selenide (Figure 5B). Control reactions demonstrated that Sec was not formed when selenide was omitted from the reaction, when seryl-tRNA[Ser]Sec was used in place of O-phosphoseryl-tRNA[Ser]Sec, or when Trx was substituted for mSecS (Figure 5B). SelA would substitute for mSecS when the substrate was seryl-tRNA[Ser]Sec or O-phosphoseryl-tRNA[Ser]Sec.
As expected, SelD could substitute for mSPS2-Cys in synthesizing SeP and generating similar amounts of Sec as mSPS2-Cys in reactions with mSecS (unpublished data), and those reactions in Figure 5B were dependent on ATP as well as selenide wherein Sec was generated (unpublished data).
Serine, alanine, and pyruvate were recovered from reactions with SelA using either seryl- or O-phosphoseryl-tRNA[Ser]Sec as substrates, wherein alanine and pyruvate were likely the deacylated, degraded products of the intermediate, aminoacrylyl-tRNA[Ser]Sec [7] (Figures 5A and 5B). In reactions with mSecS using O-phosphoseryl-tRNA[Ser]Sec as substrate, only a small amount of phosphoserine and a peak that comigrated with pyruvate were recovered as deacylated products, suggesting that pyruvate was, similar to the bacterial case, the deacylated, degraded product of the intermediate in Sec biosynthesis in eukaryotes. However, the amount of pyruvate recovered in reactions that were coupled with SeP synthesis by selenophosphate synthetase was lower (Figure 5B) than in reactions in which SeP was supplied directly as substrate (Figure 5A). Use of HSe− in these reactions required the addition of high levels of dithiothreitol (DTT) which were inhibitory to Sec synthesis (unpublished data), and apparently the intermediate that formed pyruvate as a deacylated product was unstable under these conditions. The intermediates in reactions with SelA [7] and mSecS are further considered in Discussion.
We also examined the rate of Sec synthesis with O-phosphoseryl-tRNA[Ser]Sec and SeP as substrates in the presence of mSecS (Figure 5C). As the substrate, O-phosphoseryl-tRNA[Ser]Sec, was labeled with [3H]serine and the deacylated products, O-phosphoserine, Sec and the degraded intermediate, pyruvate, migrated separately in the chromatographic system used in Figure 5A; the amounts of each could be assessed during the course of the reaction. Dephosphorylation occurred rapidly and appeared to be near completion in about 10 min. Sec synthesis increased rapidly during the initial 10 min and then appeared to proceed more slowly until completion at about 40 min. The approximate initial rate was 0.28 pmol Sec/min/pmol mSecS. Likewise, the intermediate formed rapidly during the initial stages of the reaction and then decreased over the remainder of the experiment.
In this work, we defined the pathway of Sec biosynthesis in eukaryotes. In order to carry out Sec biosynthesis, we functionally characterized two previously known enzymes, selenophosphate synthetase [22,23,25] and O-phosphoseryl-tRNA[Ser]Sec kinase [20], as well as establishing the function of SLA [14] as the eukaryotic SecS. All these enzymes were found to be required for in vitro biosynthesis of Sec, and the implications of these findings are discussed below.
The active selenium donor in bacteria is SeP [9,29] and it is synthesized from selenide and ATP by selenophosphate synthetase, also known as SelD [7,30]. Two homologs of bacterial SelD, designated SPS1 and SPS2, are present in mammals [22–25]. Interestingly, SPS2 is a selenoprotein. Direct roles of SPS1 and SPS2 in mammals have not been tested, but it was suggested that SPS2 supports the use of selenite, whereas SPS1 depends on a selenium salvage system when examined in E. coli [27]. Our results clearly demonstrate that SPS2 makes the active selenium donor, SeP, for the biosynthesis of Sec.
We recently identified the gene that phosphorylates seryl-tRNA[Ser]Sec and characterized the gene product, O-phosphoseryl-tRNA[Ser]Sec [20]. However, the precise role of O-phosphoseryl-tRNA[Ser]Sec was not determined. The current results showed that O-phosphoseryl-tRNA[Ser]Sec is the substrate of SecS, and therefore O-phosphoseryl-tRNA[Ser]Sec kinase is involved in the Sec biosynthesis pathway.
What is the intermediate produced by mSecS? Most certainly the dephosphorylated product of O-phosphoseryl-tRNA[Ser]Sec cannot be seryl-tRNA[Ser]Sec (Figure 3A). The intermediate could possibly be aminoacrylyl-tRNA[Ser]Sec (dehydroalanyl-tRNA[Ser]Sec) which could yield pyruvate on hydrolysis. The facts that the intermediate generated by SelA is aminoacrylyl-tRNA[Ser}Sec and that mSecS contains pyridoxal phosphate suggest that the Schiff base intermediate of aminoacrylyl-tRNA[Ser}Sec postulated for Sec synthesis in prokaryotes [7] is analogous to that formed in eukaryotes. However, it proved possible to trap the proposed aminoacrylyl-tRNA[Ser]Sec by reduction with KBH4 leading to the formation of alanine on hydrolysis in the prokaryotic, but not eukaryotic, case. This result may be due to differences between the enzyme to which the aminoacrylyl-tRNA[Ser]Sec is bound; that is, reduction can occur before hydrolysis in the prokaryotic, but not eukaryotic, case. Nevertheless, identification of the intermediate in Sec biosynthesis in eukaryotes must await further study.
The biosynthesis of Sec in eukaryotes is shown in Figure 6. tRNA[Ser]Sec is aminoacylated by seryl-tRNA synthetase and the seryl moiety is phosphorylated by O-phosphoseryl-tRNA[Ser]Sec kinase to form O-phosphoseryl-tRNA[Ser]Sec [20]. O-Phosphoseryl-tRNA[Ser]Sec is a substrate for SecS which replaces the phosphoryl moiety of phosphoserine, derived from the selenium donor, SeP, to yield Sec. SeP is synthesized by SPS2 in the ATP-dependent reaction. SecS does not use seryl-tRNA[Ser]Sec as a substrate (Figure 5A and 5B). Although no enzyme comparable to O-phosphoseryl-tRNA[Ser]Sec kinase has been identified in E. coli, it is of interest to note that SelA can utilize O-phosphoseryl-tRNA[Ser]Sec as a substrate. The major difference between the Sec biosynthetic pathway characterized herein and that in eubacteria is the extra step in the synthesis of O-phosphoseryl-tRNA[Ser]Sec which serves as a substrate for SecS. In E. coli, seryl-tRNA[Ser]Sec serves directly as the substrate for SelA [7]. The occurrence of SecS exclusively in selenoprotein-containing organisms in eukaryotes and archaea (Tables 1 and 2) indicates that the SecS-based pathway also operates in other animals, lower eukaryotes, and archaea where the Sec machinery occurs [3]. Considering the difficulties with identification of other components of Sec biosynthesis and insertion machinery (e.g., SBP2, EFsec), SecS might become the most characteristic feature of the Sec trait in eukaryotes and archaea.
[α-32P]ATP and [γ-32P]ATP (specific activity, approximately 6,000 Ci/mmol) and Hybond N+ nylon membranes were purchased from Amersham (http://www.amersham.com), 3H-serine (specific activity, 29.5 Ci/mmol) and 14C-pyruvate (specific activity, 19 mCi/mM) were from Perkin Elmer (http://www.perkinelmer.com), Ni-NTA agarose was from Qiagen (http://www.stratagene.com), and pfu polymerase and pBluescript II were from Stratagene (http://www.stratagene.com). pET32b vector (encoding the 109–amino acid thioredoxin with a His-tag) and BL21(DE3) competent cells were obtained from Novagen (EMD Biosciences, http://www.emdbiosciences.com), alkaline phosphatase from New England Biolabs (http://www.neb.com), T7 RiboMAX Express Large Scale RNA Production System and 3M filter paper from Whatman (http://www.whatman.com), and unlabeled amino acids, PEI TLC plates, and selenocystine from Sigma (http://www.signaaldrich.com). [(CH3)3SiO]3PSe was chemically synthesized [31]. All other reagents were commercial products of the highest grade available.
A total of 50 completely or nearly completely sequenced eukaryotic genomes (Table S1) were analyzed for occurrence of selenoproteins by TBLASTN using the set of all known selenoproteins. Twenty-six organisms were found to contain selenoproteins and 24 organisms lacked these proteins. In organisms lacking selenoproteins, the Sec insertion machinery was also missing. To identify proteins with phylogenetic profiles corresponding to selenoproteins, annotated genes in D. melanogaster were used as a query dataset. BLAST homology analyses were used to scan genomic databases using the following criteria: E-value less than 1e−06 and length of the conserved region greater than 50 amino acids. Genes present in any of the organisms lacking selenoproteins were dismissed. The remaining genes were searched against selenoprotein-containing organisms to determine their occurrences. Top candidate genes were included in Table 1; they were present in 80% of selenoprotein-containing eukaryotes. These candidates were further manually analyzed for possible function.
A similar search strategy was carried out in archaea. A total of 27 archaeal genomes (Table S1), including three selenoprotein-containing organisms (M. jannaschii, M. maripaludis, and M. kandleri), were analyzed for gene occurrence using all annotated genes in M. jannaschii. Top candidate genes are included in Table 2; these proteins were present in all three selenoprotein-containing archaea and absent in other completed archaeal genomes. We used ClustalW to generate multiple sequence alignment. Phylogenetic trees were built with PHYLIP programs.
The coding regions of E. coli selA and selD, mouse secS, sps2, sps1, and C. elegans sps2 genes were amplified from BL21 genomic DNA or mouse liver cDNA or C. elegans total cDNA using pfu polymerase, respectively [20]. The resulting product was cloned into the pET32b vector at the NdeI-XhoI cloning sites in which the vector contained a His-tag immediately downstream of, and in frame with, the open reading frame. The Sec TGA codon in sps2 was mutated to a Cys TGC codon using a site-directed mutagenesis kit (Stratagene), and the resulting gene product was designated mSPS2-Cys. The cDNA constructs were confirmed by sequencing and transformed into BL21(DE3) cells. Expression and purification of each protein were carried out as described [20]. For mSecS and SelA expression and purification, 1 nM PLP was added in the LB medium during expression and 5 μM PLP was added in the elution buffer during purification. The proteins were dialyzed against 1× TBS for 2 h and stored at −20 °C in 50% glycerol before use.
Native tRNA[Ser]Sec was purified and aminoacylated with serine and the seryl moiety phosphorylated as described [20]. Then 200 ng of purified mSecS containing a His-tag on its C-terminal was added in a total volume of 100 μl solution (20 mM Tris-HCl [pH 7.4], 0.01 mM EGTA, 1 mM DTT, 10 mM MgCl2, and 5 μg of yeast tRNA) and approximately 50 ng of purified tRNA[Ser]Sec (with either serine or phosphoserine attached, or no amino acid) added, and the reaction was incubated for 30 min at room temperature. Anti-His agarose (10 μl) was added to pull down mSecS. After washing three times with 1 ml of 1× TBS/0.1% Tween, the agarose was suspended in 40 μl of TBE-urea loading buffer (90 mM Tris-HCl [pH 8.3], 64.6 mM boric acid, 2.5 mM EDTA, 3.5 M urea), and 5 μl of each sample was loaded onto a 15% TBE-urea gel. After electrophoresis and transfer of the RNA to a nylon membrane, RNA was detected by Northern blotting with the Sec tRNA probe [20]. Of each binding reaction, 2 μl had been removed immediately after the incubation period and electrophoresed along with the reaction samples for analysis by Northern blotting that served as a loading control.
Native tRNA[Ser]Sec was aminoacylated with serine and phosphorylated with [γ-32P]ATP using O-phosphoseryl-tRNA[Ser]Sec kinase as described [20]. The 32P-labeled O-phosphoseryl-tRNA[Ser]Sec was added to a 10-μl reaction mixture containing 50 mM Tris-HCl (pH 7.5), 20 mM DTT, 10 mM KCl, 10 mM MgCl2 and 1 μg of each purified protein. Reactions were carried out for 30 min at 37 °C. Following incubation, the tRNA was deacylated by adding an equal volume of 1 M Tris-HCl (pH 8.0) and incubating at 37 °C for 45 min. Reactions were then spotted onto 3M paper (Whatman), placed in a TLC chamber, and chromatographed for 8 h using a mixture of butanol/acetic acid/water (12:3:5). The chromatogram was then exposed to a PhosphorImager screen.
ATP hydrolysis assays were carried out in a volume of 10 μl with 40 mM HEPES (pH 7.4), 20 mM KCl, 10 mM MgCl2, 5 μCi of [α-32P]ATP, and 10 mM DTT and either with or without 0.25 mM selenide. After adding 0.3 mg/ml final concentration of each selenophosphate synthetase protein, reactions were incubated at 37 °C for 1 h under anaerobic conditions. Then 0.5 μl of each reaction was loaded onto PEI TLC plates, the plates were run in 0.8 M LiCl, and the developed TLC plates were exposed to a PhosphorImager screen. For NMR analysis, ATP hydrolysis reactions were carried out under anaerobic conditions in 3-mm NMR tubes in a total volume of 200 μl with 2 mM ATP instead of [α-32P]ATP. NMR tubes were sealed and incubated at 37 °C for 4 h prior to 31P NMR spectroscopic analysis [9].
Synthetic Sec tRNA was used in all biosynthetic reactions. Synthesis, purification, and aminoacylation of Sec tRNA were carried out as described [20]. All of the reactions were set up under anaerobic conditions before chromatographic analysis. For Sec biosynthesis, the selenium donor SeP was either generated from selenide by using mSPS2-Cys or hydrolyzed from chemically synthesized [(CH3)3SiO]3PSe [31]. For generating the selenium donor with mSPS2-Cys, a 10-μl reaction containing 50 mM NH4HCO3 (pH 7.6), 10 mM DTT, 2 mM MgCl2, 2 mM KCl, 2 mM ATP, and 2 μg of mSPS2-Cys with or without 1 mM selenide was preincubated at 37 °C for 1 h. The mSPS2-Cys reaction was added to 10 μl containing 50 mM Tris-HCl (pH 7.0), 20 mM DTT, 10 mM MgCl2, 2 μM PLP, 1.0 μg of purified mSecS, and approximately 5 μg (about 30,000 cpm) of either O-phospho-[3H]seryl-tRNA[Ser]Sec or [3H]seryl-tRNA[Ser]Sec and, in a positive control reaction, SelA in place of mSecS, and in a negative control reaction, thioredoxin (with a His-tag) in place of mSecS. Reactions were incubated at 37 °C for 2 h and then heated at 75 °C for 5 min, aminoacyl-tRNAs were deacylated [20], and 1 μl of an unlabeled amino acid mix (containing 12.5 mM concentration each of serine, O-phosphoserine, Sec, and alanine in 50 mM of KBH4) was added. Each reaction, along with several control lanes containing unlabeled amino acids and 14C-pyruvate, was chromatographed on Whatman 3M filter paper (45 × 60 cm) in ethanol/acetic acid/water (12:3:5) for 16 to 20 h. Then 1.0-cm strips were cut out of the dried chromatogram and counted in a liquid scintillation counter. The locations of each amino acid were determined by staining the lanes with unlabeled amino acids in 0.3% ninhydrin in acetone or by cutting out 1.0-cm strips of lanes with 14C-pyruvate and counting in a liquid scintillation counter. For direct use of SeP as a selenium donor, reactions contained the same components as above except 1 mM SeP was used in place of mSPS2-Cys reaction solutions and DTT was omitted from the mSecS reactions, since we found that the activity of mSecS is higher without DTT. SeP was generated by hydrolysis of 20 mM chemically synthesized [(CH3)3SiO]3PSe [31]. Reactions were incubated and analyzed as above.
For measurement of the SecS synthesis rate, reactions were carried out in a total volume of 10 μl of 50 mM Tris-HCl (pH 7.0), with 10 mM MgCl2, 10 mM KCl, 0.2 mM SeP, and 4 μg (approximately 120 pmol) of O-phospho-[3H]seryl-tRNA[Ser]Sec. Reactions were initiated by adding 2 μg (approximately 35 pmol) of mSecS and were stopped at specific time points between 0 and 80 min by boiling for 2 min and then deacylating and counting as described above.
GenBank (http://www.ncbi.nlm.nih.gov/Genbank) accession numbers for the sequences used in this paper are Homo sapiens (SecS_HS), Q9HD40; Mus musculus (SecS_MM), Q6P6M7; Drosophila melanogaster (SecS_DM), NP_649556; C. elegans (SecS_CE), Q18953; Methanococcus jannaschii (SecS_MJ), Q58027; Methanopyrus kandleri (SecS_MK), Q8TXK0. SelA sequences: Escherichia coli (SelA_EC), BAE77702; Geobacter sulfurreducens (SelA_GS), P61736. SelA-like sequences: Methanococcus jannaschii (SelA-like_MJ), Q57622; Methanopyrus kandleri (SelA-like_MK), AAM01835; E. coli selA, M64177; selD, M30184; mouse secS, AL049338; sps2, NM_009266; sps1, NM_175400; and C. elegans sps2, NM_070203. The numbers for other PLP-containing proteins are Methanococcus jannaschii Sep-tRNA:Cys-tRNA synthase (SepCysS_MJ), Q59072; Medicago truncatula Orn/Lys/Arg decarboxylase (COG1982_MT), ABE83138.1; Archaeoglobus fulgidus glutamate decarboxylase (COG0076_AF), O28275; and Thermococcus kodakarensis glycine/serine hydroxymethyltransferase (COG0112_TK), 05JF06.
|
10.1371/journal.pntd.0003233 | Clinical Features and Patient Management of Lujo Hemorrhagic Fever | In 2008 a nosocomial outbreak of five cases of viral hemorrhagic fever due to a novel arenavirus, Lujo virus, occurred in Johannesburg, South Africa. Lujo virus is only the second pathogenic arenavirus, after Lassa virus, to be recognized in Africa and the first in over 40 years. Because of the remote, resource-poor, and often politically unstable regions where Lassa fever and other viral hemorrhagic fevers typically occur, there have been few opportunities to undertake in-depth study of their clinical manifestations, transmission dynamics, pathogenesis, or response to treatment options typically available in industrialized countries.
We describe the clinical features of five cases of Lujo hemorrhagic fever and summarize their clinical management, as well as providing additional epidemiologic detail regarding the 2008 outbreak. Illness typically began with the abrupt onset of fever, malaise, headache, and myalgias followed successively by sore throat, chest pain, gastrointestinal symptoms, rash, minor hemorrhage, subconjunctival injection, and neck and facial swelling over the first week of illness. No major hemorrhage was noted. Neurological signs were sometimes seen in the late stages. Shock and multi-organ system failure, often with evidence of disseminated intravascular coagulopathy, ensued in the second week, with death in four of the five cases. Distinctive treatment components of the one surviving patient included rapid commencement of the antiviral drug ribavirin and administration of HMG-CoA reductase inhibitors (statins), N-acetylcysteine, and recombinant factor VIIa.
Lujo virus causes a clinical syndrome remarkably similar to Lassa fever. Considering the high case-fatality and significant logistical impediments to controlled treatment efficacy trials for viral hemorrhagic fever, it is both logical and ethical to explore the use of the various compounds used in the treatment of the surviving case reported here in future outbreaks. Clinical observations should be systematically recorded to facilitate objective evaluation of treatment efficacy. Due to the risk of secondary transmission, viral hemorrhagic fever precautions should be implemented for all cases of Lujo virus infection, with specialized precautions to protect against aerosols when performing enhanced-risk procedures such as endotracheal intubation.
| Viral hemorrhagic fever is a syndrome often associated with high fatality and risk of secondary transmission. In 2008, an outbreak of a novel hemorrhagic fever virus called Lujo occurred in Johannesburg, South Africa, with secondary transmission from the index patient to four healthcare workers. Four of the five patients died. Lujo belongs to the arenavirus family and is only the second pathogenic arenavirus, after Lassa virus, to be recognized in Africa and the first in over 40 years. Because most viral hemorrhagic fevers occur in remote, resource-poor settings, few in-depth controlled studies of their clinical manifestations, transmission dynamics, pathogenesis, or response to treatment options are possible. We describe the clinical features of the five cases in this outbreak and summarize the clinical management, as well as providing additional epidemiologic detail. Lujo virus causes a clinical syndrome remarkably similar to Lassa fever. The treatment options used in these five cases are discussed as well as the recommended precautions to prevent secondary transmission.
| Viral hemorrhagic fever (VHF) is an acute systemic illness classically involving fever, a constellation of initially nonspecific signs and symptoms, and a propensity for bleeding and shock.
VHF may be caused by more than 25 different viruses from four taxonomic families: Arenaviridae, Filoviridae, Bunyaviridae, and Flaviviridae. Transmission of hemorrhagic fever viruses is through direct contact with blood and bodily fluids during the acute illness. Although patient isolation and specific VHF precautions (consisting of surgical mask, double gloves, gown, protective apron, face shield, and shoe covers) are advised for added security, experience has shown that routine universal and contact precautions are protective in most cases [1]. Aerosol precautions, such as the use of N95 particulate filters, are only recommended when performing specific potentially aerosol-generating procedures, such as endotracheal intubation.
South Africa has often played a role of “sentinel” for VHF in countries further to the north through the travel and admission of undiagnosed patients to South African hospitals, often with subsequent nosocomial transmission to healthcare workers. For example, cases of Marburg and Ebola hemorrhagic fevers have been reported in Johannesburg in persons initiating travel in Zimbabwe [2] and Gabon [3], respectively.
In 2008 a nosocomial outbreak of five cases of VHF occurred in Johannesburg [4], [5] (figure 1). The primary patient was a tour operator who was evacuated from Lusaka, Zambia. The etiologic agent was determined to be a novel arenavirus and the name “Lujo virus” was proposed. The source of the patient's infection is unknown, but assumed to be a rodent, as with all other pathogenic arenaviruses. Recent field studies of small mammals in Zambia did not result in isolation of Lujo virus, although another novel arenavirus was discovered [6].
Arenaviruses are divided into two groups: the New World (or Tacaribe) complex, and the Old World (or Lymphocytic Choriomeningitis/Lassa) complex, with various members of both groups causing VHF in South America and Africa, respectively [7] Lassa virus, the distribution of which is confined to West Africa, is the only other Old World arenavirus associated with VHF [8]. Lujo virus is only the second pathogenic arenavirus to be recognized in Africa and the first in over 40 years.
Some arenavirus infections, especially Lassa fever, have shown benefit with the use of the nucleoside analogue ribavirin [9]. Because of the remote and resource poor locations where Lassa fever typically occurs, as well as the history of civil unrest in West Africa in recent decades, there have been few opportunities to undertake in-depth study of the clinical manifestations or pathogenesis of Lassa fever or other VHFs, or the response of these infections to treatment options typically available in industrialized countries. We describe the clinical features of the five recognized cases of Lujo hemorrhagic fever (LHF) in the 2008 outbreak in South Africa and summarize their clinical management, as well as providing additional epidemiologic detail, with a focus on the risks for secondary transmission.
The initial description of the outbreak [4] was published primarily under the auspices of the South African National Institute for Communicable Diseases, which had a blanket ethics approval for use of all the patients' data. The same data set has been used for this publication, with ethics committee approval, with the exception of further data collated on the one survivor, who provided written consent for use of data and images related to her illness.
The five patients' ages ranged from 33 to 47 years. There were two white females, two black females, and one white male. The incubation periods of the 3 secondary and 1 tertiary cases ranged from 9-13 days. Four of the five patients died (CFR 80%).
Based on the five cases of LHF recognized to date, the clinical disease associated with LHF is remarkably similar to Lassa fever [7]. Surprisingly, the two viruses are genetically quite distinct (up to 38.1% on the nucleotide level), with Lujo virus grouping much closer genetically to Old World arenaviruses not associated with VHF [5] Lassa fever classically begins with non-specific signs and symptoms including fever, general malaise, headache, myalgia, chest or retrosternal pain, and sore throat with progressive diarrhea and other gastrointestinal involvement [7], [9]. Severe cases may progress to a capillary leak syndrome with septic shock, rash, facial and neck swelling, and multi-organ system failure. The facial and neck swelling seen in both LHF and Lassa fever appear to be specific to Old World arenavirus infection and may help differentiate it from other African VHFs. Like in Lassa fever (and despite the slight misnomer “VHF”), major bleeding was not a prominent feature in the patients with LHF, although minor bleeding was common. The AST and ALT are typically elevated in Lassa fever, with AST much greater than ALT and high levels of AST associated with a poor prognosis [7]. This same pattern was seen in all five patients with LHF, with the only survivor manifesting the lowest peak AST and AST: ALT ratio.
Some distinctive features of LHF relative to typical Lassa fever were the abrupt disease onset (typically indolent in Lassa fever) and the presence of DIC, which is generally not considered to be part of the pathogenesis of Lassa fever, although the matter has not been extensively studied [9]. Although rash is consistently seen in light-skinned persons with Lassa fever, for unknown reasons it is almost never seen in blacks. All of the white patients and one of the two black patients with LHF manifested a very prominent rash. Interestingly, the black patient without rash was HIV infected, suggesting that the rash of LHF may be immune mediated. Patient 5 also had relative bradycardia, an interesting finding given reports of depressed cardiac function in an animal model of arenavirus infection [12].
The CFR associated with this outbreak of LHF was 80%. The CFR of hospitalized patients with Lassa fever is typically in the 20–30% range, ranging up to 50% in some nosocomial outbreaks [13]. However, mild and asymptomatic Lassa virus infection is thought to be common, with mortality rates less than 5% when infection in the community is considered [7], [14]. No antibody survey of case contacts or community members in the region of origin of the index case in Zambia has been conducted to determine if mild or asymptomatic infection with Lujo virus occurs.
The four nosocomial infections of Lujo virus illustrate the risk to healthcare workers. Although no specific exposures were reported and some degree of personal protective equipment was worn by all four secondary or tertiary cases, it appears that strict barrier nursing practices were not always maintained and full VHF precautions were often implemented late in the course of treatment, if at all. Furthermore, the four infected healthcare workers generally had very close and sometimes prolonged contact with the patient, including in closed settings, such as the medical evacuation flight of Patient 1, augmenting the possibility of exposure to blood and bodily fluids. They also performed procedures that are often considered to be high risk, such as endotracheal intubation, insertion of indwelling intravascular catheters, and dialysis. The transmissibility of other emerging viruses such as SARS and MERS coronaviruses has similarly been enhanced when such procedures have been performed [15]. In addition to the 4 secondary/tertiary cases, another 94 persons were identified as contacts and monitored, including support staff (kitchen, laundry, cleaning), laboratory and radiography technicians, and nursing staff. We did not categorize contacts in terms of risk at the time, but now estimate that at least 30 of these would be reasonably categorized as high risk. Nevertheless, no suspected cases of LHF were noted in this group.
We suspect that the degree of transmissibility of Lujo virus is likely analogous to that of Lassa virus, for which, although reliable reproduction numbers and secondary attack rates are difficult to ascertain, they are generally thought to be low. Nevertheless, occasional outbreaks with secondary and tertiary cases are sometimes seen, especially when barrier nursing practices are not maintained [16], [17]. Until the matter can be studied more thoroughly, VHF precautions should certainly be implemented for all suspected and confirmed cases of LHF, with specialized precautions to protect against aerosols when performing endotracheal intubation [1].
Despite the high prevalence of HIV infection in many areas of sub-Saharan Africa, including some areas where VHF is common, data are scarce on HIV and hemorrhagic fever virus co-infection, such as was the case with our Patient 4. She was also infected with hepatitis B virus. A 68 year old Sierra Leonean man with a history of HIV infection and chronic progressive neurological deterioration was infected with Lassa virus in 2006 [18] The patient survived despite severe disease requiring intubation and mechanical ventilation. In the 2000–2001 outbreak of Ebola virus in Uganda, the CFR was not statistically different between those who were HIV positive and negative [19]. The samples were anonymously tested and no clinical data were reported. Although the clinical data on Patient 4 are also sparse, there were no obvious differences in the clinical manifestations of LHF in this patient compared to the others, with the exception of the aforementioned absence of rash. It is also interesting to note that her peak fever (38.5°C) and leukocyte count (14×109/L) were not particularly high, consistent with her compromised immune system.
There have been very few controlled studies on the management of VHF. Most recommendations represent the informal consensus of experienced clinicians and investigators. Supportive therapy is the mainstay [20]. The pathogenesis of severe cases of VHF is thought to be similar to severe sepsis, with a severe inflammatory response syndrome mediated in part by various soluble cytokines and chemokines and nitric oxide [21]. Therefore, the basic management principles of shock are also recommended for VHF [20], [22] However, since most VHFs occur in resource-poor areas with little access to advanced ICU medicine, opportunities to use and make observations on the efficacy of these or other advanced treatment options are rare.
Although obviously not a controlled trial, we were nevertheless able to make some detailed observations on the management of five patients with LHF, who were often treated in more advanced healthcare settings. The most detailed data are from Patient 5, who was the only patient for whom a specific diagnosis of VHF was considered and confirmed early in the course of disease. Despite receiving ribavirin at disease onset, Patient 5's clinical status deteriorated and her illness was severe and prolonged. Although these results could be interpreted as lack of efficacy of ribavirin against Lujo virus, this is unlikely considering the drug's proven efficacy in other arenavirus infections [8], [23]–[25] Of greater importance was probably the fact that ribavirin was administered orally for the first 6 days of treatment. Efficacy of oral ribavirin for arenavirus infection has not been definitively shown and, in light of the significant first-pass hepatic metabolism resulting in an oral bioavailability of only ∼50%, it is unlikely that oral administration reliably reaches the minimum inhibitory concentration for arenaviruses in serum [26] Serum levels are undoubtedly further diminished by decreased gut absorption, vomiting, and diarrhea in these severely ill patients.
Various adjunctive therapies with demonstrated or theoretical efficacy in severe sepsis were administered to Patient 5 and a few of the other patients, including HMG-CoA reductase inhibitors (statins), N-acetylcysteine [27], [28], recombinant factor VIIa, [29], [30], [31] mechanical ventilation, plasmapheresis, and hemodialysis. Animal models of sepsis have suggested that statin drugs may improve outcomes in septic shock [32], [10]. Furthermore, a large, population-based cohort analysis in Canada showed reduced risk of sepsis in patients with cardiovascular disease who were treated with statins [11]. Patient enrolment is currently ongoing for prospective trials of statin therapy after the development of sepsis. N-acetylcysteine is an antioxidant and free radical scavenger that resulted in decreased nuclear factor-κB and interleukin-8 in patients with sepsis, suggesting a blunting of the inflammatory response [28]. Recombinant factor VIIa is a prohaemostatic agent thought to act at the local site of tissue injury and vascular wall disruption by binding to exposed tissue factor to promote generation of thrombin and platelet activation. [29]. The drug has been used in hemophilia and other coagulation disorders, as well as in liver disease, reversal of anticoagulant therapy, and for episodes of excessive or life threatening bleeding related to surgery or trauma [30], [31]. Other therapies being explored for sepsis and, in some cases specifically for VHF, such as the recombinant inhibitor of the tissue factor/factor VIIa coagulation pathway, rNAPc2, and activated protein C, were not used in this outbreak due to lack of availability and/or risk of bleeding. The seemingly counterintuitive use of anticoagulants like rNAPc2 stemmed from work with an Ebola virus animal model to ameliorate the effects of tissue factor resulting in DIC [21].
It is difficult to assess the contribution of the various therapies to the patient outcomes. Although hemofiltration has been suggested in patients with refractory hemodynamic septic shock, with a significant decrease in ICU mortality in responders [33], and plasmapheresis appeared to have a brief positive effect in Patient 2, we are reluctant to advocate treatments or procedures that potentially increase healthcare worker exposure to blood. In fact, one explanation for the high secondary attack rate associated with this outbreak could be that such high-risk procedures were frequently undertaken.
Many of the drugs employed in the management of Patient 5 are already clinically approved. Investigation of many of these compounds in animal models of VHF is warranted, including in LHF model using strain 13/N guinea pigs [34]. Ideally, controlled clinical trials in humans would also be undertaken, although the feasibility of this is dubious for most VHFs, with the possible exception of Lassa fever, for which many infections occur across West Africa, or perhaps through a “multicenter” approach through advanced planning with Ministries of Health and other partners in endemic areas for VHFs [35], [36], [21]. Until controlled efficacy data are available, and considering the high CFR often associated with VHF, we feel that it is both logical and ethical to explore the use of these approved compounds in treatment of patients with VHF when possible. Treating clinicians should make a concerted effort to collect and publish detailed, repeated, and systematic clinical observations to facilitate objective evaluation of their efficacy.
The pace of discovery of arenaviruses has increased considerably in recent years, with over ten new viruses being isolated since 2000. Pathogenic arenaviruses will almost certainly continue to be discovered. Furthermore, rapid population growth, especially in Africa, and incursion for both economic and leisure activities into natural habitats harboring rodents will likely put humans at risk. The clinical findings and management experience reported here will be of use to clinicians faced with patients with arenavirus infections and as well as other VHFs.
|
10.1371/journal.ppat.1000859 | The Physical Relationship between Infectivity and Prion Protein Aggregates Is Strain-Dependent | Prions are unconventional infectious agents thought to be primarily composed of PrPSc, a multimeric misfolded conformer of the ubiquitously expressed host-encoded prion protein (PrPC). They cause fatal neurodegenerative diseases in both animals and humans. The disease phenotype is not uniform within species, and stable, self-propagating variations in PrPSc conformation could encode this ‘strain’ diversity. However, much remains to be learned about the physical relationship between the infectious agent and PrPSc aggregation state, and how this varies according to the strain. We applied a sedimentation velocity technique to a panel of natural, biologically cloned strains obtained by propagation of classical and atypical sheep scrapie and BSE infectious sources in transgenic mice expressing ovine PrP. Detergent-solubilized, infected brain homogenates were used as starting material. Solubilization conditions were optimized to separate PrPSc aggregates from PrPC. The distribution of PrPSc and infectivity in the gradient was determined by immunoblotting and mouse bioassay, respectively. As a general feature, a major proteinase K-resistant PrPSc peak was observed in the middle part of the gradient. This population approximately corresponds to multimers of 12–30 PrP molecules, if constituted of PrP only. For two strains, infectivity peaked in a markedly different region of the gradient. This most infectious component sedimented very slowly, suggesting small size oligomers and/or low density PrPSc aggregates. Extending this study to hamster prions passaged in hamster PrP transgenic mice revealed that the highly infectious, slowly sedimenting particles could be a feature of strains able to induce a rapidly lethal disease. Our findings suggest that prion infectious particles are subjected to marked strain-dependent variations, which in turn could influence the strain biological phenotype, in particular the replication dynamics.
| Prions are unconventional transmissible agents causing fatal neurodegenerative diseases in human and animals. They are thought to be formed from polymers of abnormal conformations of the host-encoded prion protein (PrP), but little is known about the physical organization of the infectious particles and any relationship between packing order and infectivity. As an additional layer of complexity, different PrP conformational variants associated with distinct biological phenotypes, or ‘strains’, can propagate in the same host. We subjected PrP polymers from eight different ovine and hamster prion strains to sedimentation velocity centrifugation, which allows separation of macromolecular complexes according to their size, density or shape. We showed that, whereas the PrP sedimentation profiles share common features, the infectivity profiles exhibit striking differences amongst the strains. For four of them, the infectious component was predominantly associated with slowly sedimenting particles, suggestive of small size oligomers and/or low density PrP aggregates. Such particles appeared to be a feature of strains able to induce a rapidly lethal disease in the recipient host. Our findings suggest that prion infectious particles are subjected to marked strain-dependent variations, which in turn could influence the strain biological phenotype, in particular the replication dynamics.
| Transmissible spongiform encephalopathies (TSE), such as human Creutzfeldt-Jakob disease, sheep scrapie, bovine spongiform encephalopathy (BSE) and chronic wasting disease of cervidae, are infectious, fatal, neurodegenerative disorders caused by prions [1]. Prions are unconventional pathogens primarily composed of PrPSc, a rearranged conformer of the ubiquitously expressed prion protein (PrPC), whose precise physiological function is largely unknown. Upon infection, PrPSc dictates the self-perpetuating conformational conversion of PrPC into nascent PrPSc. This conversion involves – without any apparent post-translational modification – the refolding of soluble, alpha-helix-rich PrPC molecules into beta-sheet enriched PrPSc polymers that form deposits in TSE-infected brains [2], [3] and are assumed to be responsible for the observed neurodegenerative disorders [4]. The conversion reaction may proceed through a nucleated polymerization mechanism in which PrPSc multimers recruit PrPC molecules and trigger their conformational conversion into PrPSc (for review [5]). The refolding/multimerisation process confers distinct physico-chemical properties to PrPSc, such as insolubility in non-denaturing detergents and partial resistance to proteolysis [6].
Distinct prion entities, referred to as strains, are known to self-propagate in the same host and exhibit distinguishable phenotypic traits that are heritable, such as incubation time, neuropathological and biochemical properties (for reviews: [7], [8], [9]). Accumulating experimental evidence indicates that strain-specified properties are encoded within structural differences in the conformation of the PrPSc molecules, which are faithfully imparted to host PrPC during the conversion process [10], [11], [12], [13], [14], [15], [16], [17]. However, the extent to which the aggregation state varies between different stains, and participates to strain-specific prion biology is unknown. The various fractionation methods and preparative procedures previously employed to estimate the size of the infectious particles [18], [19], [20], [21], [22], [23], [24], [25] have led to a vast range of measured sizes, making it difficult to relate any variation to potential strain differences. Of note, almost all of these studies used substantially purified PrPSc as a starting material.
In this study, we developed a specific protocol to fractionate PrP particles according to their sedimentation velocity properties in a viscous medium, characterized their relative levels of infectivity and looked for strain-specific variations. In contrast to previous reports, experiments were performed on crude brain homogenates, which a priori contain all TSE infectivity. We worked with a panel of strains that were biologically cloned on homogeneous genetic backgrounds, obtained after transmission of either classical and atypical (Nor98) sheep scrapie and BSE, or hamster scrapie infectious sources in transgenic mice expressing ovine PrP (VRQ allele; tg338 mice) and hamster PrP (tg7 line), respectively. We demonstrate that the sedimentation profile of the infectious component dramatically varies with the strain. We further show that the predominance of slowly sedimenting infectious particles that segregate from the bulk of proteinase K-resistant PrPSc particles may be a distinctive feature of strains able to induce a rapidly lethal disease.
PrPSc aggregates present in detergent-solubilised brain tissue homogenates were fractionated by sedimentation velocity centrifugation in an iodixanol gradient (Optiprep). The experimental conditions were established with brain material from tg338 mice that were infected or not with LA21K fast strain (referred to as LA21K), a prototypal, rapid strain that kills the mice within ∼2 months (see Table 1 for information on the strains used in this study). As a first step, we tested a variety of detergents for solubilization, which showed variable efficacy in terms of partition of PrPC and PrPSc species. For example, the use of standard solubilization buffers containing Triton X-100 and sodium deoxycholate or sarkosyl led to sedimentation of both isoforms throughout the gradient (Figure S1), indicating an incomplete release of total PrP from cellular constituents. In contrast, the sequential use of dodecyl maltoside and sarkosyl resulted in more efficient separation of the two PrP isoforms. Thus, in the conditions eventually employed (see Figure 1 for a summarizing flow diagram), the bulk of PrPC molecules remained in the upper fractions 1–4 (Figure 2A and D, green line), while both PrPSc (Figure 2B) and proteinase K (PK) resistant PrPSc species (Figure 2C–D, black line) were mainly detected in fractions 6–20 of the gradient. Importantly, no pelleted PrP material was observed in the selected conditions. Increasing the ultracentrifugation time caused the majority of PrPSc to sediment toward the heaviest fractions of the gradient, indicating that this material had not reached its density equilibrium (data not shown). Both dodecyl maltoside and sarkosyl are known to efficiently solubilize membrane structures, including rafts [26], [27], [28], yet PrPSc could be attached to abnormal, prion-induced structures. To address this point, brain homogenates were solubilized using these detergents in more stringent conditions, i.e. at 37°C instead of 4°C [29], however the sedimentation profile of PrPSc was affected only marginally (Figure S2A).
In order to assess the reproducibility of the partition and to enable quantitative analysis of the data, 7 independent fractionations were performed using different pooled or individual brains and the resulting data fitted (Figure 3A, black line). This revealed that ∼80% of the PK-resistant PrPSc material sedimented as one major peak (maximum in fractions 10–12) with a Gaussian-like distribution. Standard globular macromolecules and ovine recombinant PrP oligomers [30] loaded on gradients run in parallel enabled estimation of the approximate molecular mass of the PK-resistant PrPSc aggregates forming the peak in fraction 10–12: between 200 and 500 kDa (by reference to the marker proteins, the sedimentation profile of which was affected only marginally in the presence of detergents), and ∼850 kDa based on the position of the 36-mer PrP oligomer (Figure 3A).
When solubilized brain material was PK-treated prior to ultracentrifugation, the PrPres sedimentation profile resembled that observed with intact brain material (Figure S2B). However, when semi-purified PrPSc in the form of scrapie-associated fibrils [31], [32] was resolubilized and centrifuged, a markedly different profile was obtained, with peaks in fractions 22 and 30 (bottom fraction) (Figure S2C). Interestingly, fast sedimenting PrPSc material was also observed with Italian scrapie agent (referred to as SSit), which in tg338 mice produces very long incubation times and abundant plaque-like PrPSc deposits in the brain [33], in contrast to the LA21K agent. These plaques can be stained by thioflavin S (Figure S3A–B), indicating the presence of amyloid fibrils. When SSit-infected brain material was fractionated, the majority of PrPSc multimers peaked in fractions 24 to 30 of the gradient (Figure S3C).
These results suggest that the experimental conditions employed preserve potential differences in the aggregation state of PrPSc thereby enabling the comparative analysis of sedimentation properties of “close to natural” PrPSc aggregates and of associated infectivity.
The distribution of prion infectivity throughout the gradients was determined by an incubation time bioassay [34]. tg338 mice were inoculated intracerebrally with diluted aliquots from the different fractions. In terminally diseased mice, the PrPSc electrophoretic profile and regional distribution in the brain observed for representative fractions were both consistent and similar to that with the original brain material, indicating a conservation of the strain biological phenotype (Figure S4A and data not shown). The mean survival time values resulting from the analysis of 2 independent gradients are shown in Figure 3A (red line). Typically, the mice inoculated with the PK-resistant PrPSc-richest fractions (6–20) succumbed to disease in more than 80 days, whereas those inoculated with fractions 1–3 died in a markedly shorter time, ∼60–70 days. The correlation between the mean survival time values and infectivity was established by using a standard infectious dose/survival time curve previously established for this strain (Figure S5). This analysis indicated that fractions 1–2 were between 100- and 1000-fold more infectious than fractions 6–20 (Figure 3A, blue scale). These upper fractions - within the sedimentation peak of aldolase (158 kDa) and upstream of 12-mer PrP oligomer – totaled <10% of PK-resistant PrPSc molecules (Figure 3A).
There is substantial evidence to indicate that a fraction of PrPSc can exhibit low sedimenting properties and be PK-sensitive [28], [35], [36]. Recently, thermolysin has been used as a means to isolate PK-sensitive forms of PrPSc, while degrading PrPC [37]. When the upper fractions from LA21K gradients were thermolysin-digested, no enrichment in thermolysin-resistant species was observed by immunoblot as compared to unfractionated brain material (Figure S6A–C). To further analyze the forms of PrPSc present in the upper fractions, aliquots were centrifuged at 100 000 g for 1 h to produce soluble (supernatant) and insoluble (pellet) fractions, before immunoblot analysis. The ratio of soluble and insoluble PrP species in LA21K versus uninfected fractions was determined based on signal intensities. As a result, the top two LA21K fractions were reproducibly shown to contain equivalent amounts of soluble material and about 2-fold more sedimentable material as compared to the corresponding uninfected fractions (Figure S6D).
Detergents and lipids have been proposed to increase the apparent infectious titer of PrPSc preparations non-specifically [38], [39] and such compounds are relatively abundant in the upper fractions of the gradient. To test whether such an effect was responsible for the comparatively high infectivity levels of the top fractions, the PK-resistant PrPSc-enriched fractions 10 to 12 were mixed, incubated with either the top fractions 1 to 3 of a gradient made with uninfected tg338 brain or with dodecyl maltoside alone, and inoculated to mice. As a result, in either condition, the relative titer of these fractions was not significantly modified (Figure 4).
To confirm that the differences in survival times observed between mice inoculated with the various fractions were correlated with differences in infectivity content, a mouse-free, cell bioassay was used. The distribution and level of LA21K infectivity in the gradient was measured using Rov cells [40] that were exposed in parallel to fraction aliquots and to serial tenfold dilutions of a LA21K brain homogenate prepared in the same conditions. Consistent with the bioassay data, the most infectious fractions were found at the top of the gradient and were ≥100-fold more infectious than the middle fractions (Figure S7).
Overall these data indicate that the upper fractions were intrinsically highly infectious. The fact that the cumulated infectivity in the gradient fractions did not differ significantly from that present in the loaded material prior solubilization also supports the conclusion that the detergents used did not alter infectivity estimates.
Brains of tg338 mice infected by another fast ovine strain named 127S (Table 1) were also fractionated and analyzed for PrP and infectivity content. 127S PK-resistant PrPSc peaked in fractions 10–12 (Figure 3B), as in the case of LA21K agent, despite some variation of the sedimentation profile in the bottom part of the gradient. Strikingly, the sedimentation profile of infectivity again largely segregated from that of PK-resistant PrPSc as assessed by mouse bioassay. The top two fractions were at least 50–100-fold more infectious than all the other fractions, including the major PK-resistant PrPSc peak (Figure 3B).
We next examined whether the decoupling of PK-resistant PrPSc and infectivity sedimentation profiles was a general feature of ovine strains. Three more strains were studied of which the incubation time in tg338 mice is at least twice that of LA21K and 127S: LA19K, Nor98 and sheep BSE (see Table 1). Four to five independent fractionations with different pooled or individual brains were performed for each strain. The combined curves resulting from the replicate analysis of PrP content indicated that a majority of PK-resistant PrPSc peaked in fractions 10–12, similar to that seen with the two fast strains. However, faster sedimenting species were also observed, notably in fractions 16, 20 for LA19K and fractions 22–24 for Nor98 (Figure 3C–D). Remarkably, the infectivity sedimentation profile of these 3 strains, as established from bioassay of two independent gradients, tended to overlap PK-resistant PrPSc distribution, with a very small proportion of the total infectivity in the top fractions. LA19K most infectious fractions ranged from fractions 8 to 24 with a peak in fraction 20 (range of mean survival time: 152 to 163 days), while the top and bottom fractions were ∼100-fold less infectious (mean survival time ∼185 to 210 days; Figure 3C). Nor98 infectivity peaked in fraction 11 and to a lesser degree in fraction 17 and 22 (mean survival time 222, 240 and 245 days, respectively; Figure 3D). Fractions in the immediate vicinity of these peaks were among the most infectious, (except fraction 13). In contrast, the upper fractions were ∼100-fold less infectious (survival time prolonged by >40 days). The most infectious sheep BSE fractions were found in fractions 6–12, 16 and 20 (mean survival times of 155–160, 164 and 163 days) while the top and bottom fractions were about 50-fold and 100-fold less infectious, respectively (survival time of ∼175 days and >180 days; Figure 3E).
To further explore the possibility that slow sedimenting infectivity could be a specific feature of fast prion strains, we applied the same sedimentation velocity protocol to three hamster strains passaged on tg7 transgenic mice expressing hamster PrP (Table 1). For fast strains 139H and Sc237, the infectivity peaked in the top two fractions, which contained ∼10% of the total PK-resistant PrPSc material present in the gradient. The two 139H PK-resistant PrPSc peaks in fractions 11–12 and 16–18 and the Sc237 PK-resistant PrPSc peak in fraction 11–12 were ∼50-fold and <10-fold less infectious, respectively (n = 2 independent experiments made with different individual brains; Figure 5A–B). ME7H strain, characterized by a longer incubation time, produced a different picture since the mice inoculated with the PK-resistant PrPSc peak in fraction 11 were the fastest to succumb to disease, i.e. ∼180 days, whereas those inoculated with the top 3 fractions had mean survival times significantly prolonged by 20 to 40 days (Figure 5C). Therefore, much less infectivity was present in the upper region of the gradient than in the PK-resistant PrPSc containing fractions (about 50–100-fold, based on the available results of the endpoint titration of ME7H, still ongoing). Collectively, the contrasted sedimentation properties of fast and slow hamster strains were reminiscent of the results obtained with the ovine strains.
Here we compared the sedimentation velocity properties of the infectivity and of abnormal PrP amongst several prion strains, using experimental conditions aimed at preserving as much as possible the “natural” multimerization state of the prion particles while minimizing artifacts due to improper membrane solubilization. To our knowledge this is the first study that allows a rigorous comparison of phenotypically distinct strains, cloned and propagated on the same genetic background. We found striking, strain-specific differences in the sedimentation profile of the infectious prion particles, which are not reflected in the sedimentation properties of the bulk of PrPSc.
Fractionation of five tg338 mouse-passaged ovine prions revealed a major PK-resistant PrPSc population, which peaked at the same position of the gradient regardless of the strain. By comparison with standard molecular mass markers and recombinant ovine PrP oligomers [30], we estimated that this population might correspond to approximately 12–30 PrP monomers averaging ∼30 kDa each, if constituted of PrP only, with the caveat that great caution must be exercised when attributing a size to a polymer by comparing its velocity to that of molecular mass markers. This result suggests that PrPSc is not a collection of multimers with a regular continuum of size. However, the overall sedimentation profiles were not uniform among the strains, indicating that the size distribution of PrPSc aggregates is strain-dependent. tg7 mouse-passaged hamster prions showed PrPSc sedimentation characteristics resembling that of ovine strains, with the same position of the major peak and limited variation. Greater differences in the size distribution of PrPSc aggregates may however exist depending on strain and/or PrP sequence, since one amyloid-forming ovine prion (Italian scrapie; Figure S3) showed a clear shift of PrPSc toward heavier fractions of the gradient.
In two studies, larger polymers were shown to be more PK-resistant than smaller ones [28], [35], indicating that the resistance to proteolysis of PrPSc largely depends on its quaternary structure. The PrPSc associated with Nor98 agent, a newly discovered strain responsible for an atypical form of field scrapie, is highly sensitive to PK digestion, when compared to the other ovine strains studied here [41], [42]. Notwithstanding, Nor98 PrPSc exhibited not only the same predominant peak as the other strains, but also the highest proportion of faster sedimenting PrPSc species. This argues that the pronounced PK sensitivity of Nor98 PrPSc is not due to low size aggregates, but rather to its tertiary structure. Supporting this view, its C-terminal region is accessible to PK, in contrast to classical scrapie agents [43], [44].
The infectivity sedimentation profiles unexpectedly contrast with that of the PrPSc aggregates. While infectivity and PK-resistant PrPSc roughly co-fractionated for LA19K, Nor98 and sheep BSE, their respective distribution was mostly decoupled for LA21K and 127S: with these two fast strains, infectivity essentially partitioned in the upper fractions, which were 2–3 logs more infectious than the PK-resistant PrPSc–richest fractions. LA21K, 127S and Nor98 are three strains exhibiting similarly high infectious titers in tg338 mice, making the observed differences particularly striking. Remarkably, a comparable situation was observed for hamster strains. Thus, ME7H infectivity and PK-resistant PrPSc sedimentation velocity profiles were broadly congruent, whereas for the fast strains 139H and Sc237 the infectivity peaked in the upper region of the gradient. Altogether, these findings lend support to the view that the predominance of slow sedimenting particles may be a common feature of prion strains with short incubation time.
Such a decoupling between infectivity and PK-resistant PrPSc with respect to the size of the particles is to our knowledge unprecedented in the literature. Two earlier studies performed with fast hamster prions [18], [23] also reported a relatively low PrPres content of the most infectious fractions. However, the level of infectivity in these fractions did not exceed that of the PrPres-richest fractions. The preparations used were fibrillar PrPres material under the form of SAF or Rods, disaggregated by sonication in the presence of anionic detergents [18], [23]. Such a procedure is likely to destroy discrete subpopulations of infectious particles, which may explain the observed discrepancy.
Which form(s) of PrPSc could support the high infectivity of the fast strains' slow sedimenting component? One possibility is that it consists essentially of PK-resistant aggregates, with high specific infectivity [24], [45], [46], [47]. If so, then ≥99% of LA21K or 127S infectivity would be supported by ≤10% of PK-resistant PrPSc molecules. Alternatively, infectivity could be mostly associated with a form of PrPSc with low resistance to PK. While a variable, strain-dependent proportion of abnormal PrP seems fairly PK-sensitive [28], [35], [36], [48], [49], little is known about its specific infectivity, with two recent studies suggesting that it could be minimal [37], [50]. However, a PK-sensitive and soluble form of PrPSc has been shown to support a substantial fraction of infectivity [51] and to have a good in vitro converting activity [35]. The abundance of both PrPC and other components in the upper fractions impeded further characterization of the most infectious PrPSc particles in the present study. Classical approaches based on PrPSc conformation-dependent assay [36], [49], [52] are unhelpful here; as already shown by others, low sedimenting, PrPC-rich fractions contain little or no conformation-dependent immunoreactive material [28], [35]. Also, we failed to detect soluble or thermolysin-resistant PrPSc material that might be indicative of the presence of PK-sensitive molecules [48], [51] in these fractions. Additional experiments including the titration of PK-treated and then fractionated infectious material are ongoing to further assess the protease-resistance of the slow sedimenting component.
What physical properties could account for slow sedimentation of fast strains infectious component? The detergents employed are known to produce a high degree of membrane solubilization [27], [28], [53], [54], [55], [56], including for GPI-anchored proteins of detergent-resistant microdomains [26], and they led indeed to an efficient solubilization of PrPC (Figure S6E). Dodecyl maltoside is also known to preserve activity of protein complexes in the detergent-solubilized state [27]. It is still possible that a tightly bound cellular component of low density, such as lipid molecule [45], or low-density lipoprotein [57], remains part of the prion particles in the most infectious fractions. This would imply that the tightness of such an interaction specifically differs between strains. Alternatively, the low sedimenting infectivity component could involve truly small size PrPSc particles. In this hypothesis, the data would indicate a size smaller than a PrP pentamer, which is compatible with that reported for PK-sensitive PrPSc aggregates [28], [35]. As a means to distinguish between small size and lipid associated PrPSc aggregates, LA21K gradient centrifugation time was doubled. As a result, infectivity was found to peak in fraction 4 instead of fraction 2 (data not shown). While arguing against lipid floatation of the most infectious component, this does not exclude the presence of tightly associated lipids. Additional experiments will be needed to address this issue.
The neuropathology induced in mice by the fast and slow sedimenting particles did not differ for a given prion, therefore suggestive of structurally related multimers (Figure S4). The slow sedimenting infectious particles could reflect a stronger tendency of large PrPSc polymers to fragment. In the case of yeast prions [PSI+], it has been proposed that the fittest strains are those whose large fibers break more easily into smaller oligomers that in turn act as new seeds for conversion [58], a concept that was then extended to mammalian prions [59]. In this regard, our preliminary results indicate that LA21K and 127S PrPSc aggregates exhibit the lowest ‘stability’ among the ovine strains, as assayed by conformational stability assay.
In addition to providing another measurable criterion of prion strain-related phenotypic variation, this study revealed the diversity of their infectious component. Further biochemical and biophysical investigations will be crucial for a mechanistic understanding of the replication dynamics of mammalian prions, in relation with the disease phenotype.
All the experiments involving animals were approved by the INRA Jouy-en-Josas ethics committee in accordance with the European Community Council Directive 86/609/EEC.
The ovine prion strains used in this study have been obtained through serial transmission and subsequent biological cloning by limiting dilutions of classical and atypical field scrapie and experimental sheep BSE sources to tg338 transgenic mice expressing the VRQ allele of ovine PrP. The characterization of their phenotype in tg338 mice was performed as previously reported [41], [60], [61]. Pooled or individual tg338 mouse brain homogenates (20% wt/vol. in 5% glucose) were used in centrifugation analyses. Three hamster strains, 139H, Sc237 and ME7H, were also studied. These strains (kindly provided by R. Carp, Staten Island, NY, USA) were serially passaged on tg7 transgenic mice expressing hamster PrP (kindly provided by CSL-Behring (Marburg); [48], [62]). Both 139H and Sc237 were subsequently cloned by limiting dilution on this genetic background. Individual tg7 infected brains (20% wt/vol.) were used in centrifugation analyses. Non-infected brain tissue homogenates served as controls.
The entire procedure was performed at 4°C. Mouse brain homogenates were solubilized by adding an equal volume of solubilization buffer (50 mM HEPES pH 7.4, 300 mM NaCl, 10 mM EDTA, 2 mM DTT, 4% (wt/vol.) dodecyl-β-D-maltoside (Sigma)) and incubated for 30 min on ice. Sarkosyl (N-lauryl sarcosine; Fluka) was added to a final concentration of 2% (wt/vol.) and the incubation continued for a further 30 min on ice. A volume of 150 µl was loaded on a 4.8 ml continuous 10–25% iodixanol gradient (Optiprep, Axys-shield), unless specified otherwise, with a final concentration of 25 mM HEPES pH 7.4, 150 mM NaCl, 2 mM EDTA, 1 mM DTT, 0.5% Sarkosyl. Gradient linearity was verified by refractometry. In standard experiments (Figure 1), the gradients were centrifuged at 285 000 g for 45 min in a swinging-bucket SW-55 rotor using an Optima LE-80K ultracentrifuge (Beckman Coulter). Gradients were then manually segregated into 30 equal fractions of 170 µl from the bottom using a peristaltic pump. Fractions were aliquoted for immunoblot or bioassay analyses. The strains were fractionated in parallel to preserve as much as possible identical experimental conditions. To avoid any cross-contamination, each piece of the equipment was thoroughly decontaminated with 5 M NaOH followed by several rinses in deionised water after each gradient collection. Standard markers (GE Healthcare, Little Chalfont, UK) of aldolase (158 kDa), catalase (232 kDa), ferritin (440 kDa) and thyroglobulin (669 kDa) were run in parallel.
The protocol used was as described above except that PK (100 µg/ml final concentration; Euromedex, Mundolsheim, France) was added during the solubilization phase in sarkosyl (1 h at 37°C).
Brain homogenates were treated with 20 µg/ml of PK for 1 h at 37°C. The digestion was stopped by the addition of 5 mM phenylmethylsulfonyl fluoride. The solution was added to 10% sarkosyl and 10 mM Tris-HCl pH 7.4 and then centrifuged at 175 000 g for 30 min at 20°C on a 10% (wt/vol.) sucrose cushion in a Beckmann TL100 ultracentrifuge. Pellets were resuspended in 2% (wt/vol.) dodecyl maltoside and the above solubilization/fractionation protocol was followed.
Monomeric and oligomeric forms of purified ovine recombinant PrP [30] were resuspended at a final concentration of 7–10 µM in 20 mM citrate buffer. An aliquot (150 µl) was loaded onto a 10–25% iodixanol gradient in citrate buffer and centrifuged at 285 000 g for 45 min in a SW-55 rotor. Gradients were segregated as described above. Fraction aliquots (20 µl) were analyzed for PrP content by immunoblot (see below). In these conditions, recombinant, monomeric PrP was found in the upper fractions 1–3 (not shown).
Aliquots of the collected fractions were treated or not with 50 µg/ml PK before methanol precipitation. The pellet was resuspended in Laemmli buffer and denatured at 100°C for 5 min. The samples (15 µl) were run on 4–12% NuPAGE gels (Invitrogen, Cergy Pontoise, France), electrotransferred onto nitrocellulose membranes, and probed with 0.1 µg/ml biotinylated anti-PrP monoclonal antibody Sha31 as previously described [60]. Immunoreactivity was visualized by chemiluminescence (GE Healthcare). The amount of PrP present in each fraction was determined by the GeneTools software after acquisition of chemiluminescent signals with a GeneGnome digital imager (Syngene, Frederick, Maryland, United States).
All Bio-Rad TeSeE detection kit reagents were kindly provided by S. Simon (CEA, France; [63]). Briefly, aliquots (75 µl) of the collected fractions were digested with PK (50 µg/ml final concentration) for 1 h at 37°C before B buffer precipitation and centrifugation at 28 000 g for 15 min. The pellet was resuspended in 25 µl of 5 M urea before denaturation at 100°C for 10 min. R6 buffer (200 µl) was subsequently added to the samples and duplicates were analyzed in microtiter plates coated with anti-PrP antibody 11C6. The plates were left at room temperature for 2 h. After 3 washes in R2 buffer, 100 µl/well of the enzyme conjugate (Bar224 anti-PrP antibody) was added for 2 h. The substrate (100 µl) was added for 30 min and incubated in the dark. The absorbance was read at 450 nm. A dilution range of ovine, monomeric recombinant PrP was used for quantification of relative PK-resistant PrPSc levels.
The PK-resistant PrPSc sedimentation profiles obtained by either immunoblot or ELISA were normalized to units and decomposed using multiple Gaussians fits procedures with a maximum entropy minimization approach.
Fractions were methanol-precipitated. The pellet was resuspended in lysis buffer (2% sodium deoxycholate, 2% Triton X-100, 200 mM Tris-HCl pH 7.4) and mixed with an equal volume of thermolysin diluted in lysis buffer to yield a final concentration of 125 µg/ml (unless indicated otherwise) for 1 h at 70°C. The samples were analyzed by electrophoresis (4–12% gels) and immunoblotted as above. Blots were probed with either Sha31b or anti-octarepeat specific Pc248 anti-PrP antibody [64] at a final concentration of 0.1 µg/ml, before acquisition of chemiluminescent signals with a GeneGnome digital imager and analysis by the GeneTools software (Syngene, Frederick, Maryland, United States).
Aliquots (20 µl) of the fractions were added to 80 µl of 5% glucose before centrifugation at 100 000 g for 1h at 4°C in a Beckmann TL100 ultracentrifuge to generate soluble (supernatant) and insoluble (pellet) fractions. Proteins in the supernatant were precipitated with 400 µl of cold methanol, centrifuged at 16 000 g for 30 min before denaturation in 100 µl of sample buffer. The insoluble pellet was resuspended in 20 µl of Laemmli buffer before denaturation. Samples (20 µl) were analyzed by immunoblot as described above.
Fractions 1 to 4 and then every other two fractions (unless specified otherwise) were diluted extemporarily in 5% glucose (1∶5). This procedure was performed in a class II microbiological cabinet according to a strict protocol to avoid any cross-contamination. Individually identified 6- to 10-week old tg338 or tg7 recipient mice (n = 6 mice per fraction) were inoculated intracerebrally with 20 µl of the solution. Recipient mice inoculated with fractionated uninfected mouse brain were euthanized while still healthy at >400 days post-infection. Their brain was negative for PrPres content. Mice showing TSE neurological signs were monitored daily and euthanized in extremis. Brains were removed and analyzed for PrPres content by either immunoblot or histoblot (see below) as a confirmatory test. The survival time was defined as the number of days from inoculation to euthanasia.
The survival times of tg338 or tg7 reporter mice was measured for each tenfold dilution tested during endpoint titration experiments performed with all but ME7H strains. Animals inoculated with 2 mg of infectious brain tissue were assigned a relative infectious dose of 0. From these data, curves representing the relative infectious dose to survival time were established ([41] and Figure S5). The different patterns in survival time distribution among the gradients can thus be looked at as a function of relative infectious dose so as to estimate what difference in survival times between inoculated fractions means in terms of infectivity.
The scrapie cell assay technique will be fully described elsewhere. Briefly, LA21K gradient fractions aliquots (typically 20–30 µL) were methanol precipitated before resuspension in culture medium (alpha minimal essential medium supplemented with 10% fetal bovine serum, 100 U/ml penicillin and 10 µg/ml streptomycin). We verified that methanol precipitation did not affect the overall level of infectivity. Rov cell [40] monolayers established in a 96 well plate were exposed to the fractions for one week. After several washes, the cells were further cultivated for two weeks before fixation and PrPSc detection by immunofluorescence as previously described [65]. Immunofluorescence signals were acquired with an inverted fluorescence microscope (Zeiss Axiovert). A program in NIH Image J software was designed to quantify the levels of PrPSc signal per cell in each well. Serial tenfold dilutions of LA21K infected brain homogenates were prepared in the same conditions and run in parallel experiments to establish a tissue culture infectious doses curve that directly relates to the percentage of PrPSc content.
Brains were rapidly removed from euthanized mice and frozen on dry ice. Cryosections were cut at 8–10 µm, transferred onto Superfrost slides and kept at −20°C until use. Histoblot analyses were performed on 3 brains per infection, using the 12F10 anti-PrP antibody as described [60]. For thioflavin-S binding, formalin- or methanol-fixed sections were incubated with 0.01% thioflavin-S for 1 hour as previously described [66]. Sections were then incubated with nuclear marker 4′, 6-diamidino-2-phenylindole (Sigma), mounted in fluoromount-G (Interchim) before acquisition with an inverted fluorescence microscope (Zeiss Axiovert) and analysis with the Metamorph software.
The Swiss-Prot accession numbers for the proteins mentioned in the text are sheep (P23907) and hamster PrP (P04273).
|
10.1371/journal.pcbi.1006171 | Thalamocortical and intracortical laminar connectivity determines sleep spindle properties | Sleep spindles are brief oscillatory events during non-rapid eye movement (NREM) sleep. Spindle density and synchronization properties are different in MEG versus EEG recordings in humans and also vary with learning performance, suggesting spindle involvement in memory consolidation. Here, using computational models, we identified network mechanisms that may explain differences in spindle properties across cortical structures. First, we report that differences in spindle occurrence between MEG and EEG data may arise from the contrasting properties of the core and matrix thalamocortical systems. The matrix system, projecting superficially, has wider thalamocortical fanout compared to the core system, which projects to middle layers, and requires the recruitment of a larger population of neurons to initiate a spindle. This property was sufficient to explain lower spindle density and higher spatial synchrony of spindles in the superficial cortical layers, as observed in the EEG signal. In contrast, spindles in the core system occurred more frequently but less synchronously, as observed in the MEG recordings. Furthermore, consistent with human recordings, in the model, spindles occurred independently in the core system but the matrix system spindles commonly co-occurred with core spindles. We also found that the intracortical excitatory connections from layer III/IV to layer V promote spindle propagation from the core to the matrix system, leading to widespread spindle activity. Our study predicts that plasticity of intra- and inter-cortical connectivity can potentially be a mechanism for increased spindle density as has been observed during learning.
| The density of sleep spindles has been shown to correlate with memory consolidation. Sleep spindles occur more often in human MEG than EEG recordings. We developed a thalamocortical network model that is capable of spontaneous generation of spindles across cortical layers and that captures the essential statistical features of spindles observed empirically. Our study predicts that differences in thalamocortical connectivity, known from anatomical studies, are sufficient to explain the differences in the spindle properties between EEG and MEG which are observed in human recordings. Furthermore, our model predicts that intracortical connectivity between cortical layers, a property influenced by sleep preceding learning, increases spindle density. Results from our study highlight the role of intracortical and thalamocortical projections on the occurrence and properties of spindles.
| Sleep marks a profound change of brain state as manifested by the spontaneous emergence of characteristic oscillatory activities. In humans, sleep spindles consist of waxing-and-waning bursts of field potentials oscillating at 11–15 Hz lasting for 0.5–3 s and recurring every 5–15 s. Experimental and computational studies have identified that both the thalamus and the cortex are involved in the generation and propagation of spindles. Spindles are known to occur in isolated thalamus after decortication in vivo and in thalamic slice recordings in vitro [1, 2], demonstrating that the thalamus is sufficient for spindle generation. In in-vivo conditions, the cortex has been shown to be actively involved in the initiation and termination of spindles [3] as well as the long-range synchronization of spindles [4] [5].
Multiple lines of evidence indicate that spindle oscillations are linked to memory consolidation during sleep. Spindle density is known to increase following training in hippocampal-dependent [6] as well as procedural memory [7] memory tasks. Spindle density also correlates with better memory retention following sleep in verbal tasks [8, 9]. More recently, it was shown that pharmacologically increasing spindle density leads to better post-sleep performance in hippocampal-dependent learning tasks [10]. Furthermore, spindle activity metrics, including amplitude and duration, were predictive of learning performance [11–13], suggesting that spindle event occurrence, amplitude, and duration influence memory consolidation.
In human recordings, spindle occurrence and synchronization vary based on the recording modality. Spindles recorded with magnetoencephalography (MEG) are more frequent and less synchronized, as compared to those recorded with electroencephalography (EEG) [14]. It has been proposed that the contrast between MEG and EEG spindles reflects the differential involvement of the core and matrix thalamocortical systems, respectively [15]. Core projections are focal to layer IV, whereas matrix projections are widespread in upper layers [16]. This hypothesis is supported by human laminar microelectrode data which demonstrated two spindle generators, one associated with middle cortical layers and the other superficial [17]. Taken together, these studies suggest that there could be two systems of spindle generation within the cortex and that these correspond to the core and matrix anatomical networks. However, the network and cellular mechanisms whereby the core and matrix systems interact to generate both independent and co-occurring spindles across cortical layers are not understood.
In this study, we developed a computational model of thalamus and cortex that replicates known features of spindle occurrence in MEG and EEG recordings. While our previous efforts have been focused on the neural mechanisms involved in the generation of isolated spindles[5], in this study we identified the critical mechanisms underlying the spontaneous generation of spindles across different cortical layers and their interactions.
Histograms of EEG and MEG gradiometer inter-spindle intervals are shown in Fig 1C. For neither channel type are ISIs distributed normally as determined by Lilliefors tests (D2571 = 0.1062, p = 1.0e-3, D4802 = 0.1022, p = 1.0e-3), suggesting that traditional descriptive statistics are of limited utility. However, the ISI at peak of the respective distributions is longer for EEG than it is the MEG. In addition, a two-sample Kolmogorov-Smirnov test confirms that EEG and MEG ISIs are not drawn from the same distribution (D2571,4802 = 0.079, p = 1.5e-9). While the data where not found to be drawn from any parametric distribution with 95% confidence, an exponential fit (MEG) and lognormal fit (EEG) are shown in red overlay for illustrative purposes. These data are consistent with previous empirical recordings [18] and suggest that sleep spindles have different properties across superficial vs. deep cortical layers.
To investigate the mechanisms behind distinct spindle properties across cortical locations as observed in EEG and MEG signals, we constructed a model of thalamus and cortex that incorporated the two characteristic thalamocortical systems: core and matrix. These systems contained distinct thalamic populations that projected to the superficial (matrix) and middle (core) cortical layers. Four cell types were used to model distinct cell populations: thalamocortical relay (TC) and reticular (RE) neurons in the thalamus, and excitatory pyramidal (PY) and inhibitory (IN) neurons in each of three layers of the cortical network. A schematic representation of the synaptic connections and cortical geometry of the network model is shown in Fig 2. In the matrix system, both thalamocortical (from matrix TCs to the apical dendrites of layer 5 pyramidal neurons (PYs) located in the layer 1) and corticothalamic synapses (from layer 5 PYs back to the thalamus) formed diffuse connections. The core system had a focal connection pattern in both thalamocortical (from core TCs to PYs in the layer III/IV) and corticothalamic (from layer VI PYs to the thalamus) projections. Because spindles recorded in EEG signal reflect the activity of superficial layers while MEG records spindles originating from deeper layers (Fig 1 and [19]), we compared the activity of the model’s matrix system, which has projections to the superficial layers, to empirical EEG recordings and compared the activity in model layer 3/4 to empirical MEG recordings.
In agreement with our previous studies [3, 5, 20, 21], simulated stage 2 sleep consisted of multiple spindle events involving thalamic and cortical neuronal populations (Fig 3). During one such typical spindle event (highlighted by the box in Fig 3A and 3B), cortical and thalamic neurons in both the core and matrix system had elevated and synchronized firing (Fig 3A bottom), consistent with previous in-vivo experimental recordings [22]. In the model, spindles within each system were initiated from spontaneous activity within cortical layers and then spread to thalamic neurons, similar to our previous study[5]. The spontaneous activity due to miniature EPSPs in glutamergic cortical synapses led to fluctuations in membrane voltage and sparse firing. At random times, the miniature EPSPs summed such that a small number of locally connected PY neurons spiked within a short window (<100ms), which then induced spiking in thalamic cells through corticothalamic connections. This initiated spindle oscillations in the thalamic population mediated by TC-RE interactions as described before [20, 23, 24]. Thalamic spindles in turn propagated to the neocortex leading to joint thalamocortical spindle events whose features were shaped by the properties of thalamocortical and corticothalamic connections. In this study, we examined how the process of spindle generation occurs in a thalamocortical network with mutually interacting core and matrix systems, wherein the thalamic network of each system is capable of generating spindles independently.
Based on the anatomical data [16], the main difference between the modeled core and matrix systems was the radii or fanout of connections in thalamocortical and corticothalamic projections (in the baseline model, the fanout was 10 times wider for the matrix compared to the core system). Furthermore, the strength of each synaptic connection was scaled by the number of input connections to each neuron [25, 26], leading to weaker individual thalamocortical projections in the matrix as compared to the core. These differences in the strength and fanout of thalamocortical connectivity resulted in distinctive core and matrix spindle properties (see Fig 3A, right vs left). First, both cortical and thalamic spindles were more spatially focal in the core system as only a small subset of neurons was involved in a typical spindle event at any given time. In contrast, within the matrix system spindles were global (involving the entire cell population) and highly synchronous across all cell types. These results are consistent with our previous studies [5] and suggest that the connectivity properties of thalamocortical projections determine the degree of synchronization in the cortical network. Second, spindle density was higher in the core system compared to the matrix system. At every spatial location in the cortical network of the core system, the characteristic time between spindles was shorter compared to that between spindles in the matrix system (Fig 3A left vs right). In order to quantify the spatial and temporal properties of spindles, we computed an estimated LFP as an average of the dendritic synaptic currents for every group of contiguous 100 cortical neurons. LFPs of the core system were estimated from the currents generated in the dendrites of layer 3/4 neurons while the LFP of the matrix system was computed from the dendritic currents of layer 5 neurons, located in the superficial cortical layers (Fig 2). After applying a bandpass filter (6–15 Hz), the spatial properties of estimated core and matrix LFP (Fig 3C) closely matched the MEG and EEG recordings, respectively (Fig 1). In subsequent analyses, we used this estimated LFP to further examine the properties of the spindle oscillations in the core and matrix systems.
We identified spindles in the estimated LFP using an automated spindle detection algorithm similar to that used in experimental studies (details are provided in the method section). The spindle density, defined as the number of spindles occurring per minute of simulation time, was greater in the core compared to the matrix (Fig 4A) as confirmed by an independent-sample t-test (t(18) = 7.06, p<0.001 for across estimated LFP channels and t(2060) = 19.2, p<0.001 across all spindles). The results of this analysis agree with the experimental observation that MEG spindles occur more frequently than EEG spindles. While the average spindle density was significantly different between the core and matrix, in both systems the distribution of inter-spindle intervals peaks below 4 seconds and has a long tail (Fig 4B). A two sample KS test comparing the distributions of inter-spindle intervals confirmed that the intervals were derived from different distributions (D1128,932 = 0.427, p<0.001). The peak ISI of the core was shorter than that of the matrix system, suggesting that the core network experiences shorter and more frequent quiescence periods than the matrix population. Furthermore, maximum-likelihood fits of the probability distributions (red line in Fig 4B) confirmed that the intervals of spindle occurrence cannot be described by a normal distribution. The long tails of the distributions suggest that a Poisson like process, as oppose to a periodic process, is responsible for spindle generation. This observation is consistent with previous experimental results [18, 27] and suggests that our computational model replicates essential statistical properties of spindles observed in in vivo experiments.
Several other features of simulated core and matrix spindles were similar to those found in experimental recordings. The average spindle duration was significantly higher in the core compared to the matrix system (Fig 4C) as confirmed by independent-sample t-test (t(2060) = 16.3, p<0.001). To quantify the difference in the spatial synchrony of spindles between the core and matrix systems, we computed the spatial correlation [28] between LFP groups at different distances (measured by the location of a neuron group in the network). The correlation strength decreased with distance for both systems (Fig 4D). However, the spindles in the core system were less spatially correlated overall when compared to spindles in the matrix system.
Simultaneous EEG and MEG measurements have found that about 50% of MEG spindles co-occur with EEG spindles, while about 85% of EEG spindles co-occur with MEG spindles [29]. Further, a spindle detected in the EEG signal is found to co-occur with about 66% more MEG channels than a spindle detected in MEG. Our model generates spindling patterns consistent with these features. The co-occurrence probability revealed that during periods of spindles in the matrix system, there was about 80% probability that core was also generating spindles (Fig 4E). In contrast, there was only a 40% probability of observing a matrix spindle during a core system spindle. An independent-sample t-test confirmed this difference between the systems across estimated LFP channels (t(14) = 31.4, p<0.001). Furthermore, we observed that the number of LFP channels that were simultaneously activated during a spindle event in the core system was higher when a spindle co-occurred in the matrix versus times when the spindles only occurred in the core (Fig 4F, t(14) = 67.2, p<0.001). This suggests that the co-occurrences of spindles in both systems are rare events but lead to the wide spread activation in both the core and matrix when they take place.
Finally, we examined the delay between spindles in the core and matrix systems (Fig 4G). We observed that on average (red line in Fig 4G), the spindle originated from the core system then spread to the matrix system with a mean delay of about 300 ms (delay was measured as the difference in onset times between co-occurring spindles within a window of 2,500 ms; negative delay values indicate spindles in which the core preceded matrix). The peak at -750 ms corresponds to spindles originating from the core system, while the peak at +750 ms suggests that at some network sites, spindles originated in the matrix system and then spread to the core system. While there were almost no events in which the matrix preceded the core by more than 1 sec (right of Fig 4G), many events occurred in which the core preceded the matrix by more than 1 sec (left of Fig 4G).
In sum, these results suggest that spindles were frequently initiated locally in the core system, then propagate to and spread throughout the matrix system. This can trigger spindles at the other locations of the core, so eventually, even regions in the core system that were not previously involved become recruited. These findings explain the experimental result that spindles are observed in more MEG channels when they also co-occur in the EEG [29].
We leveraged our model to examine factors that may influence spindle occurrence across cortical layers. The main difference between the core and matrix systems in the model was the breadth or fanout of the thalamic projections to the cortical network. Neuroanatomical studies suggest that the core system has focused projections while matrix system projects widely [16]. Here, we assessed the impacts of this characteristic by systematically varying the connection footprint of the thalamic matrix to superficial cortical regions, while holding the fanout of the thalamic core to layer 3/4 projections constant. We also modulated the corticothalamic projections in proportion to the thalamocortical projections. Using the estimated LFP from the cortical layers corresponding to core and matrix system, respectively, we quantified various spindle properties as the fanout was modulated.
Spindle density (the number of spindles per minute) in both layers was sensitive to the matrix system’s fanout. ANOVA confirmed significant effects of fanout and layer location, as well as an interaction between layer and fanout (fanout: F(6,112) = 66.4; p<0.01, Layer: F(1,112) = 65.18; p<0.01 and interaction F(6,112) = 22.8; p<0.01). When the matrix and core thalamus had similar fanouts (ratio 1 and 2.5 in Fig 5B), we observed a slightly higher density of spindles in the matrix than in the core system. This observation is consistent with the properties of these circuits (see Fig 2), wherein the matrix system contains direct reciprocal projections connecting cortical and thalamic subpopulations and the core system routes indirect projections from cortical (layer III/IV) neurons through layer VI to the thalamic nucleus. When the thalamocortical fanout of the matrix system was increased to above ~5 times the size of the core system, the density of spindles in the matrix system was reduced to around 4 spindles per minute. Interestingly, the density of spindles in the core system was also reduced when the thalamocortical fanout of the matrix system was further increased to above ~10 times of that in the core system (ratio above 10 in Fig 5B). This suggests that spindle density in both systems is determined not only by the radius of thalamocortical vs. corticothalamic projections, but also by interactions between the systems among the cortical layers. We further expound on the role of these cortical connections in the next section.
We also examined the effect of thalamocortical fanout on the distribution of inter-spindle intervals (Fig 5C). Although the mean value was largely independent of the projection radius, a long tailed distribution was observed for all values of fanout in the core. Contrastingly, in the matrix system the mean and peak of the inter-spindle interval shifted to the right (longer intervals) with increased fanout. With large fanouts, the majority of matrix system spindles had very long periods of silence (10-15s) between them. This suggests that thalamocortical fanout determines the peak of the inter-spindle interval distribution, but does not alter the stochastic nature of spindle occurrence.
The degree of thalamocortical fanout also influenced the co-occurrence of spindles in the core and matrix systems (Fig 5D). Increasing the fanout of the matrix system reduced spindle co-occurrence between two systems. This reduction resulted mainly from lower spindle density in both layers. However, the co-occurrence of core spindles during matrix spindles was higher for all values of fanout when matrix thalamocortical projections were at least 5 times broader than core projections. This suggests that the difference in spindle co-occurrence between EEG and MEG as observed in experiments [14] depends mainly on the difference in the radius of thalamocortical projections between the core and matrix systems, while overall level of co-occurrence is determined by the interaction between cortical layers.
We examined how spatial correlations during periods of spindles vary depending on the fanout of thalamocortical projections. The spatial correlation quantifies the degree of synchronization in the estimated LFP signals of network locations as a function of the distance between them. As expected, increasing the distance reduced the spatial correlation (Fig 4D). We next measured the mean value of the spatial correlation for each fanout condition. The mean correlation increased as a function of the fanout in the matrix system (Fig 5A). However, the spatial correlation within the core, and between the core and matrix systems, did not change with increases in the fanout, suggesting that the spatial synchronization of core spindles is largely influenced by thalamocortical fanout but not by interactions between the core and matrix systems as was observed for spindle density.
Does intra-cortical excitatory connectivity between layer 3/4 of the core system and layer 5 of the matrix system affect spindle occurrence? To answer this question, we first varied the strength of excitatory connections (AMPA and NMDA) from the core to matrix pyramidal neurons (Fig 6A and 6B). Here the reference point (or 100%) corresponds to the strength used in previous simulations, i.e. half the strength of a within-layer connection. The spindle density varied with the strength of the interlaminar connections (Fig 6A). For low connectivity strengths (below 100%), the spindle density of the matrix system was reduced significantly, while at high strengths (above 140%) the matrix system spindle density exceeded that of the control (100%). There were significant effects of connection strength and layer on the spindle density, as well as an interaction between the two factors (connection strength: F(5,96) = 24.7; p<0.01, layer: F(5,96) = 386.6; p<0.01 and interaction F(5,96) = 36.9; p<0.01) that suggests a layer-specific effect of modulating excitatory interlaminar connection strength. Similar to the spindle density, spindle co-occurrence between the core and matrix systems also increased as a function of interlaminar connection strength, reaching 80% for the both core and matrix at 150% connectivity. In contrast, changing the strength of excitatory connections from layer 5 to layer 3/4 had little effect on the spindle density, (Fig 6C). Taken together, these results suggest that the strength of the cortical core-to-matrix excitatory connections is one of the critical factors in determining spindle density and co-occurrence among spindles across both cortical lamina and the core/matrix systems.
Using computational modeling and data from EEG/MEG recordings in humans we found that the properties of sleep spindles vary across cortical layers and are influenced by thalamocortical, corticothalamic and cortico-laminar connections. This study was motivated by empirical findings demonstrating that spindles measured in EEG have different synchronization properties from those measured in MEG [14, 29]. EEG spindles occur less frequently and more synchronously in comparison to MEG spindles. Our new study confirms the speculation that anatomical differences between the matrix thalamocortical system, which has broader projections that target the cortex superficially, and the core system, which consists of focal projections which target the middle layers, can account for the differences between EEG and MEG signals. Furthermore, we discovered that the strength of corticocortical feedforward excitatory connections from the core to matrix neurons determines the spindle density in the matrix system, which predicts a specific neural mechanism for the interactions observed between MEG and EEG spindles.
There were several novel findings in this study. First, we developed a novel computational model of sleep spindling in which spindles manifested as a rare but global synchronous occurrence in the matrix pathway and a frequent but local occurrence in the core pathway. In other words, many spontaneous spindles occurred locally in the core system but only occasionally did this lead to globally organized spindles appearing in the matrix system. As a result, only a fraction of spindles co-occurred between the pathways (about 80% in matrix and 40% in core pathway). This is consistent with data reported for EEG vs MEG in vivo (Fig 1). In contrast, in our previous models [3, 5], spindles were induced by external stimulation and always occurred simultaneously in the core and matrix systems, but with different degrees of internal synchrony. In addition, these studies did not examine how the core and matrix pathways interact during spontaneously occurring spindles. Second, in this study we found that the distribution of the inter-spindle intervals between spontaneously occurring spindles in both the core and matrix pathways had long tails similar to a log-normal distribution. This result is consistent with analyses of MEG and EEG data reported in this study and in our prior study [18]. In our previous models [3, 5], spindles were induced by external stimulation and the statistics of spontaneously occurring spindles could not be explored. Third, we demonstrated that the strength of thalamocortical and corticothalamic connections determined the density and occurrence of spontaneously generated spindles. The spindle density was higher in the core pathway as compared to the matrix pathway with high co-occurrence of core spindles with matrix spindles. These findings were corroborated with experimental evidence from EEG/MEG recordings. Finally, we reported that laminar connections between the core and matrix could be a significant factor in determining spindle density, suggesting a possible mechanism of learning. When the strength of these connections was increased in the model, there was a significant increase in spindle occurrence, similar to the experimentally observed increase in spindle density following recent learning [10].
The origin of sleep spindle activity has been linked to thalamic oscillators based on a broad range of experimental studies [2, 30, 31]. The excitatory and inhibitory connections between thalamic relay and reticular neurons are critical in generating spindles [20, 23, 32, 33]. However, in intact brain, the properties of sleep spindles are also shaped by cortical networks. Indeed, the onset of a spindle oscillation and its termination are both dependent on cortical input to the thalamus [3, 34, 35]. In model studies, spindle oscillations in the thalamus are initiated when sufficiently strong activity in the cortex activates the thalamic network, and spindle termination is partially mediated by the desynchronization of corticothalamic input towards the end of spindles [3,32]. However, in simultaneous cortical and thalamic studies in humans, thalamic spindles were found to be tightly coupled to a preceding downstate, which in turn was triggered by converging cortical downstates [36]. Further modeling is required to reconcile these experimental results. In addition, thalamocortical interactions are known to be integral to the synchronization of spindles [5, 33]. In our new study, the core thalamocortical system revealed relatively high spindle density produced by focal and strong thalamocortical and corticothalamic projections. Such a pattern of connectivity between core thalamus and middle cortical layers allowed input from a small region of the cortex to initiate and maintain focal spindles in the core system. In contrast, the matrix system had relatively weak and broad thalamocortical connections requiring synchronized activity in broader cortical regions in order to initiate spindles in the thalamus.
We previously reported [5] that (1) within a single spindle event the synchrony of the neuronal firing is higher in the matrix than in the core system; (2) spindle are initiated in the core and with some delay in the matrix system. The overal density of core and matrix spindle events was, however, the same in these earlier models. In the new study we extended these previous results by explaining differences in the global spatio-temporal structure of spindle activity between the core and matrix systems. Our new model predicts that the focal nature of the core thalamocortical connectivity can explain the more frequent occurrence of spindles in the core system as observed in vivo. The strength of core-to-matrix intracortical connections determined the probability of core spindles to “propagate” to the matrix system. In our new model core spindles remained localized and have never involved the entire network, again in agreement with in vivo data.
We observed that the distribution of inter-spindle intervals reflects a non-periodic stochastic process such as a Poisson process, which is consistent with previous data [18, 27]. The state of the thalamocortical network, determined by the level of the intrinsic and synaptic conductances, contributed to the stochastic nature of spindle occurrence. Building off our previous work [21], we chose the intrinsic and synaptic properties in the model that match those in stage 2 sleep, a brain state when concentrations of acetylcholine and monoamines are reduced [37–39]. As a consequence, the K-leak currents and excitatory intracortical connections were set higher than in an awake-like state due to the reduction of acetylcholine and norepinephrine [40]. The high K-leak currents resulted in sparse spontaneous cortical firing during periods between spindles with occasional surges of local synchrony sustained by recurrent excitation within the cortex that could trigger spindle oscillations in the thalamus. Note that this mechanism may be different from spindle initiation during slow oscillation, when spindle activity appears to be initiated during Down state in thalamus [35]. Furthermore, the release of miniature EPSPs and IPSPs in the cortex was implemented as a Poission process that contributed to the stochastic nature of the baseline activity. All these factors led to a variable inter-spindle interval with long periods of silence when activity in the cortex was not sufficient to induce spindles. While it is known that an excitable medium with noise has a Poisson event distribution in reduced systems [41], here we show that a detailed biophysical model of spindle generation may lead to a Poission process due to specific intrinsic and network properties.
Layer IV excitatory neurons have a smaller dendritic structure compared to Layer V excitatory neurons [42]. Direct recordings and detailed dendritic reconstructions have shown large post-synaptic potentials in layer IV due to core thalamic input [42, 43]. We examined the role of thalamocortical and corticothalamic connections in a thalamocortical network with only one cortical layer (S1 Fig). We found that increasing the synaptic strength of thalamocortical and corticothalamic connections both increased the density and duration of spindles, however it did not influence their synchronization (S1A Fig). In contrast, changing fanout led to an increase in spindle density, duration, and synchronization. Furthermore, we examined the impact of thalamocortical and corticothalamic connections individually without applying a synaptic normalization rule (see Methods). We observed that the thalamocortical connections had a higher impact on spindle properties than corticothalamic connections (S1B Fig). In our full model with multiple layers, which included a weight normalization rule and wider fanout of the matrix pathway (based on experimental findings[16]), the synaptic strength of each thalamocortical synapse in the core pathway was higher than that in the matrix pathway. The exact value of the synaptic strength was chosen from the reduced model to match experimentally observed spindle durations, as observed in EEG/MEG and laminar recordings [17].
The simultaneous EEG and MEG recordings reported here and in our previous publications [14, 29] revealed that (a) MEG spindles occur earlier compared to the EEG spindles and (b) EEG spindles are seen in a higher number of the MEG sensors compared to the spindles occurring only in the MEG recordings. This resembles our current findings, in which the number of regions that were spindling in the core system during a matrix spindle was higher than when there was no spindle in the matrix system. Further, the distribution of spindle onset delays between the systems indicates that during matrix spindles some neurons of the core system fired early, and presumably contributed to the initiation of the matrix spindle, while others fired late and were recruited. Taken together, all the evidence suggests a characteristic and complex spatiotemporal evolution of spindle activity during co-occurring spindles, where spindles in the core spread to the matrix and in turn activate wider regions in the core leading to synchronized activation across cortical layers that is reflected by strong activity in both EEG and MEG. Thus, the model predicts that co-occurring spindles could lead to the recruitment of the large cortical areas, which indeed has been reported in previous studies [28, 44]. At the same time, local spindles occurring in the model within deep cortical layers may correspond to the local spindles observed in some studies [45], or may be even hidden from empirical recordings because of their localized and low amplitude properties. Finally, regional differences in thalamocorical and corticocortical connections could explain the characteristic regional and spatial patterns of spindles observed in human recordings [15].
The correspondence of the matrix vs core thalamocortical system to EEG vs MEG recordings was proposed originally to explain the differences between properties of the EEG and MEG spindles [14, 46]. Several lines of evidence support this hypothesis and was reviewed by Piantoni et al [15]. MEG and EEG share neural generators, but differences arise due to the biophysics of their cancellation patterns as they project from cortex to sensor. We have recently applied a biophysical forward model of MEG and EEG generation from a large-scale thalamocortical model that is similar to the model used in this study [47]. As hypothesized, in this combined neural/biophysical model, core-dominant spindles were more MEG-weighted than matrix-dominant spindles.
Human EEG studies have previously reported the existence of two types of spindles based on frequency—slow and fast spindles (9–12 and 12-15Hz). In our study, there was small difference (core-13.8 (subharmonic at 6.7Hz) and matrix-14.4 Hz (subharmonic-7.2Hz)) in the frequency between core and matrix spindles. However, this difference was much smaller than between fast and slow spindles (which are 9–12 and 12-15Hz) reported in vivo. These results are consistent with laminar recordings of spindles in humans [17], where the average frequency of middle vs upper layer spindles does not differ significantly. Furthermore, in intracranial recordings, both slow and fast spindles occur after down states [35, 36], as opposed to the reported occurrence of slow spindles before down states in the EEG [48]. Taken together, these findings suggest that the properties of the thalamocortical and corticothalamic connections explored here are not sufficient to explain the origin of fast vs. slow spindles as observed in rodent studies.
The relationship between spindle phase and spike timing of cortical neurons has been examined by previous studies, though their findings have been contradictory. While some studies have shown a preference in phase [45], others have shown no such preference [49]. In a recently published analysis of spindles recorded in the thalamus and cortex of humans [35], we found that thalamic spindling appears to drive cortical spindling, including both LFP and high gamma activity. This suggests that the nature of thalamic connections could influence the phase preference of spiking during spindle oscillations. In this new study, we measured the phase of cortical neurons’ spiking during spindles (S2 Fig) and we observed that, in our model, neurons both in the core and the matrix systems had a higher preference to spike at the peak of the spindle oscillation (corresponding to the oscillation phase values close to pi or–pi). We also compared the variability of the spiking phase in the matrix pathway versus the core one (S2C Fig). The statistical test comparing the phase of spiking in the core versus matrix systems aggregated for all cortical neurons was not significant (two sample KS test of phase distribution, KS statistic = 0.12, p = 0.44). However, normalized probability of spiking for the different spindle oscillation phases (obtained from the normalized histogram binned at 100 intervals) was significantly different between the core and the matrix for many values of phase. This suggests that the phase of spiking in the matrix pathway has a trend for being more variability than in the core pathway.
Increase in spindle density following learning is a robust experimental finding that suggests a role for sleep in memory consolidation [2, 6, 7, 10, 50, 51]. However, the neural mechanisms that increase spindle density after learning are not known. The hippocampal CA1 region projects to both superficial and deep layers of the rodent prefrontal cortex [52, 53]. In addition, experiments with simultaneous recordings from cortex and hippocampus report that during NREM sleep, the two structures show coordinated activity [54, 55], underlying spike sequence replay and the reactivation of memories that were recently learned while awake [56]. In our study, we found that activation of layer 3/4 of the neocortex triggers spindles that propagate between the core and matrix systems and eventually lead to spindle recruitment in wide regions of the neocortex. Based on these findings, we predict that hippocampal input to superficial cortical layers (layer 2/3/4) during NREM sleep can induce local activation and spindles in the core system, which then propagate to the matrix system thereby activating large cortical regions in both layers, potentially contributign to memory consolidation.
Elevated spindle density may arise due to the changes in the cortical microcircuit, of which excitatory interlaminar connections form the main component [57]. This circuit is implicated as a site of sensory coding and learning [58]. In this study, we identified that connection strength from the core to the matrix, but not vice verse, was critical in determining spindle density. This predicts that the increase in spindle density following a hippocampal-dependent task may arise from the strengthening of feedforward projections from middle to superficial cortical layers.
In sum, our study identified a rich set of the local and global network mechanisms involved in the propagation and interactions of spindles across different cortical structures. While spindle activity in the model arises from thalamic circuits, our study supports the idea that thalamocortical and intracortical projections significantly shape the properties of spindling activity and that this may explain the characteristic changes of spindle density associated with sleep-related memory replay.
The human research reported in this study was approved by the institutional review board at Partners Healthcare Network. Written informed consent was directly obtained from all subjects prior to their participation.
Extracranial electromagnetic fields were recorded in 4 healthy adults (3 female). Subjects did not report any neurological problems including sleep disorders, epilepsy, or substance dependence. In addition, subjects did not consume caffeine or alcohol on the day of recording. A whole-head MEG system with integrated EEG cap (Elekta Neuromag) was used to collect 204 planar gradiometers and 60 EEG channels. EEG data were referenced to an averaged mastoid. Additional details concerning data collection can be found in [46]. Sleep staging was performed by three neurologists according to standard criteria (Rechtschaffen and Kales, 1968). Data analyzed came from a 17.5 ± 3.4 (mean ± SD) minute period of stage 2 sleep.
Data were acquired at 603.107 Hz. Gross artifacts and bad channels were excluded manually. The continuous data were band-pass filtered to between 0.1 and 30 Hz and ICA (Delorme and Makeig, 2004) was used to remove the ECG component. Spindles were automatically detected in each MEG and EEG channel using a method modified from [45].The 10–16 Hz analytic signal was extracted from the data using the Hilbert transform and the envelope obtained by computing its elementwise modulus. The spindle-band envelope was smoothed with a Gaussian kernel (300 ms width, 40 ms σ). Putative spindles were initially marked as contiguous regions of the smoothed spindle-band envelope where the envelope amplitude was more than 2 standard deviations above the mean. Marked regions were then expanded until amplitude dropped below 1 standard deviation above the mean. Putative spindles shorter than 500 ms and longer than 2 s were excluded from further analysis. Inter-spindle intervals (ISIs) were computed from spindle center to center. Outlying ISIs longer than 20 seconds were excluded, as these are likely caused by false negatives in spindle detection. ISIs from all subjects and all channels were pooled together to form a single distribution for EEG and gradiometer data, respectively.
|
10.1371/journal.pntd.0004575 | A Two-Year Review on Epidemiology and Clinical Characteristics of Dengue Deaths in Malaysia, 2013-2014 | Dengue infection is the fastest spreading mosquito-borne viral disease, which affects people living in the tropical and subtropical countries. Malaysia had large dengue outbreaks in recent years. We aimed to study the demographics and clinical characteristics associated with dengue deaths in Malaysia.
We conducted a retrospective review on all dengue deaths that occurred nationwide between 1st January 2013 and 31st December 2014. Relevant data were extracted from mortality review reports and investigational forms. These cases were categorized into children (<15 years), adults (15–59 years) and elderly (≥60 years) to compare their clinical characteristics.
A total of 322 dengue deaths were reviewed. Their mean age was 40.7±19.30 years, half were females and 72.5% were adults. The median durations of first medical contact, and hospitalization were 1 and 3 days, respectively. Diabetes and hypertension were common co-morbidities among adults and elderly. The most common warning signs reported were lethargy and vomiting, with lethargy (p = 0.038) being more common in children, while abdominal pain was observed more often in the adults (p = 0.040). But 22.4% did not have any warning signs. Only 34% were suspected of dengue illness at their initial presentation. More adults developed severe plasma leakage (p = 0.018). More than half (54%) suffered from multi-organ involvement, and 20.2% were free from any organ involvement. Dengue deaths occurred at the median of 3 days post-admission. Dengue shock syndrome (DSS) contributed to more than 70% of dengue deaths, followed by severe organ involvement (69%) and severe bleeding (29.7%).
In Malaysia, dengue deaths occurred primarily in adult patients. DSS was the leading cause of death, regardless of age groups. The atypical presentation and dynamic progression of severe dengue in this cohort prompts early recognition and aggressive intervention to prevent deaths.
National Medical Research Registry (NMRR, NMRR-14-1374-23352)
| Dengue infection, especially severe dengue, affected more of adults from working age groups in the society. They can present with non-specific symptoms mimicking many other febrile illnesses, or severe symptoms suggestive of sepsis, with low suspicion of dengue. The clinical progression in severe dengue can be dynamic and sometimes unanticipated, whereby patients can deteriorate rapidly in a short period of time and succumb to death. Although children tend to have central nervous system involvement, where they presented with confusion and/ or seizure, and more elderly had heart involvement, the primary cause of dengue death, dengue shock syndrome, did not differ across different age groups.
| Dengue infection has been identified as the fastest spreading mosquito-borne viral disease by World Health Organization (WHO) [1]. At least 3.6 billion people living in the subtropical and tropical regions are at risk of contracting dengue, imposing substantial socioeconomic impact to affected populations [2].
Since the first outbreak in Malaysia in 1902, dengue has become a notifiable infection by 1973 [3]. Malaysia is endemic with all four dengue virus serotypes co-circulating over the past two decades [4], with predominantly DENV1 and DENV2 responsible for the escalating number of cases over the recent years [5]. Major outbreaks occurred in a cyclical pattern of approximately 8 years [6]. However, the reported cases had increased from 7,103 in 2000 to 43,346 in 2013, with a drastic escalation to 108,698 in 2014 [5]. While dengue mortality increased over the years, the case fatality ratio remained between 0.16% and 0.30% over the past decade [5]. Studies have revealed more severe manifestations with multisystem involvement for dengue infections, and an epidemic shift of age range incidence from children to adults in Malaysia [3, 4].
To date, there is no cure for dengue infections, apart from supportive treatments. There have been attempts to look at the risk factors associated with dengue haemorrhagic fever and dengue shock syndrome (DSS) focusing on dengue patients in single tertiary centres [7, 8]. Local published data on epidemiological characterisation and clinical spectrum of the fatal dengue cases remain limited. This review aims to study the epidemiology, clinical characteristics and cause of deaths on fatal dengue cases nationwide for 2013 and 2014.
This was a retrospective observational study reviewing all the dengue mortality cases which occurred in Malaysia between 1st January 2013 and 31st December 2014. Notification of all dengue deaths in Malaysia are mandatory. All cases were hospitalized and deaths occurred in hospital setting. Doctors managing each case were required to complete and submit the dengue specific investigational forms to the Ministry of Health. All these cases were reviewed by a group of independent consultants across different specialties at dengue mortality meetings held by each state health of departments. All relevant details pertaining to patient’s initial presentation and diagnosis, course of management in relation to the dynamic disease progression, and subsequent cause of death were reviewed extensively together with the managing teams during the meeting. It was aimed to identify areas for improvement and strengthen patients’ care to reduce mortality in future cases, apart from educational purpose. Detailed discussion was then documented in mortality review report.
The lists of patients’ names were retrieved from the Vector Borne Disease Control Division, Ministry of Health, Malaysia. Predesigned case report form (S1 Fig) was employed to collect all basic demographics, pre-existing co-morbidities, clinical and outcome data from the source documents, which included all mandatory investigational forms as well as mortality review reports from hospitals and state health departments, whichever available.
All cases were categorized into three age groups, namely children aged <15 years, adults aged 15–59 years, and elderly aged ≥60 years. Two cases with missing age were excluded from subsequent analyses involving comparison across the different age groups.
The case definition for dengue infection, definitions of warning signs, severe dengue and dengue shock syndrome (DSS) were referred to the WHO 2009 guideline [9]. All cases were confirmed seropositive for dengue non-structural glycoprotein 1 (NS1), and/or Immunoglobulin M (IgM) antibodies, and/or by polymerase chain reaction (PCR). The presence of NS1 antigen, IgM and IgG antibodies in the sera was detected either using the dengue-specific combinational rapid test (Standard Diagnostics BIOLINE Dengue Duo, Korea) or enzyme-linked immunosorbent assay (ELISA) test available at local hospitals, or samples were sent to the National Public Health Laboratory in Selangor to proceed with ELISA test if neither test was available in the hospital.
For the classification of various organ involvement in this cohort, liver involvement refers to high titre of alanine aminotransferase and/ or aspartate aminotransferase beyond 1000 iu/L, or acute liver failure. Kidney involvement refers to acute renal failure or acute exacerbation on chronic kidney disease. Heart involvement refers to myocarditis, pericarditis and/ or heart failure, either acute onset or exacerbation of the underlying heart disease, while central nervous system involvement refers to encephalopathy, encephalitis as well as intracranial bleeding.
This study was registered under National Medical Research Registry (NMRR, NMRR-14-1374-23352) and approved by Medical Research Ethics Committee, Ministry of Health.
Data were compiled and analysed using Statistical Package for Social Sciences program version 21 (SPSS: Inc. Chicago. Il. USA). All data were anonymized during analysis. Categorical variables were expressed in frequency and proportion. The normally distributed continuous variables were reported in means ± standard deviation, while the non-normally distributed continuous variables were reported in medians (25th percentile, 75th percentile). Comparisons across different age groups for initial presentation with warning signs, pre-morbid illnesses, criteria for severe dengue and causes of death were performed using Pearson Chi-Square tests. The two-sided statistical significance level, p-value, was set at 0.05 for all inferential analyses in this study.
This study included 322 fatal dengue cases, with 94 in 2013 and 228 in 2014. The mean age of this cohort was 40.7 ± 19.30 years. One in ten of the cases were children aged <15 years while 17.5% were elderly aged 60 years and beyond (Table 1). The youngest case was a 10-day old female infant who died of DSS while the eldest was a 90-year old lady with multiple co-morbidities who died of severe dengue with upper gastrointestinal bleeding. Majority were adults from the working age group. Six of them were pregnant women and succumbed due to DSS. Half of our cohort (51.2%) were females. More than half of them (54%) were of Malay ethnic, followed by Chinese (21.7%), Indian (11.2%) and other minor ethnics (4.7%), reflecting closely the ethnic distribution in the country. Up to 8.1% of the deceased persons were of foreign nationalities, with 14 Bangladeshi, three Indonesians, three Nepalese, three Burmese, two Thai and one Somalian.
There were no differences in gender (ᵡ2 = 0.82, df = 2, p = 0.664) and ethnic distributions (ᵡ2 = 10.86, df = 6, p = 0.093) across the age groups. Although the proportion of gender differed by only 2.4% in this cohort, there was a slight shift of mortality distribution to the right, where more females were affected with a peak at 45–59 age group (Fig 1). On the other hand, mortality among males peaked at a younger age group of 30–44 years.
West coast of Malaysia reported the highest dengue mortality cases, with more than one-third from Selangor (34.5%), followed by Johor (15.2%), Kuala Lumpur (8.7%), Perak (8.1%), Penang (6.5%), Malacca (5.3%), Negeri Sembilan (3.4%), Kedah (1.6%) and Perlis (0.3%). East coast contributed 10% to this study cohort, with 6.5% from Kelantan, 2.8% from Pahang and 0.6% from Terengganu, apart from the Borneo states, Sabah (3.4%) and Sarawak (3.1%). Only 38.9% of these cases documented they stayed at dengue at-risk areas, where there were more than one suspected dengue cases from the same area within the last one month prior to their symptoms onset.
Patients did seek treatment early with a median duration of one day from symptom onset to first healthcare contact (Table 2). Majority (84%) presented themselves to local healthcare facilities, either clinic or hospital, in public or private setting, within the first three days of their symptom onset, with 22.3% presented within the first 24 hours. About 14.1% sought treatment between four and seven days of their illness, with the remaining 1.9% presented late after one week.
Although all patients were subsequently admitted to hospital for further management, there was a slight delay of approximately two days between first medical contact and hospital admission. Up to half of them were admitted within first three day, and 44.7% between four and seven days of their illness. Majority (76.4%) were managed in intensive care units (ICU), with 51.8% of them warranted immediate transfer to ICU upon hospital admission for close monitoring and stabilization.
About 8.4% experienced rapid deterioration and died within 24 hours of admission. Meanwhile, 6.9% of deaths occurred early within first three days of illness, while 66.5% occurred between four and seven days which fell within the critical phase of dengue. The remaining quarter died more than a week after their symptoms onset. There was no significant difference in the duration from symptom onset to the first medical contact, hospital admission, ICU admission and death, across the age groups. Among 283 cases with location of deaths available, 89% occurred at ICU, 5% at emergency departments, 3.5% at high dependency units and 2.5% in general wards.
Among 11 children aged <15 years with their body mass indexes (BMI) available, 36.4% were obese (≥95th centile of gender specific BMI-for-age), 36.4% were normal weight (5th-85th centile) and the remaining underweight (<5th centile). Out of the 114 adults, 26.3% were overweight (25–29.9kg/m2) with 42.1% were obese (≥30kg/m2). In the contrary, only 10% of the 20 elderly were obese and 30% were overweight, while more than half were of normal weight.
Nearly 90% of the elderly had underlying comorbid illness, but half of the children and 38.8% of the adult patients were free from any comorbid (ᵡ2 = 20.51, df = 2, p<0.001) (Table 3). Hypertension and diabetes mellitus were the most common comorbids in this cohort, especially among the elderly. In addition, dyslipidemia, heart diseases (ischaemic cardiomyopathy, heart failure and heart block), lung diseases (bronchial asthma and chronic obstructive pulmonary disease) and renal disease (chronic kidney disease) were more frequently observed among the elderly. Six adults had chronic liver disease, while another three had underlying malignancy.
All patients presented with non-specific symptoms of febrile illness. Although 78.1% presented with at least one warning sign, one in every five patients with severe dengue in our cohort did not present with any warning sign (Table 4). Among those presented with dengue warning signs, up to 31.2% had only one warning signs which could be non-specific to dengue. The most common warning signs were lethargy or restlessness (42.5%) followed by persistent vomiting (38.6%) and laboratory markers indicating haemo-concentration (34.7%). Lethargy or restlessness (ᵡ2 = 8.509, df = 2, p = 0.014) were commoner in children while abdominal pain or tenderness was observed more frequently in the adults (ᵡ2 = 7.533, df = 2, p = 0.023).
Only 109 patients (34%), including a case of unknown age, were clinically diagnosed of dengue fever during their first contact with local healthcare facilities. Among those suspected of dengue, nearly half (48.6%) were diagnosed with dengue fever without warning signs, followed by dengue fever with warning signs (21.1%), severe dengue (17.8%) and DSS (11.5%). As compared to children, more adults and elderly presented with symptoms leading to early diagnosis dengue fever with warning signs, severe dengue and DSS. Non-specific viral febrile illness and upper respiratory tract infections were the commonest clinical impressions for patients across all the three age groups (Table 5). Two children presented with symptoms suggestive of meningoencephalitis. More adults and elderly presented with symptoms leading to the diagnoses of acute gastroenteritis and pneumonia.
During the clinical assessment upon hospital admission, the index of suspicion for dengue had increased to 73.6%, among which 46.8% were diagnosed as severe dengue or DSS, 28.7% as dengue fever with warning signs and 24.5% were dengue infection without warning signs. Approximately a quarter of cases were remained diagnosed as non-dengue conditions, with a small proportion (2.8%) were treated as severe sepsis (Table 6). Some patients first sought medical care at hospitals leading to immediate admission, hence giving rise to overlapping of clinical diagnoses in Tables 5 and 6.
Among 285 cases (69.9%) had their sera samples tested reactive for dengue NS1 antigen, two-third had their tests performed between days three and five from their symptoms onset. However, about 17.3% were tested reactive within 48 hours of their illness, while 14.2% between days 6 and 10, with the remaining 1.8% detected between days 11 and 16.
IgM was detected in the sera of 196 patients (60.9%). Nearly 20% had their dengue IgM detected within 48 hours of their symptoms onset, 44.4% between days three and five, 32.1% between days 6 and 10, and 3.6% beyond day 10 of illness. One-third of the patients were immunoglobulin G (IgG) positive. Two patients with NS1 and IgM seronegative had their diagnoses confirmed with PCR.
As the disease progressed, all patients fulfilled at least one of the components for the severe dengue. Nearly 80% had severe organ involvement with 65% experienced severe plasma leakage and a third suffered from severe bleeding (Table 7). More adults had severe plasma leakage, as compared to children and elderly (ᵡ2 = 8.016, df = 2, p = 0.018). Neurological involvement in severe dengue were commoner in children (ᵡ2 = 9.392, df = 2, p = 0.009). On the other hand, heart involvement were observed more frequently among elderly (ᵡ2 = 7.802, df = 2, p = 0.020), especially those with underlying heart disease.
Among 140 patients with bleeding tendencies, 112 (80%) experienced severe bleeding. Acute gastrointestinal bleeding, involving either upper and/or lower gastrointestinal tract, was the commonest haemorrhagic complications (37.8%), followed by pulmonary haemorrhage (11.4%), occult bleeding (5%), intracranial bleeding (3.6%) and vaginal bleeding (3.6%). Hematuria (0.7%), intraabdominal haemorrhage (0.7%), and post-partum haemorrhage after Caesarean section (0.7%) were the rare complications reported. Ten patients developed disseminated intravascular coagulopathy. Other sites with milder bleeding included mucosa (gum and nasal cavity) (11.4%), venous puncture sites (4.3), and abdominal wall (0.7%).
In our cohort, 46.6% of the patients received transfusion of blood products. Seventy-eight patients (24.1%) required emergency dialysis for acute renal failure, with only nine of them had pre-existing renal disease. Nearly half (49.2%) of them required mechanical ventilatory support. Six patients had their bone marrow biopsy showing reactive haemophagocytosis activity.
More than half (55.7%) of these patients had DSS. More adults (59.9%) suffered from DSS, as compared to children (50%) and the elderly (42.9%) (ᵡ2 = 5.833, df = 2, p = 0.054).
Dengue IgG positivity was associated with higher proportion of severe plasma leakage (ᵡ2 = 6.889, df = 2, p = 0.032), but not DSS (ᵡ2 = 3.015, df = 2, p = 0.221), severe bleeding (ᵡ2 = 0.530, df = 2, p = 0.767) and severe organ involvement (ᵡ2 = 2.451, df = 2, p = 0.294).
Classifying the cause of death for each case remained challenging, as so for some patients, more than one cause might have contributed to their deaths. Overall, DSS was the most common cause of death among our cohort (72.9%), followed by severe organ dysfunction (69%) and severe bleeding (29.7%) (Fig 2). There was no significant difference in causes of death across different age groups in our cohort. Approximately half (51.9%) of the mortality were contributed by combinational causes, and 16.6% died of DSS complicated with both severe bleeding and severe organ involvement.
Dengue infection is associated with severe disease, and deaths occur despite supportive management. This is the first study at national scale to investigate epidemiology, clinical characteristics and causes of death of fatal dengue cases in Malaysia to date. This study also provides additional information to fill in the gap of understanding in term of the warning sign manifestations, clinical diagnoses at initial presentation, severe dengue and causes of death across three different age groups.
Over the recent years, there is a shift of the infection from affecting primarily children to adults [7, 8, 10]. The Malaysian Cohort Study in 2008 reported 91.6% of adults aged 35–74 years were seropositive for dengue IgG [11]. This rate increased with age, and reached nearly 100% by 60 years of age [9]. The mean age of the fatal cases in our cohort was 40.7 ±19.3 years, with 72.5% of them were from working age group of 15–59 years. We postulate the higher proportion of secondary dengue infection among adults which might have contributed partly to the age shift in mortality. Among 112 IgG seropositive patients in this cohort, 73% were from the adult group and 19% from the elderly. Secondary dengue infection has been associated with severe outcome of dengue via antibody-dependent enhancement [11] and T-cell original antigenic sin [12].
Some studies have observed that early care-seeking behaviour may determine the possibility of receiving optimum treatment and thereby avoiding fatal outcomes [13–15], and gender might affect the care-seeking behaviour patterns [16]. However, we were unable to demonstrate any association between the care-seeking behaviour pattern and gender, as well as age groups in this cohort. Majority of our patients sought treatment early regardless of gender and age groups. The median duration from onset of illness to first healthcare facilities visit was 1 day for children and elderly while 2 days for adults. Patients in this cohort were hospitalized early by a median duration of illness of 3 days as compared mean of 4.7 days as reported by Sam SS et al. [15]. Due to rapidly progressive clinical deterioration, they were also admitted to ICU earlier on a median of 3 days of illness in contrast to a mean of 5.6 days as reported in Singapore [13], but similar to 3.75 days in Cuba [17].
Dengue death was observed more frequently among patients who had sought care after the fourth or fifth day of fever, compared to those who sought care during the first three days of fever [17, 18]. However, 84% in this study was found to seek treatment early within first 3 days of illness. The slight delay of median 2 days between duration of fever onset to first visit to health care facilities and subsequent hospital admission reflected not only the challenges in identifying dengue in the early course of disease, but the dynamic progression of the infection itself, as 21.9% did not experience any warning signs at initial presentation. This observation is inconsistent with the expectation that timely treatment following early identification of the illness might have resulted in better survival outcome. In fact, only 35% of adults and 37.5% of elderly were clinically diagnosed with dengue at their first presentation to healthcare facilities. This was consistent with study by Jenny Low GH et al., who reported early clinical diagnosis based on WHO classification is more difficult in older adults [19]. Despite WHO had developed a set of guidelines to aid in diagnosis of dengue infection and disease classification, it remains challenging to differentiate dengue infection clinically from other causes of febrile illness, especially during the early phase of illness [20]. It may also be contributed by the atypical presentations among adults age >18 years as the frequency of symptoms and signs reported by adult population in accordance to the WHO classification schemes (1997 and 2009) reduced significantly with increasing age [19].
Using the WHO recommendations for BMI, the national prevalence of overweight and obesity among adults aged 18 years and above were 33.6% and 19.5% respectively [21]. Besides, the National Health Morbidity Survey III reported 19.9% of Malaysian children aged 7–12 years were overweight [22]. The fatal cases in our cohort appeared to have higher proportion of overweight and obesity as compared the national prevalence.
The presence of co-morbidities and other chronic illnesses might contribute to increased dengue mortality in adult population [13, 23, 24]. In this review, 64.9% had underlying co-morbid illness. Nonetheless, 38.8% adults and 10.7% elderly were healthy individuals without any co-morbids. This observation highlighted that underlying co-morbidities might not be the main factor contributed to the dengue mortality in our cohort.
This study also reviewed the warning signs of fatal dengue cases. It has been suggested that many warning signs are uncommon on their own, but lethargy or restlessness, abdominal pain or tenderness and mucosal bleeding were the three commonest occurring before the development of severe dengue [25]. This review reported the similar finding that lethargy was the commonest warning sign observed in the cohort. However, none of the warning signs presented in more than half of these fatal cases. Interestingly, 21.9% did not have any warning signs upon hospital admission. This finding was lower compared to a study by Thein TL et al. where 58% of their cohort did not have any warning signs prior to development of severe dengue [25]. This gap was probably attributed to the different study populations whereby our study focused only on fatal cases and excluded survivors of severe dengue. While lethargy or restlessness was commoner among children and abdominal pain or tenderness was observed more often in adults, there was no difference in other warning signs at presentation across the three age groups. This is in accordance to abdominal pain and lethargy being well-established warning signs which contributed to severe dengue and dengue deaths [25–27], but there is limited literature illustrating the association of warning signs with different age groups.
Being age >40 years was an independent risk factor for the development of severe dengue [28]. More adults in this review developed severe plasma leakage as compared to children and elderly. There is anecdotal evidence suggesting both hepatic dysfunction and bleeding manifestations are more common in older age groups [13, 29, 30], but not in our study population. In contrast, we observed a higher proportion of heart involvement among the elderly whom with underlying heart diseases, which may have been exacerbated by the effects of acute infections. Neurological involvement was observed more frequently among children, which is consistent with a review by Puccioni-Sohler M et al. [31], yet the neuropathogenesis of dengue infection remains poorly understood. Some proposed possible mechanisms for neurological involvement are direct viral infection of central nervous system, autoimmune reaction, metabolic and hemorrhagic disturbances [32].
In this study, we attempted to classify the primary cause of deaths to DSS, severe bleeding and/or severe organ dysfunction, based on the causes of death retrospectively determined. DSS was the commonest cause of death (72.9%) in our cohort, followed by severe organ involvement (69%) and severe bleeding (29.7%). Our findings are similar to another review on fatal cases from Singapore which reported shock as the commonest cause of death followed by organ impairment (71.4% with acute renal impairment, 57.1% with impaired consciousness and 53.6% with severe hepatitis) [33]. There was slightly higher proportion of deaths attributable DSS alone among children as compared to adults and elderly. Gamble J et al. reported children are more prone to develop hypovolaemic shock as compared to adults in any conditions characterized by increased microvascular permeability because they have a larger microvascular surface area per unit volume of skeletal muscle, and the proportion of the developing vessels is greater during development [34]. These developing microvessels are known to be more permeable to water and plasma proteins than when they mature [34].
The strength of this study lies in our attempt to review all dengue-related mortality cases reported countrywide between 2013 and 2014, according to different age groups. Each case was reviewed extensively and causes of death were retrospectively determined and revised by an independent panel of consultant physicians at dengue mortality meeting. In our attempt to report the basic epidemiological characteristics of these fatal cases and their clinical course of their illness from initial presentation until deaths ensued, we recognize the lack of access to the individual case record imposed a quality limitation to clinical data collection. Although we attempted to collect relevant data in details, we acknowledge the possibility of recall bias by the managing team, information and reporting biases during documentation process. The absence of detailed information with regards to organ involvement in organ failure and those resulted in death, as well as management plan for patients during their illness have led to limited analyses in this study. We have identified some but not all the co-morbidities, especially in the elderly as limited by the retrospective nature of this review. We were unable to assess the risk factors for severe dengue and associated mortality due to lack of a control group. All patients in this study were hospitalized and the current surveillance system is yet to be strengthened to capture dengue-related deaths occur at home or as out-patients.
Dengue is a dynamic disease which its clinical progression can evolve beyond anticipation. In Malaysia, majority of the dengue-related deaths occurred in adults of working age group between 15–59 years. Although many have sought treatment early, the index of suspicion for diagnosing dengue remained low at their initial presentation. More than 60% subsequently succumbed to DSS and severe organ involvement refractory to aggressive intervention.
|
10.1371/journal.pcbi.1002223 | Transcriptomic Coordination in the Human Metabolic Network Reveals Links between n-3 Fat Intake, Adipose Tissue Gene Expression and Metabolic Health | Understanding the molecular link between diet and health is a key goal in nutritional systems biology. As an alternative to pathway analysis, we have developed a joint multivariate and network-based approach to analysis of a dataset of habitual dietary records, adipose tissue transcriptomics and comprehensive plasma marker profiles from human volunteers with the Metabolic Syndrome. With this approach we identified prominent co-expressed sub-networks in the global metabolic network, which showed correlated expression with habitual n-3 PUFA intake and urinary levels of the oxidative stress marker 8-iso-PGF2α. These sub-networks illustrated inherent cross-talk between distinct metabolic pathways, such as between triglyceride metabolism and production of lipid signalling molecules. In a parallel promoter analysis, we identified several adipogenic transcription factors as potential transcriptional regulators associated with habitual n-3 PUFA intake. Our results illustrate advantages of network-based analysis, and generate novel hypotheses on the transcriptomic link between habitual n-3 PUFA intake, adipose tissue function and oxidative stress.
| A fundamental goal in the field of nutritional genomics is defining the molecular link between diet and health. Human nutritional genomic studies are frequently hindered by a high level of unexplained variation in gene expression, protein and metabolite abundance, and clinical parameters – potentially attributable to variation in genotype, background diet, anthropometry, physical activity and health status. In our present study, relationships between adipose tissue gene expression, habitual diet and clinical markers of metabolic health were investigated in a cohort of individuals with impaired metabolic health, typical of the Metabolic Syndrome (MetS). Using multivariate statistics in conjunction with a novel approach to metabolic network analysis, we identified regions of the human metabolic network showing coordinated transcriptomic response to variation in n-3 PUFA intake and correlation with markers of metabolic health.
| Dietary fat intake has profound effects on molecular processes of metabolic health. These effects are diverse and often subtle, representing a considerable analytical challenge in reaching system-level understanding. Transcriptomics has become a central technology in the development of molecular nutrition, having the capacity to produce expression data for every gene in a given genome. However, the major challenge is to apply appropriate techniques for extracting information from high-throughput datasets. Differentially expressed gene lists are an intuitive first choice, but they are hard to interpret in a biological context. Pathway analysis – typically implemented using gene set enrichment analysis – has become a standard method in the field of transcriptomic analysis [1], [2]. It is easy to implement and can simplify and contextualize large lists of differentially expressed genes, although this approach possesses technical limitations due to inherent redundancy among pathways and interconnectedness between one pathway and the next. Failure to appropriately account for these features can substantially limit biological interpretation of high-throughput datasets.
Network-level analysis has revealed detailed insight on metabolic regulation in type 2 diabetes and insulin resistance [3], [4]. Del Sol et al., have proposed that in the emerging systems-level view of molecular biology, diseases should be viewed as a function of network perturbation rather than as isolated local changes [5]. Molecular networks may be classified in two categories: metabolic networks and protein interaction networks. Metabolic networks are inclusive, intuitive abstractions for representing system-level metabolism, as they incorporate all known metabolic interactions in a given species. Due to their size and complexity, however, they are analytically challenging. Previously applied analytical approaches include topological analysis (e.g., identification of hub nodes and functional modules) [6] and reporter metabolite analysis [4].
A number of methods exist for analyzing transcriptomic data in the context of a global interaction network [7]. The majority of these methods focus on protein interaction networks, and aim to partition a global network into clusters/modules of genes, and identify clusters showing coordinated transcriptomic response. When modelling transcriptomic activity in metabolic networks, however, it is instructive to use path (rather than cluster) constructs, because paths match the native pattern of energy flux through a metabolic network. Cluster-based analysis of metabolic network activity performs well in identifying regions of a network with altered transcriptomic activity, but the identified clusters may contain disconnected sections of different paths of metabolite conversion. A given path of interest may thus be fractioned across several different neighbouring clusters, making it difficult to identify coordinated alteration of activity across that entire path. We therefore defined a method that identifies altered transcriptomic activity in the context of network paths rather than clusters. With this approach, we identified local coexpressed paths in the metabolic network showing covariance with recorded dietary intake of n-3 PUFA, and correlation with a urinary marker of oxidative stress. In a parallel analysis, investigation of the promoter regions of n-3 PUFA-correlated genes highlighted significantly over-represented binding sites for transcription factors related to adipogenesis.
The LIPGENE human dietary intervention study was a randomized, controlled trial that complied with the 1983 Helsinki Declarations, approved by the local ethics committees of the 8 intervention centres (Dublin, Ireland; Reading, UK; Oslo, Norway; Marseille, France; Maastricht, The Netherlands; Cordoba, Spain; Krakow, Poland; Uppsala, Sweden). Written informed consent was attained from every participant as approved by each institutional ethical committee.
The current study was conducted within the framework of the LIPGENE Integrated Project “Diet, genomics and the metabolic syndrome: an integrated nutrition, agro-food, social and economic analysis” (Clinical Trials. gov number: NCT00429195) and NuGO, The Nutrigenomics Organization (www.nugo.org); both European Union FP 6 initiatives. The subjects participated in the LIPGENE human dietary intervention study [8], although only baseline, pre-intervention samples were used in the present study. Samples were collected under standardised conditions according to a strict SOP [9]. Briefly, volunteers attended the clinics following a 12 h overnight fast; they were asked to abstain from alcohol, medications or vigorous exercise in the 24 h prior to assessment. For inclusion in the study, volunteers were required to be age 35–70 years, BMI 20–40 kg/m3 and show 3 or more of the following MetS criteria (based on slightly modified NCEP ATP-III): fasting plasma glucose 5.5–7.8 mmol/L, serum TAG ≥1.5 mmol/L, serum HDL-cholesterol <1.0 mmol/L in males, and <1.3 mmol/L in females, waist circumference >102 cm in males and >88 cm in females, and elevated blood pressure (systolic blood pressure ≥130 mmHg, diastolic blood pressure ≥85 mmHg or on prescribed blood pressure lowering medication). Habitual dietary intake was monitored for each volunteer by a 3 day weighed food dietary record, and assessed for daily intake of energy, carbohydrate, protein, fat, saturated fat (SFA), monounsaturated fat (MUFA), polyunsaturated fat (PUFA), and n-3 and n-6 PUFA [10]. Extensive metabolic profiling including plasma markers of inflammation, fatty acid pattern, plasma lipoproteins and apolipoprotein profiles, and markers of insulin sensitivity (Table 1), was performed as described by Tierney et al. [9]. Means and standard deviations for all dietary and plasma marker variables in our cohort are provided in supplementary Tables S1 and S2.
Subcutaneous adipose tissue samples were taken from the periumbilical area of 19 volunteers from the Norwegian and Spanish cohorts (10 female, 9 male) after an overnight fast. Needle biopsies were obtained after a 5 mm transdermal incision under local anaesthesia. Samples were rinsed in saline, put in RNA later and frozen immediately (−80°C) for subsequent analysis. Total RNA was extracted from adipose tissue using the RNeasy lipid tissue mini kit (Qiagen, U.K.). Briefly, 100 mg of adipose tissue was homogenised in Qiazol lysis reagent. After addition of chloroform, the homogenate was centrifuged to separate the aqueous and organic phases. Ethanol was added to the upper aqueous phase, and applied to the RNeasy spin column, where the total RNA was bound to the membrane, and phenol and other contaminants were washed away. RNA was then eluted in RNase-free water.
Extracted RNA was sent to ServiceXS (a high-throughput data service provider; www.servicexs.com) for labelling with the 3’ IVT express kit and hybridization to Affymetrix arrays. The microarray platform used in this study was custom designed by NuGO, and contained 16554 probe sets. This platform is designated ‘nugohs1a520180’, and we used the ‘entrezg’ version 12.1.0 annotation from the MBNI custom cdf database, reflecting the latest remapping of Affymetrix probes based on current data in the NCBI database (http://brainarray.mbni.med.umich.edu). Raw and GCRMA-normalized data are available from the Gene Expression Omnibus database, under accession GSE28070.
Raw microarray data were first assessed for quality using a set of standard QC tests, including array intensity distribution, positive and negative border element distribution, GAPDH and β-actin 3’/5’ ratios, centre of intensity and array-array correlation check. All QC tests were implemented in the R programming language (Version 2.11.1l, R Foundation for Statistical Computing), using the affyQCReport library. A batch effect was noted due to the arrays being hybridized on two separate days; thus, all subsequent analyses accounted for this effect by including batch number as a covariate in statistical models. It was also noted that the β-actin 3’/5’ ratios were higher than recommended (i.e., greater than 3-fold intensity difference) in most samples, although this ratio has been shown to be higher when cRNA is synthesized using the Affymetrix 3’ IVT express kit, particularly with low input RNA quantities (http://media.affymetrix.com/support/technical/whitepapers/3_ivtexpresskit_whitepaper.pdf). Furthermore, all samples except 2 showed 3’/5’ GAPDH ratios within expected range of 1.25-fold, and no samples appeared suspect in RNA degradation plots. The 2 samples that did not meet the GAPDH ratio recommendation were removed from further analysis. All QC-verified samples were background corrected and normalized using the GCRMA normalization method, which accounts for nucleotide specific differences in hybridization efficiency. The normalized dataset was then filtered to remove genes with Mas5 ‘absent’ call on all arrays, and those showing the lowest 10% variance, resulting in a final dataset of 10618 genes.
Diet and plasma marker variables were first normalized with log or square root transformation as appropriate to reduce skewness and kurtosis. Sparse partial least squares regression (sPLS; [11]) and regularized canonical correlation analysis (rCCA; [12]) were used to assess relationships between dietary components and gene expression levels, and between clinical markers and gene expression levels, respectively. The mixOmics library of R functions was used to carry out the analysis [12]. Specifically, the spls function was used to fit the sPLS model, and the network function to produce the network of interactions. An sPLS model was fitted using dietary variables, sex, nationality and array batch number as predictors (sex, nationality and array batch number were included in order to identify and control for correlations between gene expression and these variables), and gene expression as response variables. Choice of PLS dimensions was determined using the Qh2 variable previously proposed by Tenenhaus [13] which measures the relative contribution of each dimension h to the predictive power of the PLS model (see 11 and 13 for further details on sPLS and use of Qh2). With this approach, we retained 5 dimensions in the model, and retained all diet-gene pairs showing a similarity score >0.7 (using ‘threshold’ argument of the network function in the mixOmics library). This similarity score is a convention used in multivariate statistical methods; it ranges from 0 to 1, and corresponds to the distance between two given variables in the number of chosen dimensions [14].
The mixOmics library was used for rCCA modelling of plasma marker and gene expression data. The rcc function was used to define the canonical correlations and the canonical variates, estim.regul for estimation of regularization parameters and the network function to produce the network of interactions. In this case, datasets were not interpreted as predictors or responses given the more complex two-way relationship between plasma marker profile and tissue gene expression. Initial rCCA modelling including all plasma markers showed that correlations between gene expression and plasma fatty acid and lipid profile were so strong that they masked more subtle correlations between the remaining plasma markers and expression data. Consequently, separate rCCA analyses were performed: first, comparing adipose tissue gene expression to plasma fatty acids, lipids and apolipoprotein profile (including sex, nationality and array batch number as variables in the model); and second, comparing gene expression to plasma cytokines, IVGTT measurements, prostaglandin and urinary isoprostane (as before, including sex, nationality and array batch number as variables in the model). For comparison of gene expression vs. plasma fatty acids, lipids and apolipoproteins, the first 11 dimensions were retained in the model (as subsequent dimensions did not provide additional information to the model) and all gene-plasma marker pairs with a similarity score >0.75 were retained for subsequent analysis. Due to the very strong relationship between gene expression and plasma fatty acids, we observed that using a similarity score threshold of 0.7 resulted in a very high number of plasma marker-gene correlations (571 plasma marker-gene pairs passing threshold). Therefore, the higher threshold was chosen in this comparison in order to highlight only the strongest plasma marker-gene correlations, thereby facilitating downstream biological interpretation. For the second comparison (gene expression vs plasma cytokines, IVGTT measurements, prostaglandin and urinary isoprostane) the first 6 dimensions and all gene-plasma marker pairs with a similarity score >0.7, were retained in the model.
We used the Edinburgh human metabolic network reconstruction [15]; (www.ehmn.bioinformatics.ed.ac.uk/), which contains reaction information for 1627 unique metabolites and 1371 unique metabolic enzymes. In its native form, this reconstruction is a metabolite-centred network (i.e., nodes represent metabolites and edges are the enzymes that catalyze reactions between metabolites). For our analysis, the network was first transformed to an enzyme-centric construction (where 2 genes/proteins are linked if gene 1 produces a metabolite that is used as a metabolic substrate by gene 2). As is the norm in topology-based network analyses, we excluded currency metabolites (such as H2O, ATP and O2) from the network [16].
To identify paths of interest in the global interaction network, Dijkstra's shortest paths [18] were calculated from each diet-sensitive node to all others in the coexpressed subset of the global network (as determined above), taking into consideration directionality of node pair interactions. An algorithm was written in R to evaluate metabolic feasibility of each putative path – i.e. whether an unbroken path of metabolite conversion could be traced from one end to another (most recent scripts available on request). This concept of metabolic feasibility is an important consideration in analysis of global networks, because a connected path through the network does not necessarily indicate an unbroken path of metabolite conversion. Figure 1 illustrates the rationale behind metabolic feasibility (see supplementary Figure S1 for detailed description of the algorithm). The output of this algorithm is a list of feasible paths of metabolite conversion, wherein each path is strongly coexpressed (i.e., between each gene pair in the path) and possesses a diet-correlated gene at the upstream end. To our knowledge, this is the first metabolic network analysis algorithm that explicitly considers metabolic feasibility and adjacent pair-wise coexpression in analysis of network paths. This represents an informative alternative to network clustering analysis.
The TFM-explorer tool [19] was used to identify significantly over-represented transcription factor binding sites (TFBSs) among genes with expression showing positive correlation with n-3-PUFA intake. Using the promoter regions spanning -2000+200 bp relative to the transcription start site and all vertebrate transcription factor matrices from the Jaspar database (jaspar.cgb.ki.se), TFM-explorer returned all TFBSs that were significantly over-represented at a level of p<0.0001.
Results from sPLS indicated that among all dietary variables, the registered dietary intake of n-3-PUFA showed the strongest covariance with adipose tissue gene expression. Of the 53 n-3-PUFA-correlated genes identified in the sPLS analysis, 41 positively correlated and 12 negatively correlated with n-3-PUFA intake (Figure 2; Supplementary Table S3). Dietary intake of MUFA also showed strong covariance with expression of three of these genes (one positive: GALNTL1; two negative: CDIPT, PRPS1). It was also noted that the expression level of these three genes covaried in opposing direction with intake of n-3-PUFA and MUFA, reflecting the inverse relationship between habitual dietary consumption of these two dietary fatty acids in our population. To assess if > = 53 n-3 PUFA-correlated genes would be detected by chance alone, we permuted the sample labels and re-ran the sPLS analysis 100 times. These permutation tests yielded an average of 2.38 and median of 0 n-3 PUFA-correlated genes, suggesting that the 53 genes identified in our original dataset were unlikely to be identified by chance alone.
rCCA results showed that among the measured plasma lipids, fatty acids and apolipoproteins, plasma DHA, stearic acid and EPA correlated most strongly with adipose tissue gene expression (Figure 2; Supplementary Table S4). At the chosen threshold of 0.75, DHA [C22:6 n-3] correlated with the expression of 113 genes, followed by plasma stearic acid [C18:0]: 60 genes, and EPA [C20:5 n-3]: 21 genes. Comparison of sPLS and rCCA results highlighted 26 genes that were related to dietary n-3-PUFA intake as well as plasma DHA levels, reflecting the expected correlation between dietary fat intake and plasma fatty acid profile [20]. Among markers of inflammation, oxidative stress and insulin resistance, urinary 8-iso-PGF2α correlated most strongly with adipose gene expression, resulting in 96 gene correlations, 48 of which also correlated with plasma DHA (Figure 2; Supplementary Table S5). Any variables not included in Figure 2 (e.g., IVGTT) did not correlate with any adipose tissue genes at the chosen threshold.
The complete metabolic network included 1371 nodes and 65637 directed edges; the transcriptionally coexpressed (TC) subset contained 602 nodes and 5414 directed edges (supplementary Figure S2). To identify the biological functions predominant in this network, the largest connected subset of the TC subset network was partitioned into topological modules using a simulated annealing approach [21], as implemented by the spinglass.community function in the igraph library in R. This modular partitioning identified 3 topological modules. Hypergeometric tests were performed using the Category library in R to identify significantly over-represented gene ontology (www.geneontology.org) ‘biological process’ terms in each module. Briefly, over-represented terms in the first module related primarily to phosphatidylinositol and lipid metabolism; those in second related to nucleic acid metabolism; and the third module was more heterogeneous, consisting of cellular ketone metabolism, red-ox processes, and lipid and protein catabolism terms (see supplementary Table S6 for expanded results).
Diet-sensitive path extraction from the TC network revealed 755 unique paths greater than length 2 originating from 30 n-3 PUFA-sensitive genes, although paths leading from each diet-sensitive gene collapsed into tree-like structures (Figure 3). The most complex n-3 PUFA-sensitive path (in terms of path size and link density) centred on the AK1 gene (Figure 3A). The genes in the AK1 path are mostly involved in the highly redundant processes of energy and nucleotide metabolism, explaining the high link density in this region of the network. The majority of metabolic links in the AK1 path are different nucleotides and energy metabolism cofactors such as ATP, ADP and AMP. These metabolites are normally classified as currency metabolites and were removed from the rest of the network, although they were retained in this region where they act as primary reactants and products. Of the nodes in this path, an additional 7 correlated with dietary intake of n-3 PUFA, 28 strongly correlated with plasma fatty acid levels, and 14 with urinary 8-iso-PGF2α, suggesting that activity in this region of the metabolic network is sensitive to dietary intake of n-3 PUFA and correlated with metabolic health.
The ANXA3, PTEN and MTMR12-linked paths (Figure 3B-D) are interesting from a biological perspective because they each incorporate elements of lipid metabolism. The ANXA3-linked path primarily includes reactions involved in metabolism of glycerolipids, glycerophospholipids, arachidonic acid (AA) and linoleic acid (LA). ANXA3 is connected to ADH5 and LPL via the glycerol metabolite, which is further metabolized by LIPA and DGKA to form 1,2-diacyl-sn-glycerol (1,2-diacylglycerol) and phosphatidate on one branch of the path, and by ALDH isoforms to form d-glyceraldehyde and d-glycerate on the other. Five genes in this path correlated strongly with plasma levels of DHA, stearic acid, dihomo-gamma-linolenic acid and/or EPA.
The PTEN-linked path includes reactions that metabolize inositol phosphate- and lipid-related metabolites. PTEN is linked to OXSM and PLCL2 via the 1-phosphatidyl-D-myo-inositol 4,5-bisphosphate, which is metabolized by these enzymes to form 1-phosphatidyl-myo-inositol 5-phosphate and 1,2-diacyl-sn-glycerol. This 1,2-diacyl-sn-glycerol is further metabolized by DGKA to form phosphatidate. Two genes in this path – PTEN and OXSM – were inversely correlated with urinary 8-iso-PGF2α, and three –PLCL2, DGKA and CDS2 – were positively correlated with plasma DHA level.
The MTMR12 path (Figure 3D) is also linked to lipid metabolism via phosphatidylcholine, through a more complex upstream path involving inositol phosphate derivatives. At the downstream end of this path, cytochrome p450 enzymes (MYP19A1, CYP2B6 and CYP1B1) act on AA and LA as substrates to form diverse epoxyeicosatrienoic acids (EET), hydroxyeicosatrienoic acids (HETE) and epoxyoctadecenoic acids (EpOME), involved in the resolution of inflammation with subsequent relevance to cardiovascular disease [22], [23]. Two genes in this path – PIK3CAand PIP5K1A – were negatively correlated with urinary 8-iso-PGF2α; PIKFYVE, PIK3CA, and CEPT1 were positively correlated with plasma DHA, and PIP5K1A with plasma stearic acid and dihomo-gamma-linolenic acid.
To assess whether a similar group of paths would be extracted from any TC network – e.g., due to higher connectivity in certain regions of the network – we generated such a TC network from publicly available muscle tissue microarray data from obese individuals (GEO accession: GSE474), and extracted paths leading from the n-3 PUFA-sensitive genes identified in the present study. In this muscle tissue TC network we found only 24 paths of maximum length three leading from ten of the n-3-PUFA-sensitive genes (supplementary Table S7). Furthermore, these paths did not intersect to form a larger sub-network.
To compare our network analysis with a standard approach to pathway analysis, hypergeometric tests were performed to identify KEGG pathways significantly enriched (using the hyperGTest function in the R ‘Category’ library) for the n-3PUFA-sensitive genes identified in our sPLS analysis. This analysis returned four KEGG pathways greater than length four (Table 2). Of these pathways, the top three - biosynthesis of plant hormones, biosynthesis of terpenoids and steroids and biosynthesis of alkaloids derived from terpenoid and polyketide - have 45 genes in common. Accordingly, the same 7 n-3 PUFA-sensitive genes (PMVK, FDFT1, ALDOA, IDH3G, PGK1, SDHB, GGPS1) were present in each pathway. Thus, the apparent enrichment of the biosynthetic plant hormones pathway is probably an artifact of the high degree of overlap with other pathways in the database. Closer inspection of these pathways in the KEGG database showed that they are large, diverse and disjointed pathways, including many parallel processes. For example, the pathway for biosynthesis of terpenoids and steroids includes of subsets of glycolysis, limonene and pinene degradation, terpenoid backbone biosynthesis, carotenoid biosynthesis and geraniol degradation. The n-3 PUFA-correlated genes were distributed across these processes rather than occurring in a single one.
Figure 4 illustrates over-represented TFBSs among genes showing positive correlation with the intake of n-3 PUFA (from our sPLS results). The three most significantly over-represented TFBSs were those of Krüppel-like factor 4 (KLF4), specificity protein 1 (SP1) and E2F1 transcription factors. KLF4 is involved in adipogenesis, specifically by binding to the promoter of the CEBPB (C/EBPβ) gene [24]. CEBPB was not present in the Edinburgh human metabolic network (as it is not a metabolic enzyme). Thus, an expanded sPLS analysis was performed, comparing dietary intake to all genes on the Affymetrix microarray. This analysis identified CEBPB expression to be positively correlated with n-3 PUFA intake (data not shown). Although no additional adipogenic genes were identified at this correlation threshold of 0.7, reducing the threshold to 0.6 revealed a number of additional adipogenic factors showing positive correlation with n-3 PUFA intake – including ADIPOQ, BMP2, CFD, FABP4, LIPE, LPL and PLIN.
SP1 is a broadly acting transcription factor operating in conjunction with NF-YA, SREBP and PPARγ in promoting lipogenesis. The NF-YA and PPARγ TFBSs were also significantly over-represented in our group of n-3 PUFA-sensitive genes, although SREBP TFBS was not. E2F1 is a transcription factor involved in early adipogenesis, and positively regulates transcription of PPARγ [25]. Interestingly, all but one PPARγ TFBS-containing genes in our sample also contained an E2F1 TFBS.
An emerging limitation to pathway analysis of transcriptomic data is that documented pathway models tend to overlap and intersect, yielding analytical results that are biased, incomplete or both [26]. Our network approach did not segregate metabolic processes into discrete pathway models, thereby revealing inherent overlaps and intersections between pathways. Examples of this pattern are seen in the ANXA3, PTEN and MTMR12 paths, which incorporate connected reactions from the metabolic pathways of inositol phosphate derivatives, glycerolipids, glycerophospholipids, arachidonic and linoleic acid. Intersections of these pathways are clear when viewed in the network context, but less evident when each canonical pathway is assessed separately. Cross-talk between metabolism of lipids/fatty acids and inositol phosphate derivatives plays an important role in the induction of signalling cascades by dietary and fat [27]. Figure 3B-D illustrates paths of triglyceride metabolism including formation of intermediate metabolites such as diacylglycerol and phosphatidate. These metabolites act as signalling molecules that affect a wide range of cellular functions like insulin signalling, inflammation, cellular differentiation and proliferation and oxidative stress [28], [29]. Thus, it is of particular interest that genes in the PTEN and MTMR12 paths show strong inverse correlation with urinary 8-iso-PGF2α – a marker of systemic oxidative stress.
Previous work has described an inverse relationship between n-3 PUFA intake and n-6 fatty acid-derived prostaglandins (e.g., PGF2α) in the plasma of Alzheimer's disease patients [30], urine of healthy males [31] and plasma and urine of dyslipidaemic and type 2 diabetic individuals [32]. A prevalent hypothesis for this relationship is that n-3 and n-6 PUFA compete for the same enzyme systems, and consequently, increased n-3 PUFA intake precludes production of n-6 fatty acid-derived prostaglandins [33]. In addition, n-3 PUFA may exert independent anti-inflammatory effects through unique receptors and enzyme systems, regardless of n-6 fatty acid intake [34]. Our analysis identified 17 genes showing opposing direction of correlation with n-3 PUFA intake and urinary 8-iso-PGF2α (Figure 2). Furthermore, results from network analysis identified precise coexpressed regions of the metabolic network showing positive correlation with dietary n-3 PUFA intake and plasma DHA, and negative correlation with urinary 8-iso-PGF2α in the cohort of MetS subjects (Figure 3). The sub-network illustrated in Figure 3A is interesting in this context because it contains numerous members of the electron transport chain, including 20 isoforms of ATPase/ATP synthase, six of which were negatively correlated with 8-iso-PGF2α. Further investigation of diet-dependent energy flux through these network regions may provide insight on the precise relationship between n-3 PUFA intake and adipose tissue oxidative stress. Comparing n-3 PUFA intake directly to urinary 8-iso-PGF2α resulted in only a near-significant trend (p = 0.077; p = 0.0828 after adjusting for sex); a recent publication of findings from the larger LIPGENE cohort reported a similar near-significant trend [35]. This may be due to the number of molecular intermediates between dietary intake and urinary output, highlighting the increased clarity provided by analysis of tissue-level high throughput data in the framework of a global metabolic network.
To understand the potential regulatory consequences of dietary n-3 PUFA intake on adipose tissue biology, we analysed the promoter regions of n-3 PUFA-correlated genes to identify significantly over-represented transcription factor binding sites. Results from this analysis highlighted significantly over-represented transcription factors related to adipogenesis. The most strongly over-represented transcription factors were KLF4, SP1 and E2F1. SP1 and KLF4 share similar GC-rich target binding sites [36], which is evident in their overlapping binding sites in Figure 4. Although limited work has focused on joint activity of these transcription factors in adipose tissue, KLF4 has been shown to inhibit SP1 activity in the gut by competitive TFBS binding [37]. PPARγ was also identified as significantly over-represented among genes correlated with habitual n-3 PUFA intake. PPARγ is arguably the most well-studied transcription factor in the field of diet-related transcriptomic regulation, and is the subject of many reviews on the subject [38], [39], [40], [41]. In addition to its role as a central regulator of adipogenesis, lipid storage and combustion, PPARγ also protects against oxidative stress and inflammation [42]. Dietary n-3 PUFA are potent inducers of PPARγ expression in adipocytes [43] and preadipocytes [44]. Accordingly, n-3 PUFA supplementation has yielded positive effects on weight gain in cancer [45] and Alzheimer's patients [46], and reduced lipotoxicity in a range of experimental models [47]. Future work should clarify the contribution of these adipogenic transcription factors to adipose tissue function, particularly given the positive correlation in our present study between dietary n-3 PUFA intake and expression of additional adipogenic genes, including CEBPB, ADIPOQ, BMP2, CFD, FABP4, LIPE, LPL and PLIN1.
In conclusion, we have taken a joint multivariate and network-based approach to transcriptomic analysis, relying on known metabolic reaction information to reveal coordinated paths of metabolite conversion. This approach highlighted coexpressed regions of the metabolic network with opposing direction of correlation with habitual n-3 PUFA intake and urinary isoprotane levels - relationships that were not identified using a traditional pathway enrichment test. Promoter analysis further highlighted adipogenic transcription factors as potential transcriptional regulators of n-3 PUFA-correlated genes.
|
10.1371/journal.pcbi.1005406 | A magnesium-induced triplex pre-organizes the SAM-II riboswitch | Our 13C- and 1H-chemical exchange saturation transfer (CEST) experiments previously revealed a dynamic exchange between partially closed and open conformations of the SAM-II riboswitch in the absence of ligand. Here, all-atom structure-based molecular simulations, with the electrostatic effects of Manning counter-ion condensation and explicit magnesium ions are employed to calculate the folding free energy landscape of the SAM-II riboswitch. We use this analysis to predict that magnesium ions remodel the landscape, shifting the equilibrium away from the extended, partially unfolded state towards a compact, pre-organized conformation that resembles the ligand-bound state. Our CEST and SAXS experiments, at different magnesium ion concentrations, quantitatively confirm our simulation results, demonstrating that magnesium ions induce collapse and pre-organization. Agreement between theory and experiment bolsters microscopic interpretation of our simulations, which shows that triplex formation between helix P2b and loop L1 is highly sensitive to magnesium and plays a key role in pre-organization. Pre-organization of the SAM-II riboswitch allows rapid detection of ligand with high selectivity, which is important for biological function.
| The presence of positively charged metal ions is essential to maintain the structural fold and function of RNA. Among different metal ions, magnesium is particularly important for the stability of RNA because it can efficiently support a close assembly of negatively charged phosphate groups in an RNA fold. The SAM-II riboswitch is an example of a classical pseudoknot fold, which binds S-adenosyl methionine, stabilizing an alternate folded form to inhibit gene expression. In our early 13C- and 1H-chemical exchange saturation transfer (CEST) experiments, we found a conformational transition between a minor, partially closed and a major, open state conformation in the absence of ligand. Our CEST experiments at different magnesium concentrations now suggest that magnesium ions can induce a conformational pre-organization in the apo SAM-II riboswitch, which is expected to facilitate ligand binding. To understand the microscopic details of this magnesium-induced transition, we perform all-atom structure-based molecular simulations including electrostatics and explicit magnesium ions. Our free energy calculations reveal that the partially closed pre-organized state is further stabilized with increasing magnesium concentration. This is in excellent agreement with our 13C-CEST profile, SAXS, and size-exclusion chromatographic data, and with recent single molecule FRET experiments. Our results suggest that a sufficiently high concentration of magnesium is essential to pre-organize the apo SAM-II riboswitch.
| Non-coding RNAs are currently thought to account for over 75% of the human genome [1]. In bacteria, non-coding RNAs play important roles in gene regulation. One such class of RNAs, riboswitches, regulates metabolite production. Here, a single RNA sequence folds into one of two or more mutually exclusive folds depending on the metabolite concentration [2,3]. In some cases, such as the S-adenosylmethionine-I (SAM-I) riboswitch, the RNA contains a transcriptional terminator that forms when ligand is present, in effect silencing genes important for ligand production [4–6]. When the ligand is not present, the terminator does not form, allowing gene expression, and therefore ligand production, to continue efficiently. In other cases, such as the SAM-II riboswitch, ligand binding may lead to sequestration of the Shine/Dalgarno sequence, likely blocking ribosome binding and, as a consequence, protein synthesis [7]. While these examples of ligand-dependent secondary structure switches have been known for some time, a detailed thermodynamic understanding at the atomistic level, including the indispensable effect of the RNA’s ion-atmosphere, has not been achieved. In recent years, riboswitches have become canonical systems for studies of diverse RNA behaviors, as they possess quintessential characteristics of many RNA systems: ligand binding, Magnesium ion (Mg2+) sensitivity, conformational changes, secondary structure remodeling, and regulatory functions. Chemical footprinting, NMR, small-angle X-ray scattering (SAXS) and single molecule FRET techniques are being exploited to elucidate the folding kinetics, thermodynamics and the magnesium ion sensitivity in RNA systems such as the TPP riboswitch [8], glycine-dependent riboswitches [9], different variants of P4-P6 RNA [10–12] and P5abc subdomain of the Tetrahymena group I intron ribozyme [13]. Other work focuses on more complex functions, such as splicing and ligand recognition and their associations with proteins or different metabolites [14,15]. Success in understanding the structural, dynamical and functional aspects of riboswitch systems requires an integrated experimental and theoretical approach. Traditional crystallographic techniques produce static snapshots of the riboswitch. SmFRET, NMR, and SAXS methods obtain kinetic information and overall distributions of conformations. Molecular simulation allows one to integrate disparate experimental data into a single coherent picture, characterizing transitions in atomistic detail and the free energy landscape with fine resolution. A large number of riboswitches have been crystallized and have also been investigated via fluorescence and single molecule techniques [16–28].
Molecular simulations have also been used to study a number of riboswitches, including but not limited to the SAM-I, SAM-II, pre-Q, and adenine riboswitches [29–32]. Some of these studies replaced the essential atmosphere of divalent Mg2+ ions with other, monovalent ions. Others used a repulsive Debye-Hückel interaction between the phosphate groups of the RNA. Such implicit treatments of the ions, however, neglect important near-field effects that occur inside the core of the riboswitch, where the ions are strongly coupled to the RNA.
Although the role of Mg2+ in stabilizing the RNA tertiary structure has long been realized [33,34], the molecular basis of ion-RNA interactions, in terms of structure and function, is not well understood. In a pioneering study, Draper and co-workers distinguished three classes of ion environments: (i) the diffuse ions, which are not restrained to any particular region, (ii) water-surrounded ions separated from the RNA by a single hydration layer, which we call outer-sphere ions, and (iii) chelated ions in the inner sphere, which form direct contacts with at least two different phosphate groups of the RNA [34,35]. While the potential role of chelated ions has been emphasized in many studies, recent work using explicit solvent molecular dynamics instead highlights a dense layer of outer-sphere Mg2+ ions, which are primarily responsible for anchoring the RNA structure [36]. These outer-sphere Mg2+ ions are only transiently bound but nonetheless strongly coupled to the RNA dynamics. Such highly correlated Mg2+ ions may even reside in the core region of riboswitch RNAs [36,37]. This dynamic cloud of Mg2+ has also been simulated in all-atom reduced models that combine Manning theory with a background of monovalent ions, represented by Debye-Hückel interactions [37,38]. In addition to these native basin simulations, studies of metabolite recognition and specificity have also been initiated using conformational ensemble sampling, again in the absence of Mg2+ ions [39].
While most riboswitch studies have focused on riboswitches in the 5’-UTR of mRNA that control transcription, less attention has been paid to translational control by riboswitches through ligand-dependent sequestration of the ribosome binding site (i.e., the Shine/Dalgarno sequence). The SAM-II riboswitch is one such relatively small RNA element which regulates methionine and SAM biosynthesis. A single hairpin, classic H-type pseudoknot and triplex interaction near the ligand binding site make this RNA an interesting system to study RNA control over the translation initiation process [21,40–44].
A previous single molecule fluorescence resonance energy transfer (smFRET) [21] imaging study sheds light on the dynamic nature of ligand-free SAM-II riboswitch, which becomes conformationally restrained upon ligand binding. The flexibility of such highly transient conformations is tuned to ensure a viable time scale for conformational transitions in the absence of ligand. More rapid sensing could, however, be achieved if the riboswitch adopted a binding competent conformation in the ligand-free state. Mg2+ ions can act as effective anchors, aiding in the preservation of the structural integrity of the RNA. The emergence of two distinct FRET configurations in the presence of 2 mM Mg2+ in the ligand free system suggests that Mg2+ has the ability to compress the RNA structure, in such a way that it might pre-organize the RNA to form a binding competent conformation. In a series of small angle X-ray scattering (SAXS) experiments, we observed an analogous signature of Mg2+ induced structural collapse that can facilitate subsequent ligand binding [40]. Our studies provided direct insight into the global rearrangement induced by both Mg2+ and ligand. The compaction of RNA by Mg2+ was also studied by size-exclusion chromatography (SEC): changes in the measured elution volume suggested a decrease in the particles’ hydrodynamic radius [40].
Pre-organization by Mg2+ has also been observed in other riboswitches [21,28,45]. In the SAM-I system, our biochemical studies have shown that addition of Mg2+ yields the pre-organized partially folded state. In addition, we have shown that, in the absence of Mg2+, the fully folded state cannot be achieved, even at high ligand concentrations [23,28,46]. The presence of both, Mg2+ and ligand are required for the stabilization of the fully folded ligand-bound configuration.
While previous smFRET and SAXS data revealed that ligand free RNA undergoes substantial structural changes upon variation of Mg2+ concentration, these structural changes often remained undetected by traditional NMR and X-ray crystallography techniques because of the transient nature and low population levels of such intermediates [47]. The newly developed Chemical Exchange Saturation Transfer (CEST) measurements are now capable of probing these sparsely and transiently populated RNA conformations. Earlier we studied NMR dynamics of the SAM-II system with this new method [47]. The data indeed confirmed that SAM-II riboswitch can access a sparsely populated but bound-like pre-organized state even in the absence of ligand [40,47].
In the present study, we performed molecular simulations to predict the effect of Mg2+ on the conformational landscape of the SAM-II riboswitch. We then tested these predictions with 13C-CEST data. Analysis of our simulations yields the free energy landscape of the SAM-II riboswitch, the effect of Mg2+ on this landscape and insight into the microscopic origins of these effects. More specifically, reappraisal of the 13C-CEST data for the ligand-free SAM-II riboswitch at different Mg2+ concentrations enabled us to probe the influence of Mg2+ on sparsely populated bound-like pre-organized states. We then revisited earlier smFRET, SAXS and SEC elution profiles and compared with our present equilibrium simulation results to integrate these data into a unified scenario of Mg2+-induced collapse. We calculated the free energy landscape of the SAM-II riboswitch using our recently developed all-atom structure-based model (SBM) that includes explicit Mg2+ ions and the effects of Manning condensation and Debye-Hückel Potassium and Chloride interactions. We specifically predict that, as Mg2+ concentration is increased from 0.25 mM to 2 mM, the SAM-II riboswitch collapses from an extended, partially unfolded state to a highly compact, pre-organized state, in agreement with the 13C-CEST studies, where we observe a shift in population towards a bound-like conformation. In addition, our simulations characterize this collapse transition in terms of the radius of gyration as a function of Mg2+ concentration, which is qualitatively similar to previous SAXS measurements. This agreement gives us confidence in the microscopic details of our simulations, showing that the triplex formation between helix P2b and loop L1 plays an important role in the collapse process.
Simulations of the SAM-II riboswitch were performed with and without SAM at different Mg2+ concentrations. The simulation started from the crystal structure of SAM-II riboswitch (pdb accession code: 2QWY) [20]. A global view of this structure and details of the secondary structure are presented in Fig 1. As ligand-free conformations of SAM-II are still under investigation in experiments, we explore the full folding free energy landscape both in the presence of ligand (SAM) and in the absence of ligand to characterize the entire accessible conformational space. To cover the folding landscape, we employed umbrella sampling over the reaction coordinate, Q, which is the fraction of intra-molecular native contacts, present in both the free and bound states, formed by the riboswitch.
As mentioned earlier, CEST experiments are able to capture transiently populated dynamic conformations [13]. This strategy was applied to the ligand-free SAM-II riboswitch in the presence of 0.25 mM and 2 mM Mg2+. The 13C-CEST profiles of the labeled ribose C1’ and base C6 carbons of C43 were recorded at three different B1-fields (17.5, 27.9, 37.8 Hz) with a mixing time of 0.3 s at 298 K [47]. We compare the data for 0.25 mM and 2 mM Mg2+ concentrations at B1-field of 17.5 Hz (Fig 2a). The data were fit with a two-state model where the low population of the partially closed state (peaks around 300 Hz spin-lock offset) appears to increase with addition of 2 mM Mg2+. Consistent results were obtained from CEST profiles for other B1-fields of 27.9 and 37.8 Hz (Fig S1 in the S1 Text). Trajectory plots of the fraction of native contacts extracted from the generalized Manning equilibrium simulations of ligand-free SAM-II at these two concentrations clearly show the hopping between different conformations (Fig 2b). Furthermore, the dynamic transitions between the two major states (bound-like: Q≈0.9 and open: Q≈0.7) visit native-like conformations (Fig 2c) more frequently at 2 mM than at 0.25 mM Mg2+, as summarized in the corresponding contact histograms, P(Q) (Fig 2d). Both CEST experiments and simulation data indicate that the equilibrium shifts from the open conformations toward the native bound-like state as we increase Mg2+ concentration.
The signature of the existence of such Mg2+ induced bound-like states has also been reported in previous smFRET experiments (Fig 2e) [22]. To support our observations we have revisited some of these smFRET efficiency assessments [22] and compared them with theoretical FRET predictions obtained from our generalized Manning model simulations under similar buffer conditions. The equation used for theoretical FRET prediction is described in section S1 in the S1 Text. We tracked the dynamics of positions 14 and 52, where acceptor (cy5) and donor (cy3) fluorophore labels were placed in the smFRET experiments (Fig 2f). Both experimental and simulation FRET confirm the coexistence of two states at 2 mM Mg2+ (Fig 2g) [22].
Previous SAXS data corroborates well the existing smFRET observations [22,40]. The SAXS data also indicated both ligand and Mg2+ ions are required to effectively fold this riboswitch. To microscopically understand their mutual and stand-alone effects from the present simulations, we studied the conformational differences of this riboswitch in four extreme buffer conditions and compared our computational results with experimental SAXS data. For this comparison, we extracted multiple snapshots from several long trajectories and computed ensemble averaged SAXS profiles using the Debye formula for spherical scatterers parameterized in the FoXS web server [48,49] as described in section S2 in the S1 Text. The predicted SAXS curves here show qualitative agreement with experiments (Fig 2h) [40]. The Kratky representation of SAXS data presented in Fig s2 in the S1 Text shows a pronounced peak, indicating the emergence of more extended conformations with decreasing Mg2+ concentrations. We note that capturing the entire conformational heterogeneity of an extended state is computationally challenging. This mostly applies for the extreme case where neither ligand nor Mg2+ is present. In this case, the correlation between theoretical and experimental SAXS profiles leaves room for improvement. Values for chi-square reflect that and are shown in Table S1 in the S1 Text. These analyses indeed suggest the potential impact of both, ligand and Mg2+, stabilizing the closed conformations, which we characterize further below with contact data to describe the pre-organization and the ligand-organized closing.
A significant Mg2+ induced collapse transition, as indicated by the SAXS data, has been followed over a wide concentration range (up to 100 mM) of Mg2+ in SEC elution volume profile (Fig 2i). Here RNA elutes after longer retention times with increasing Mg2+ concentration ([Mg2+]) in the mobile phase [40]. Bigger elution volume signifies decreasing hydrodynamic radius of a monomeric RNA molecule. The folding transitions, both from experimental elution volume data and from average Rg measured from the equilibrium simulation analysis as functions of [Mg2+], follow sigmoid curves with transition midpoint, Mg1/2 at 6 mM (Fig 2i).
At this point, a range of experimental techniques and simulation data support the existence of pre-organized states. Here we aim to obtain a thermodynamic description of how Mg2+ governs the energy landscape of RNA from our model simulation study. In Fig 3a, we show the free energy landscape for the folding transition of SAM-II riboswitch in SAM-bound (in the presence of explicit ligand) and SAM-free (in the absence of ligand) conditions near the physiological concentration of Mg2+ ([Mg2+] = 2.0 mM). During this folding transition, each secondary structural segment folds sequentially illuminating the pathway of folding (Fig 3b). The free energy profile, in the presence of explicit SAM has a distinct bound-state-well, reflecting the ligand-induced stabilization of the closed conformations (designated as (i) in Fig 3c). In the apo-form of the riboswitch, the fully closed bound state does not correspond to a minimum in the landscape. At lower Q than this ligand-bound state, the free energy profile for apo-SAM-II riboswitch reveals three distinct minima. They involve: a ligand-free partially closed state (state (ii) in Fig 3c), which has a substantial overlap with the ligand-bound closed conformation. In this state, the nonlocal contacts (involving base-pairing contacts) including base-stacking contacts in P1, the P1-L3 pseudo-knot interaction, and major segments of P2b and the L1-P2b triplex interactions remain secured, while the contacts involved in Shine/Dalgarno sequence (AAAG50G51A523´), and in the part of L1-P2b are disrupted (state (ii) in Fig 3c).
Recent fluorescence and NMR spectroscopic data also indicated that C16 in P2a helix remains mostly unpaired in the absence of SAM [22]. The data also suggested that formation of the pseudoknot in the absence of SAM is highly transient in nature. Intermediate states, (iii) and (iv) in Fig 3c, although marginally separated by a small barrier, effectively belong to a broad, flat basin which involves an ensemble of partially folded open configurations. A representative unfolded structure ((v) in Fig 3c) is shown to describe the unfolded minimum. As we increase the concentration of Mg2+ we find enhanced stabilization of the pre-organized partially closed conformations (state (ii) in Fig 3d) relative to the open conformations. Our latest 13C-CEST chemical exchange data anticipated that the emergence of Mg-induced pre-organization can have immense consequences for rapid ligand recognition [47]. In the context of the simulation, Mg2+ induced thermodynamic stabilization is reflected by the difference in stability, ΔGPC-PO between the bound-like partially closed (PC) conformation and the partially open (PO) conformation and by ΔGU-O between the unfolded (U) and the open conformation (O). These two stability differences, ΔGPC-PO and ΔGU-O vary upon increasing [Mg2+] until they reach their saturation limits. Both ΔGPC-PO and ΔGU-O plotted as functions of [Mg2+], are fitted well to sigmoid curves with a Mg2+1/2 value around 6 mM (inset of Fig 3d) which again correlates well with the SEC elution volume data (Fig 2i) [40].
To address the open question of how Mg2+ ions regulate structural collapse, we have determined the Mg2+ distribution in the ion-solvation layer of SAM-II, which accommodates increasing numbers of Mg2+ up to 8 mM Mg2+ content (Fig 4a). Subsequent additions of Mg2+ beyond 8 mM do not effectively add to the 1st layer of Mg2+ solvation. How we characterize the ion-solvation layer from our simulated trajectories is described in section S3 in the S1 Text (Fig s3 in S1 Text). We have further classified the outer sphere Mg2+ present in the ion-solvation layer into two categories based on their number of associated phosphate groups: (i) Single phosphate coordinating Mg2+ (Fig 4b), which efficiently neutralize the negative charge of the adjacent phosphate (Fig 4d), and (ii) multiple phosphate coordinating Mg2+ (Fig 4c). The key role in stabilizing the structure is played by such Mg2+ bridging multiple phosphates, which can act as glue in compact structures by holding a number of negatively charged phosphates together in close proximity.
The population shift coincident with multiple coordinated Mg2+ ions with increasing Mg2+ concentration directly supports their role in stabilizing the structure (Fig 4e). We have also investigated the thermodynamic impact of Mg2+-mediated phosphate contacts (PHOSCont: total number of pair-wise phosphate-phosphate contacts) on the energy landscape as a function of overall folding progress, expressed by the number of native contacts (NCont), as shown in Fig 4f–4h. As we increase the Mg2+ concentration the broad minimum that appeared around NCont~800, involving partially folded open conformations, gradually becomes more stabilized. Concurrent enrichment of phosphate-phosphate contacts extends the contour of the minimum asymmetrically toward higher PHOSCont. Additionally, by 8 mM [Mg2+], the bound-like pre-organized state grows with substantial population, stabilized again by phosphate connections (Fig 4h).
We analyzed long equilibrium trajectories of the apo- and bound-forms of SAM-II slightly below the folding temperature in order to capture the essential characteristics of the pre-organized state, and also to compare this state with the fully folded ligand bound state. We have evaluated the distribution of native contact formation in each segment of secondary structure as a function of the total Q at different Mg2+ concentrations. Plots show contact formation in P2b (Fig 5a–5d) and the triplex interaction between helix P2b and loop L1 (Fig 5e–5h), which are most affected by Mg2+ concentration. Data for the nonlocal contacts of P1, L3-P1, P2a, which appear only marginally affected by Mg2+ concentration, are shown in Fig s4 in S1 Text. The two distinct basins visible at low [Mg2+], for the P2b helix and L1-P2b triplex contacts correspond to the pre-organized (at higher Q) and open states (at lower Q). At increasing [Mg2+], the populations gradually shift towards the pre-organized state.
Around 8 mM Mg2+, the dominant contribution arising from this pre-organized triplex to the conformational space is evident from Fig 5c and 5g. Ligand binding also strongly favors structure formation, even at moderate [Mg2+], as the ligand bridges the gap between L1 strand and P2b helix, producing the fully formed triplex.
Motivated by our 13C-CEST profiles for SAM-II and their [Mg2+] dependence we have explored the free energy landscape of the SAM-II riboswitch using a recently developed all-atom SBM that includes explicit Mg2+ ions, Debye-Hückel treatment of implicit KCl interactions, and the effects of Manning condensation to accurately account for the ion atmosphere around the RNA. Our results support a mechanism involving Mg2+ induced pre-organization followed by conformational selection by the ligand, SAM, as we speculated in an early study [47]. The free energy analysis validates the observations of that pre-organization, providing an atomistic and thermodynamic basis for the enhanced population of a partially collapsed, pre-organized ensemble at sufficiently high Mg2+ concentration in the absence of ligand. We observe three distinct sets of conformations in the folding free energy landscape of ligand-free SAM-II riboswitch: (i) an ensemble of unfolded conformations, (ii) a broad ensemble of partially folded open conformations, and (iii) an ensemble of pre-organized bound-like conformations. As we increase magnesium concentration beyond 2–4 mM, the bound-like ensemble is further stabilized, shifting the equilibrium toward the pre-organized states. All the experimental results from our 13C-CEST profile, recent SAXS, single molecule FRET, and size-exclusion chromatographic studies are assembled and found to be in good agreement with the present simulation results (Fig 2). At higher concentrations, Mg2+ stabilizes compact structures by coordinating multiple charged phosphate groups of RNA in close proximity. The experimental results, together with free energy landscapes confirm that sufficient Mg2+ can indeed promote stable ligand binding in the SAM-II riboswitch, and is likely the structural basis for the switching control of protein translation.
While this structural pre-organization of SAM-II can assist in rapid ligand recognition, our study suggests that a sufficiently high concentration of Mg2+ is necessary to capture those pre-organized states. Only when the system achieves a well-organized ion solvation layer at high [Mg2+], the effect of additional Mg2+ seems limited. This layer involves a number of Mg2+ ions, each coordinating with multiple phosphate groups. Mg2+ ions thus serve as glue to the negatively charged phosphates and facilitate the structural compaction. We note that while chelated Mg2+ may play an important role in other riboswitch RNAs, no specific chelated ions have been reported so far in the SAM-II system.
Our molecular simulation trajectories also allow us to pinpoint the structural basis of the effect, revealing that triplex interaction between the helix P2b and its association with the L1 strand dominate the process of pre-organization as summarized in Fig 5a–5c and 5e–5g, showing the gain of structure with increasing [Mg2+]. In the final step, ligand binding firmly bridges the extended gap between L1 and P2b, which seems otherwise not achievable through the addition of small, dynamic Mg2+ alone. But although P2b and its connection with L1 can be secured by the ligand, its presence again alone cannot fully stabilize the overall structure without addition of significant amount of Mg2+ (Fig 2h). These findings suggest that a sufficiently high concentration of Mg2+ is necessary to stabilize the pre-organized triplex and then the presence ligand promotes the native triplex formation, as summarized in Fig 5d and 5h. We note that triplexes have recently emerged as important players in gene regulation by non-coding RNAs [50–53]. Base triples also play a role in RNase P and the Diels-Alder ribozyme [54]. Heroic calculations, as such recent microsecond explicit solvent simulations of riboswitches, will also shed light on these effects, especially regarding the role of solvation [55].
Nucleic acid-ion interactions make a substantial energetic contribution in the stabilization of the native state of RNAs, including complex formation with proteins and other macromolecules [56]. The dynamics of nucleic acids are also found to be strongly influenced by the motion of their ion atmospheres. Relative to other ionic species, Mg2+ can efficiently support a close assembly of negatively charged phosphates by mediating favorable interactions among them. Other earth alkali metals/divalent ions (e.g. Ca2+) and even monovalent ions are also able to induce similar transitions, albeit at higher concentration. Our early SEC elution profiles for SAM-II show that the transition midpoint in presence of Potassium (K+) alone occurs only at [K]1/2 ≈ 25 mM. The midpoint for Calcium (Ca2+) is [Ca]1/2 ≈ 8 mM, compared to 6 mM for the Mg2+ ion [40]. This is a direct result of the larger charge/radius ratio of magnesium [40,57].
Thus, having these special characteristics, Mg2+ efficiently helps pre-organize the system and enables access to the partially collapsed states that are further stabilized by ligand binding. The general importance of Mg2+ for the stability of compact RNA structures supports a possibly universal role of conformational selection in ligand-binding RNAs, such as riboswitches, aptamers, and possibly protein-binding RNAs. A detailed thermodynamic understanding of the underlying landscape will indeed enable greater control of riboswitch regulation, highly sought after by researchers in synthetic biology who are currently employing riboswitches as ligand-dependent ‘knobs’ to control desired gene expression [58].
Our all-atom structure-based model (SBM) has proven successful in describing the dynamics of numerous proteins and macromolecular complexes [59–63]. To elucidate RNA free energy landscapes under the influence of Mg2+, models capable of quantitatively describing the ion atmosphere are needed, including ionic condensation around the negatively charged phosphate groups of RNA. Early studies have simply included electrostatic effects in SBM of RNA via repulsive Debye-Hückel interactions, thus treating all ions implicitly [29,30]. Recently, our group developed a more detailed model of RNA electrostatics and applied it within all-atom structure-based molecular dynamics simulations. Our model treats Mg2+ ions explicitly to account for ion-ion correlations neglected by mean-field theories [38]. The KCl buffer, which completes the experimental setup, is treated implicitly by a generalized Manning counter ion condensation model [38,64], since mean-field theories correctly assess the charge densities of monovalent K+ and Cl- ions. Classical Manning counter-ion condensation theory was originally developed for understanding the low concentration limiting behavior of polyelectrolyte chains modeled with an infinite line of charge. Folded RNA, however, is not a line of charge. To account for the compact and irregular structures of RNAs and the effects of varying ion concentrations, we improve the Manning counter ion condensation model to handle electrostatic heterogeneity, making the condensed charge density a dynamical function of each phosphate coordinate. KCl screening is characterized by a Debye-Hückel potential. Removal of the continuum screening ions from the inaccessible volume of RNA is a substantial extension to Manning counter-ion condensation. The model has been tested against experimental measurements of excess Mg2+ associated with RNA, characterizing the Mg2+-RNA interaction free energy. This hybrid SBM has opened up new possibilities to study various structural and functional processes of RNA that are essentially controlled by ions [38]. In the present study we used this recently developed all-atom hybrid SBM to understand the conformational transition of SAM-II and the corresponding Mg2+ sensitivity. The energy function used in this model is given below,
Φ=ΦSBM+ΦMg-Size+Φion-effect
(1)
where, ΦSBM is the all-atom SBM potential ensuring a global minimum in the landscape for the native state of RNA. The SBM potential is composed of two general types of interactions:
ΦSBM=Φlocal+Φnon-local
(2)
where, Φlocal characterizes the local interactions that encode covalent bonds and torsional angles, maintaining the correct local geometry and chirality. Φnon-local comprises two non-local contributions: (i) an attractive term that is applied specifically to all tertiary interactions determined from the native structure, (ii) the general repulsive interactions, that describe the excluded volume by symmetric hard potentials (to avoid any unwanted chain crossing). ΦMg-Size adds the excluded volume interactions involving the explicit Mg2+ ions, regulating RNA-Mg2+ and Mg2+-Mg2+ interactions. Φion-effect accounts for all interactions between charges in the system which consist of the fixed charge distribution of the RNA and the dynamic contribution from the ions. Mg2+ and phosphate charges interact via a Debye-Hückel potential with a screening term that depends, in turn, on the distribution of the monovalent ions. The monovalent ions, K+ and Cl- from the added salt, fall into two categories: screening ions and Manning condensed ions. The screening ion density is obtained using Debye-Hückel electrostatics. The density of the Manning-condensed ions is modeled as the sum of two normalized Gaussian distributions where the center of each Gaussian is located on the position of the negatively charged phosphate group. All the condensation variables along with the explicit Mg2+ and RNA coordinates are evolved with Langevin dynamics [38]. The mathematical formulations of all the terms and the related parameterizations are discussed in depth in section S4 in the S1 Text.
The umbrella sampling method [65] was used to sample the conformational space of SAM-II riboswitch along the reaction coordinate, Q, which is the fraction of intra-molecular native contacts in the riboswitch. The Weighted Histogram Analysis Method [66] was then used to calculate the thermodynamic quantity, G(Q). The detail is described in section S5 in the S1 Text.
CEST data were collected using a pseudo-3D HSQC experiment with the B1 field offsets (-600 to 600 Hz) incremented in an interleaved manner with 3 references (no CEST period) [47]. A total of 1024x16 complex points were recorded [40] with 32 transients with a recovery delay of 1.5 s for a total experimental time of approximately 12 hr for each spin-lock field. A CEST saturation period of 100 ms was used for the base and 200 ms for ribose. The pulse program used was an adaptation of a previously published one without the need for selective pulses [47].
We used a two-state model to fit each of the three profiles of the selectively labeled carbon (ribose C1’ and base C6) and quantitatively extracted the carbon chemical shift (Δω), the exchange rate, and the population of the minor state based on the Bloch−McConnell 7x7 matrix [47]. The CEST data was plotted as I(t)/I(0) versus spin-lock offset (Hz) and was fit by numerically solving the matrix exponential for the CEST spin-lock period based on this 7x7 two-state Bloch-McConnell equation as described earlier [47,67].
In experiments, Fluorescence Resonance Energy Transfer (FRET) efficiency is the quantum yield of the energy transfer where a donor chromophore from its excited electronic state may transfer its energy to an acceptor chromophore through a non-radiative dipole-dipole coupling. The FRET efficiency varies with the separation between donor and acceptor fluorophores following the Fӧrster relation. For theoretical FRET predictions we use the Fӧrster relation where the value of Fӧrster radius is taken as 53Å [68]. We described it in detail in section S1 in the S1 Text.
In SAXS experiments, the scattering intensity is measured from the electron density difference between the purified sample and that of the solvent/buffer. FoXs is a method that uses the Debye formula by which a theoretical scattering profile of a structure can be computed [48,49]. The detail is discussed in section S2 in the S1 Text.
|
10.1371/journal.pmed.1002032 | Prices, Costs, and Affordability of New Medicines for Hepatitis C in 30 Countries: An Economic Analysis | New hepatitis C virus (HCV) medicines have markedly improved treatment efficacy and regimen tolerability. However, their high prices have limited access, prompting wide debate about fair and affordable prices. This study systematically compared the price and affordability of sofosbuvir and ledipasvir/sofosbuvir across 30 countries to assess affordability to health systems and patients.
Published 2015 ex-factory prices for a 12-wk course of treatment were provided by the Pharma Price Information (PPI) service of the Austrian public health institute Gesundheit Österreich GmbH or were obtained from national government or drug reimbursement authorities and recent press releases, where necessary. Prices in Organisation for Economic Co-operation and Development (OECD) member countries and select low- and middle-income countries were converted to US dollars using period average exchange rates and were adjusted for purchasing power parity (PPP). We analysed prices compared to national economic performance and estimated market size and the cost of these drugs in terms of countries’ annual total pharmaceutical expenditure (TPE) and in terms of the duration of time an individual would need to work to pay for treatment out of pocket. Patient affordability was calculated using 2014 OECD average annual wages, supplemented with International Labour Organization median wage data where necessary. All data were compiled between 17 July 2015 and 25 January 2016. For the base case analysis, we assumed a 23% rebate/discount on the published price in all countries, except for countries with special pricing arrangements or generic licensing agreements.
The median nominal ex-factory price of a 12-wk course of sofosbuvir across 26 OECD countries was US$42,017, ranging from US$37,729 in Japan to US$64,680 in the US. Central and Eastern European countries had higher PPP-adjusted prices than other countries: prices of sofosbuvir in Poland and Turkey (PPP$101,063 and PPP$70,331) and of ledipasvir/sofosbuvir in Poland (PPP$118,754) were at least 1.09 and 1.63 times higher, respectively than in the US (PPP$64,680 and PPP$72,765). Based on PPP-adjusted TPE and without the cost of ribavirin and other treatment costs, treating the entire HCV viraemic population with these regimens at the PPP-adjusted prices with a 23% price reduction would amount to at least one-tenth of current TPE across the countries included in this study, ranging from 10.5% of TPE in the Netherlands to 190.5% of TPE in Poland. In 12 countries, the price of a course of sofosbuvir without other costs was equivalent to 1 y or more of the average annual wage of individuals, ranging from 0.21 y in Egypt to 5.28 y in Turkey. This analysis relies on the accuracy of price information and infection prevalence estimates. It does not include the costs of diagnostic testing, supplementary treatments, treatment for patients with reinfection or cirrhosis, or associated health service costs.
Current prices of these medicines are variable and unaffordable globally. These prices threaten the sustainability of health systems in many countries and prevent large-scale provision of treatment. Stakeholders should implement a fairer pricing framework to deliver lower prices that take account of affordability. Without lower prices, countries are unlikely to be able to increase investment to minimise the burden of hepatitis C.
| New medicines for hepatitis C are very effective but also very expensive, prompting debates about pricing and affordability and limited access.
Our study is a global analysis to compare public information about the prices of sofosbuvir and ledipasvir/sofosbuvir, taking account of probable confidential discounts on prices, in order to calculate the potential total cost of these medicines for different national health systems and individual patients in 30 countries.
We obtained 2015 prices for a 12-week course of treatment with sofosbuvir and ledipasvir/sofosbuvir for as many countries as possible. We used three sources: the Pharma Price Information service of the Austrian public health institute Gesundheit Österreich GmbH, national government and drug reimbursement authority websites, and press releases.
We estimated how many patients in each country are infected with hepatitis C, based on existing studies.
We analysed and compared the medicine prices, adjusting for currency differences, the confidential price discount that might be negotiated by purchasers in each country and the national wealth of countries. We then calculated the likely total cost to each country of treating all of their patients infected with hepatitis C. We compared this to the annual total expenditure on medicines for each country, and we also calculated how long a person would need to work in each country to pay for treatment out of pocket, based on each country’s average wage.
We found the following:
The prices of the medicines for hepatitis C vary significantly across countries, particularly when adjusted for national wealth. Poorer countries may be paying higher adjusted prices than richer countries.
The total cost of treating all patients with hepatitis C would be equal to at least a tenth of the current annual cost for all medicines in all of the countries studied. In some countries where prices are high and the burden of disease is large, the total cost of treating all infected patients would be more than the cost of all other medicines put together.
If a patient had to pay for the treatment out of pocket, the total cost of a full course of sofosbuvir alone would be equivalent to one year or more of average earnings for individuals in 12 of the 30 countries analysed.
The prices of the medicines for hepatitis C vary significantly across countries, particularly when adjusted for national wealth. Poorer countries may be paying higher adjusted prices than richer countries.
The total cost of treating all patients with hepatitis C would be equal to at least a tenth of the current annual cost for all medicines in all of the countries studied. In some countries where prices are high and the burden of disease is large, the total cost of treating all infected patients would be more than the cost of all other medicines put together.
If a patient had to pay for the treatment out of pocket, the total cost of a full course of sofosbuvir alone would be equivalent to one year or more of average earnings for individuals in 12 of the 30 countries analysed.
Paying for sofosbuvir and ledipasvir/sofosbuvir in national health systems would consume large proportions of their total pharmaceutical budget. The potential total cost of treatment presents a financial and ethical dilemma for payers and physicians. Some national health systems have therefore restricted access to these medicines to small groups of patients, despite the fact that almost all patients with chronic hepatitis C infection are likely to benefit from treatment with these medicines.
Our analysis is limited by the accuracy of the price information that was accessible, which may be inaccurate because of confidentiality agreements between manufacturers and purchasers. The estimates of the numbers of people infected with hepatitis C are also uncertain. Also, we have considered only the costs of the medicines themselves and not the costs of other parts of the treatment for hepatitis C, so we have most likely underestimated the true total cost.
If countries are to try to pay for treating all patients with hepatitis C, governments and industry stakeholders will need to jointly develop and implement fairer pricing frameworks that lead to lower and more affordable prices.
| Hepatitis C virus (HCV) infection is a significant global public health problem. Although the precise prevalence is uncertain, a recent analysis estimated that 80 million people globally were living with HCV viraemia (also known as chronic HCV infection), with a range of 64–103 million people [1]. If left untreated, chronic HCV infection can cause liver cirrhosis and cancer, leading to an estimated 700,000 deaths per year worldwide [2].
Until recently, standard guidelines for treating HCV infection recommended combination therapy with pegylated interferon and ribavirin. While resulting in sustained virological response (SVR) in 54% to 63% of clinical trial participants [3], this combination requires 24 to 48 wk of therapy and has severe side effects such as haemolytic anaemia and flu-like symptoms [4]. Recently developed direct-acting antivirals (DAAs) have markedly improved treatment efficacy and shortened and simplified the treatment regimen. In late 2013, sofosbuvir was approved in the United States (US) as the first once-daily orally administered therapy for HCV without interferon. Clinical trials showed SVR in up to 93% of trial participants following 12 wk of treatment with sofosbuvir [5]. In 2014, a combination product of ledipasvir/sofosbuvir was approved in the US based on evidence of a SVR rate of up to 99% after 12 wk of treatment [6]. Since then, other DAA regimens have gained regulatory approval, but sofosbuvir and ledipasvir/sofosbuvir dominate the market [7].
The initial published list prices in the US for 12-wk courses of treatment with sofosbuvir and ledipasvir/sofosbuvir were US$84,000 and US$94,500, respectively. The manufacturer claimed the new regimen was equal to or less expensive than prior standard of care regimens because the new regimen had much higher cure rates and would reduce the total treatment costs of HCV, including the costs of medications, side effects, complications, and additional health services required [8,9]. A number of published economic studies from high-income countries supported the cost-effectiveness of sofosbuvir at the proposed price, although the budget impact was substantial [10,11]. A recent study estimated the cost of production of sofosbuvir to be US$68–US$136 for a 12-wk course of treatment based on the same manufacturing methods used in the large-scale generic production of HIV/AIDS medicines [12], and its findings have not been challenged. The difference between the estimated cost of production and the marketed prices raises questions about the fairness of pricing medicines of public health importance like sofosbuvir and ledipasvir/sofosbuvir [13], echoing similar discussion about new cancer medicines [14–16].
As a consequence of the high prices of the new HCV medicines, payers in high-income countries have been restricting coverage (e.g., United States) [17], negotiating public deals and private discounts and rebates with the manufacturer (e.g., France and Germany [18]), or delaying reimbursement until a reasonable price has been negotiated (e.g., Australia [19]).
A recent report suggested that 73% of people with chronic HCV live in middle-income countries [20]. For some low- and medium-income countries (LMICs), tiered pricing agreements have been negotiated. As a result, countries such as Mongolia, Egypt, and Pakistan have published prices of about US$900 per 12-wk course of sofosbuvir. Voluntary licensing agreements have been established with 11 India-based generic pharmaceutical manufacturers for production and distribution of the medicines in 101 countries [21]. The Indian licensees are selling generic sofosbuvir at prices between US$161 and US$312 per 28-tablet pack [22]. However, the license agreement does not include 39 middle-income countries, including Brazil, China, Mexico, and Turkey [21,23].
This study seeks to systematically compare the prices of sofosbuvir and ledipasvir/sofosbuvir across countries, including generic versions where prices are available. We assessed the affordability and budgetary impact of these treatments, both to health systems and to individual patients paying for the treatment fully out of pocket in the absence of reimbursement from public or private health insurance. We did not include other DAAs (e.g., daclatasvir, simeprevir, and combination products by other manufacturers) in this analysis because they are still only available in a small number of countries.
We conducted a comparative study of the published prices of sofosbuvir and ledipasvir/sofosbuvir in countries where published price information was available. Ex-factory prices were used because prices paid by the consumer include different amounts of taxes, mark-ups, and distribution costs that make them difficult to compare. We included Organisation for Economic Co-operation and Development (OECD) member countries and LMICs for which we had access to reliable, publicly available price information. If price information was unavailable for a country, the country was not included in the analysis. In the case of some European countries, this was usually because the medicines had not yet been marketed, or prices were not publicly disclosed. In India, the ex-factory price was “commercial in confidence”, and we used the publicly listed retail price without further adjustments.
In order to compare prices between countries, we converted prices in national currencies to US dollars using OECD 2014 period average exchange rates [24], without and with adjustment for purchasing power parity (PPP) [24]. We report these as the “nominal price” (US dollars) and the “PPP-adjusted price” (PPP dollars), respectively. PPP adjustment is important when comparing the prices of goods across countries to account for the differences in national income levels and purchasing power in buying goods and services [25,26]. The theory of PPP, based on the “law of one price”, states that, after accounting for transaction costs and trade barriers, identical goods will be sold for the same price in trading countries when their prices are expressed in a common currency [27,28]. Accordingly, the price of a medicine in different countries should in theory be the same when expressed in a common currency (e.g., US dollars) using the PPP exchange rates. S1 Text lists the prices and rates used in the analysis where prices in national currency were divided by the corresponding period average or PPP exchange rate of the national currency per US dollar published by the OECD.
We ranked the nominal and PPP-adjusted prices for all countries. We compared the PPP price for each country according to gross domestic product (GDP) per capita published by OECD [29] and, for Egypt and Mongolia, published by the World Bank [30].
We then estimated the total budget impact of sofosbuvir and ledipasvir/sofosbuvir, allowing for different levels of treatment coverage. To estimate the population requiring HCV treatment, we used the prevalence estimates of viraemic HCV infection by Gower and colleagues [1] rather than the number of persons with anti-HCV antibodies [31] because approximately 25% of persons who acquire HCV clear the infection spontaneously [32]. Based on the WHO treatment guidelines [33], we assumed in our base case analysis that all adults with viraemic HCV infection were eligible for treatment, and patients would receive a 12-wk treatment regimen of sofosbuvir or ledipasvir/sofosbuvir because a majority of patients have non–genotype 3 HCV. Patients with genotype 3 infections were assumed to receive 24 wk of treatment with sofosbuvir in sensitivity analysis.
We compared the total estimated budget impact of sofosbuvir and ledipasvir/sofosbuvir treatment with the country’s total expenditure on pharmaceuticals, based on the reported PPP-adjusted total pharmaceutical expenditure (TPE) per capita [34] multiplied by the population size [35]. For consistency in adjustments, we used the PPP-adjusted prices instead of the nominal prices of medicines when estimating budget impact. For the countries with negotiated tiered prices, we assumed no further rebate.
To assess the affordability of sofosbuvir and ledipasvir/sofosbuvir for individual patients without insurance coverage, we estimated the duration of time that an individual would need to work—earning the average wage of the general population—to obtain sufficient income to pay for a full course of treatment fully out of pocket. We performed a sensitivity analysis using minimum wage because people with HCV infection may have incomes below population average wages due to illness and socioeconomic status [36,37]. Previous studies [38–40] have measured the affordability of medicines using minimum wage as an indicator of patient income.
For the European Union (EU) member states, Norway, and Switzerland, published ex-factory prices for a 12-wk course of treatment were provided by the Pharma Price Information (PPI) service of the Austrian public health institute Gesundheit Österreich GmbH. The PPI service offers medicine price data for all price types (ex-factory price, pharmacy purchasing price, and net and gross pharmacy retail price) for all 28 EU member states, Norway, and Switzerland based on data collection from official national databases. The PPI service provided data for sofosbuvir as of July 2015, and for ledipasvir/sofosbuvir as of September 2015. To validate the collected data and to complete the dataset in other countries, we sought additional information from national government or drug reimbursement authorities and recent press releases (S1 Text).
Insurance agencies and reimbursement organisations often negotiate purchase prices lower than drugs’ published prices (list prices). These arrangements may take the form of confidential discounts, rebates, or refunds after purchase [41]. To account for the impact of these undisclosed agreements on prices, in our base case analysis, we assumed a 23% price reduction from the published price in all countries, except for the countries with a negotiated tiered price as noted above. This price reduction estimate was based on the legislated rebate obtained by the US Centers for Medicare & Medicaid Services [17]. To ensure the appropriateness of the estimated budget impact of sofosbuvir and ledipasvir/sofosbuvir, we conducted two sensitivity analyses assuming (1) no price reduction and (2) a 50% price reduction based on the rebate received by the US Department of Veterans Affairs [42].
Relevant demographic and economic statistics for each country were extracted from the OECD database: population [35], pharmaceutical expenditure per capita [34], GDP per capita [29], currency exchange rates [24], and average annual wages in 2014 PPP dollars (constant price) per full-time- and full-year-equivalent employee [43]. The OECD dataset contained wage information for all countries of interest except for Brazil, Egypt, Iceland, and Turkey. For these countries, we used the median nominal monthly earnings (converted to annual earnings) of employees reported in the United Nations International Labour Organization Global Wage Database [44], with linear extrapolation of historical data to 2014 values and with PPP adjustments.
All descriptive statistical analyses and data visualisation were performed in Microsoft Excel 2010 (version 14.0).
Table 1 lists the main assumptions used in the analysis and their rationale, data sources, estimates of uncertainty, and likely impact on the outcomes of the analysis.
As noted above, there are confidential agreements between manufacturers and insurance agencies/reimbursement organisations, leading to the published prices being higher than the actual prices. We applied a 23% price reduction to estimate the impact of this uncertainty and undertook a sensitivity analysis using a 50% price reduction (S2 Text). The levels of supply chain remuneration and taxes also differ between countries, affecting the pharmacy retail prices. To ensure a conservative approach and to exclude possible bias due to different levels of mark-up and remuneration, we used ex-factory prices for all analyses. This underestimates the total expenditure and costs to the patient.
We estimated TPE based on the mean and range of values of prevalence of hepatitis C viraemia presented in Gower et al. [1].
We also carried out an analysis to adjust the total expenditure on sofosbuvir and ledipasvir/sofosbuvir for the proportion of patients in each country infected with genotype 3 HCV, based on the genotype distribution estimated by Gower et al. [1] and assuming that this group of patients would receive the recommended 24-wk treatment with sofosbuvir (S3 Text). Other factors influencing the duration of treatment are primarily the presence or absence of cirrhosis [33], but reliable data for this factor are not yet available from all countries in our analysis. However, adjusting for this factor would increase the estimates for total expenditure on sofosbuvir and ledipasvir/sofosbuvir because the recommended duration of treatment for patients with cirrhosis is longer (24 wk).
There are two sources of uncertainty when estimating the duration of full-time paid employment required for a patient to pay for a full course of treatment: the income level of HCV patients and the level of out-of-pocket payment. We used the published average or median wage in different countries as the indicator of patient income, assuming that people with HCV infection have the same income as the general population. This is a conservative assumption because some HCV patients are intravenous drug users who typically have low incomes, low employment, and low levels of education [36,37]. To test this assumption, and also for comparability with previous studies on medicine affordability [38–40], we performed a sensitivity analysis using the minimum wage data reported in Country Reports on Human Rights Practices for 2014 [45].
For this analysis of the patient affordability sofosbuvir and ledipasvir/sofosbuvir, we assumed that patients would pay for the treatment fully out of pocket because the level of coverage and co-payments varies considerably across health systems, types of insurance and benefits packages, and characteristics of HCV patients. It could be expected that in high-income countries with universal health coverage, such as European OECD countries, prices of HCV medicines would be fully covered by public payers if the eligibility criteria for reimbursement were met. However, while information from the PPI service showed that some European high-income countries provided full coverage, national price data sources indicated no coverage in other countries. This suggests either full out-of-pocket payments or some specific coverage arrangements beyond regular reimbursement.
We did not include in this analysis the prices of other medicines used in various HCV treatment regimens, the cost of diagnostic tests, and other health service costs. This, again, is a conservative assumption.
Prices of sofosbuvir were obtained for 26 OECD countries and 4 LMICs: Brazil, India, Egypt, and Mongolia (Fig 1A). Ledipasvir/sofosbuvir prices were available in 21 OECD countries, India, Egypt, and Mongolia (Fig 1B). The prices in India were the retail prices of generic products.
Assuming a 23% price reduction on the list ex-factory price, the median nominal price of sofosbuvir for a 12-wk course across all OECD countries was US$42,017, with the price ranging from US$37,729 in Japan to US$64,680 in the US. The 23% reduction was not applied in Brazil, Egypt, India, and Mongolia, where there were special pricing arrangements or generic licensing agreements. The nominal prices of sofosbuvir in Brazil (US$6,875), Egypt (US$932), Mongolia (US$900), and India (US$539) were up to 120 times lower than the nominal price in the US (US$64,680) due to negotiated pricing arrangements and licensing agreements.
In countries with stronger purchasing power, PPP adjustment resulted in lower prices. For example, as illustrated in Fig 1A, the PPP-adjusted price in Norway was 0.67 times less than its nominal price (US$42,148, PPP$28,092). In contrast, countries with weaker purchasing power had a significant increase in price with PPP adjustment. For example, the PPP-adjusted price of sofosbuvir in India (PPP$1,861) was 3.45 times more than the nominal price (US$539). In the OECD, the PPP-adjusted prices were higher in Poland (PPP$101,063), Turkey (PPP$70,331), Slovakia (PPP$63,815), Portugal (PPP$57,384) than in higher-income European economies, particularly the Nordic countries (i.e., Norway, Sweden, Denmark, and Finland). The PPP-adjusted prices for sofosbuvir in Brazil, Egypt, Mongolia, and India were PPP$9,708, PPP$3,117, PPP$2,604, and PPP$1,861, compared to the nominal prices of US$6,875, US$932, US$900, and US$539, respectively.
The nominal price of ledipasvir/sofosbuvir was the highest in the US (US$72,765 with 23% price reduction) and lowest in the UK (US$43,215) among the OECD countries. Assuming no price reduction, the nominal price of ledipasvir/sofosbuvir in Egypt and Mongolia (US$1,200) was 47 times lower than in the United Kingdom (US$56,123) due to negotiated tiered pricing. The nominal price of ledipasvir/sofosbuvir in India (US$655) was 1.2% of the price in the United Kingdom. The price in Norway was the lowest after adjustment for PPP (PPP$31,255), reflecting the stronger purchasing power of its currency. The PPP-adjusted price of a 12-wk course of ledipasvir/sofosbuvir was again higher in Poland (PPP$118,754) than in other countries, with the price being 3.8 times higher than the price in Norway (PPP$31,255).
Fig 2 shows the relationship between PPP-adjusted price, GDP per capita, and estimated market size. Prices do not increase, and in some cases decrease, with increased standard of living. This pattern is not seen in Brazil, Egypt, India, and Mongolia, however, where prices of sofosbuvir were low because of existing pricing arrangements. In the US, prices of both products were much higher than in countries with comparable GDP per capita.
There was no observable relationship between the PPP-adjusted price and potential market size (Fig 2). For example, Nordic countries had fewer people requiring HCV treatment (based on point estimates reported by Gower et al. [1]) and had higher GDP per capita than countries such as Japan, Italy, and Spain, but the PPP-adjusted prices of sofosbuvir and ledipasvir/sofosbuvir were much lower. Although the GDP per capita in Turkey was only about $3,000 higher than in Brazil, the PPP-adjusted price of sofosbuvir in Turkey (PPP$70,331) was 7.2-fold higher than in Brazil (PPP$9,708). Luxembourg had the smallest estimated HCV population and the highest GDP per capita, but its price for sofosbuvir was less than the price in half of the other OECD countries.
Fig 3 and Table 2 show the budget impact of treating all infected patients with sofosbuvir or ledipasvir/sofosbuvir for a 12-wk course of treatment, at the PPP-adjusted prices with a 23% reduction, based on the point estimates and range estimates for HCV prevalence from Gower et al. [1]. The budget impact estimates vary from PPP$100.9 million (UI: PPP$56.7 million, PPP$174.1 million) in Luxembourg to PPP$166.6 billion (UI: PPP$153.7 billion, PPP$307.5 billion) in the US. As noted, these estimates do not include the costs of diagnostic testing, ribavirin or other medicines, or other associated health service costs. Fifteen of the 30 countries analysed would require more than PPP$5 billion to provide treatment coverage for the entire infected patient population of their country (Fig 3A). For Poland, Turkey, Spain, and Italy, the point estimates of total budget impact vary from PPP$20 billion to PPP$35 billion. For Japan, the point estimate of the total budget impact of providing treatment for 1.25 million people with HCV viraemia is close to PPP$50 billion. The PPP-adjusted expenditure for treating 2.58 million people with HCV viraemia (point estimate) in the US would be PPP$166.6 billion. These estimates do not include the cost of retreatment for patients who fail treatment or become reinfected. A similar level of financial impact is observed across countries for ledipasvir/sofosbuvir (Fig 3B).
Compared to the PPP-adjusted TPE in each country, treating the entire HCV viraemic population (based on point estimates reported by Gower et al. [1]) with sofosbuvir at the PPP-adjusted price with a 23% price reduction would amount to at least one-tenth of the current TPE in all countries (Table 3). In Poland, treating the whole HCV viraemic population would amount to as much as 1.6 times the current TPE. If only 10% of the HCV viraemic population were treated, the expenditure on sofosbuvir would still be high in proportion to the TPE in Poland (16.2%; UI: 11.1%, 40.6%), New Zealand (15.5%; UI: 8.9%, 26.8%), Portugal (13.3%; UI: 8.7%, 23.0%), Italy (11.1%; UI: 7.5%, 21.4%), and Spain (10.0%; UI: 5.4%, 16.7%) because these countries have relatively high sofosbuvir prices and HCV prevalence. In contrast, in the Netherlands, it would cost an amount equal to 1% (UI: 0.4%, 1.8%) of its current TPE to provide sofosbuvir to 10% of its infected population—even though the price of sofosbuvir is similar to in other OECD countries—because the estimated number of people with hepatitis C is relatively low. The analyses of prices of ledipasvir/sofosbuvir produced similar results, where the expenditure on ledipasvir/sofosbuvir would cost an amount equal to a considerable proportion of the country’s current TPE, particularly in Poland (19.1%; UI: 13.0%, 45.1%), Portugal (16.5%; UI: 9.9%, 28.5%), New Zealand (15.1%; UI: 9.3%, 26.1%), Italy (12.3%; UI: 8.5%, 25.2%), and Spain (11.2%; UI: 5.4%, 17.0%) (Table 3).
The PPP-adjusted price of a full course of sofosbuvir alone would be equivalent to at least 1 y (365 d) of the PPP-adjusted average earnings for individuals in 12 of the 30 countries analysed (Fig 4A). In Poland, Slovakia, Portugal, and Turkey, a course of sofosbuvir alone would cost at least 2 y of average annual wages. This analysis is conservative because prices were ex-factory prices with an assumed 23% price reduction, and did not include supply chain mark-ups and other costs such as the cost of diagnosis, daclatasvir, ribavirin, and health service costs. We also assumed that all wages were disposable for purchasing these medicines. Assuming no price reduction, a HCV patient in Poland would have to spend 5.55 y and 6.52 y of earnings on a 12-wk course of treatment with sofosbuvir and ledipasvir/sofosbuvir, respectively. Similarly, in ten of the 21 countries where annual wage data and ledipasvir/sofosbuvir prices were available, a person who earned an average wage would need at least 1 y of income to afford a course of ledipasvir/sofosbuvir if no subsidy were offered (Fig 4B). Due to the availability of tiered prices for ledipasvir/sofosbuvir, less than 1 y’s median wage earnings are required in Egypt to pay for treatment. The duration of time that an individual would need to work to pay for a course of treatment out of pocket would be higher with less conservative assumptions.
S2 Text lists the findings of univariate sensitivity analyses assuming 0% and 50% rebates on the prices of sofosbuvir and ledipasvir/sofosbuvir. Based on these assumptions, the estimated percentages of current TPE that would be required for countries to provide treatment for different percentages of the HCV-infected population were within the corresponding range of the values presented in Table 3. The estimated number of years individuals would need to work to pay for a full course of treatment using the minimum wage is an average of 1.55-fold (SA range: 0.8; 2.7) longer compared to the base case findings using average/median wage.
S3 Text shows the findings assuming that patients with genotype 3 HCV would receive the recommended 24 wk of treatment with sofosbuvir. Under this scenario, the estimated percentages of current TPE at different levels of treatment coverage and the number of equivalent income years required to pay for full treatment increased by an average of 1.26-fold (SA range: 1.00; 1.54) compared to the base case estimates.
Our analyses show significant price variation for sofosbuvir and ledipasvir/sofosbuvir across countries, especially when accounting for local purchasing power. The lowest and highest nominal prices of sofosbuvir and ledipasvir/sofosbuvir in OECD countries varied, respectively, by 1.71 times between Japan and the US and 1.68 times between the United Kingdom and the US. If the prices in LMICs under tiered pricing arrangements or licensing agreements are included, the prices vary by more than 100-fold. Our analysis also shows that the PPP-adjusted prices of these medicines in Central and Eastern European countries are considerably and consistently higher than in other OECD countries, particularly compared to Nordic countries. Countries that benefit from tiered pricing arrangements and are included in licensing agreements (such as Egypt, Mongolia, and India) have lower prices, which are more affordable compared to their average/median wages. Assuming minimal transaction costs and trade barriers, these are substantial price disparities for goods that are identical in all markets.
We are aware of four recent publications that have also analysed the budget impact and price variation of new HCV medicines [22,46–48]. The studies of budget impact [46–48] were specific to the US and Ireland only, whereas our analysis provides budget impact estimates for a significantly larger sample of countries. Our estimates of budget impact for the US are comparable to those in these studies, given the difference in populations used in each study. Médecins Sans Frontières (MSF) undertook a survey on the prices of six DAAs, including sofosbuvir and ledipasvir/sofosbuvir [22], although our study differs from the MSF study in a number of ways. Our analysis includes a larger range of high-income countries, while the MSF study included a larger range of LMICs. The MSF study obtained price data from key informants, while we obtained our price information from publicly accessible sources. We also obtained data in local currency and then adjusted for exchange rate, purchasing power, and potential price reductions using consistent methods. However, the MSF study found similar variability in prices among high-income countries, “with little correlation between drug prices and gross national income” [22]. Additionally, price disparities in relation to national wealth indicators also occurred among LMICs. The prices of sofosbuvir and ledipasvir/sofosbuvir were higher in Malaysia than in some high-income countries, and the originator version of sofosbuvir in India was less expensive than generic sofosbuvir in Côte d’Ivoire. The MSF study also found that the availability of most DAAs in low-income countries was low at the time of the survey, with the exception of sofosbuvir. For countries included in both studies, foreign exchange rates, time of data collection, and differences in sources are the primary reasons for the differences in prices. Notwithstanding these differences, we concur with the MSF authors in noting that having access to updated and reliable information on prices is essential to allow decision-makers to negotiate prices.
The price disparities in OECD countries may be explained partly by the pharmaceutical price-setting policies used in different countries. For example, some countries set prices according to explicit cost-effectiveness thresholds based on GDP per capita [49], average monthly wage [50], or comparative assessment [51] against interferon-based therapy. These thresholds signal the purchasers’ willingness to pay and may result in the highest possible price that satisfies the threshold, without consideration of budget impact. Another policy for price setting used by many OECD countries is external reference pricing. This is the practice of using the prices of a medicine in one or several countries in order to derive a benchmark, or reference, price for the purpose of setting or negotiating the price. However, the methods used may not explicitly incorporate local purchasing power and status of economic development. For example, the reference countries used by Turkey are France, Greece, Portugal, and Spain, which are selected on the basis of similar “product variety, communicable and common diseases, population and age distribution, and health/disease status” [52,53]. The nominal price of sofosbuvir in Turkey, assuming a 23% price reduction, was lower (US$38,518) than in its reference countries—France, Greece, Spain, and Portugal (ranging from US$41,885 to US$44,731). However, the PPP-adjusted price of sofosbuvir in Turkey (PPP$70,331) was 1.8 times higher than the price in France (PPP$38,077). Turkey’s GDP per capita (US$19,363) is more comparable to Brazil (US$16,320) than its current reference countries, but the nominal and PPP-adjusted prices of sofosbuvir in Turkey were 5.6 and 7.2 times higher, respectively, than the prices in Brazil. Similarly, Slovenia, Poland, and Slovakia use reference prices from countries with mostly richer economies. In the case of sofosbuvir, the nominal price in Norway is comparable to that in most of the other countries it references (i.e., Austria, Belgium, Germany, Denmark, Finland, Ireland, Netherlands, Sweden, and the UK). However, because Norway has greater purchasing power, PPP-adjusted prices indicate that the medicine may be more affordable for Norwegians in comparison to other countries in the OECD. This highlights the challenges of external reference pricing, which include the need to select comparable countries and to consider local purchasing power during price negotiation. There may also be differences between list prices and actual prices that may be hidden due to confidentiality agreements.
Our analysis suggests that sofosbuvir and ledipasvir/sofosbuvir are not “affordable” for most OECD countries at the nominal and PPP-adjusted prices, with Central and Eastern European countries being the most affected. While determining what is affordable or not is a value judgement, funding these treatments in these national health systems would consume large proportions of their TPE and increase pressure on existing budgets. We calculated that even funding treatment for only 10% of the potential population requiring sofosbuvir treatment would amount to at least 1% (UI: 1%, 16.2%) of current TPE in all countries analysed. However, treating only 10% of the infected population is unlikely to be ethically defensible or acceptable to the patient community. The cost of treatment increases substantially if treatment uptake is higher than 10%: if half of the eligible patient population is treated, five countries would spend an amount equivalent to more than half of their current TPE on sofosbuvir. Thus, the potential total cost of treatment presents a dilemma for payers and physicians, with some systems currently restricting access to these medicines to small groups of patients, despite the fact that almost all patients with chronic HCV infection are likely to benefit [17].
Where patients do not have access to subsidised treatment, individuals are unlikely to be able to pay for the medicines out of pocket. Based on ex-factory prices, the price of a 12-wk course of sofosbuvir would be equivalent to at least 1 y of income for the average income earner in 12 of the 30 countries analysed. It is not surprising that, given the price differences, HCV patients in high-income countries have been reported to import sofosbuvir at lower prices or even devise plans to receive treatment in India [54,55]. Assuming a 23% price reduction of nominal prices in both countries, consumers in the US pay US$64,680 for 12 wk of treatment, but if they instead obtained 12 wk of treatment of sofosbuvir in India, inclusive of airfare, hotels, and travel insurance (based on searches in common travel sites), they would pay only approximately US$6,000–US$7,000–10% of the price paid in the US.
Our analysis is limited by the accuracy of the estimates of the numbers of people infected and of the price information that was accessible. We have also not included all likely costs, such as the costs of combination treatment with ribavirin, other health care services, and increases in the duration of treatment in patients with cirrhosis; thus, our budget impact estimates are underestimates of the cost of treatment. We are also aware that in some countries, the prices are probably lower than the publicly accessible prices because of confidential discounts or rebates negotiated with the manufacturer. To minimise overestimation of price and budgetary impact, we made a conservative assumption that all countries, except for countries with special pricing arrangements, had the same price reduction as two of the largest payers—the US Centers for Medicare & Medicaid Services and the US Department of Veterans Affairs. However, neither this price reduction nor the sensitivity analysis using a 50% price reduction changes our overall conclusions about total expenditure on these HCV drugs in relation to TPE. There are also different types of discounts being offered that we have not included, such as rebate schemes that provide the medicines for free after 12-wk of treatment. We also did not attempt to adjust the estimates for pricing agreements based on treating only subgroups of patients. For example, in Portugal, the government has agreed to pay US$28,287 per patient treated irrespective of the duration of treatment. However, this arrangement is limited to less than 10% of the total eligible patient population over 3 y [56,57].
We have analysed the current situation for medicines for hepatitis C, but they are not the only group of medicines where high prices have affected or are affecting patient access. When the antiretrovirals were first launched for HIV, there were similar problems with price and affordability, due to the high price of medicines and the burden of disease. A combination of strategies and interventions—including global purchasing mechanisms, use of voluntary licensing agreements, compulsory licenses, price reduction through tiered pricing, and the rapid development of many high-quality generic products—contributed to improving the situation [58,59]. However, an analysis by Wirtz et al. [60] found persisting and substantial variability in antiretroviral prices in Latin American and Caribbean countries. They suggested that those middle-income countries with comparatively higher prices could afford to treat more patients if prices were lower. They also suggested the need to ensure effective procurement and price negotiation by procurement agencies. Currently, however, there is no global procurement or funding mechanism for hepatitis C medicines. This has limited the potential economies of scale that could be achieved by large-scale generic production. Existing licensing agreements, similar to those that are in place for antiretrovirals, exclude the upper-middle income and OECD countries that are currently paying the highest prices. Perhaps in contrast to the HIV epidemic, the burden of disease due to hepatitis C is more evenly distributed across countries, and the expectations about the outcome of treatment—a cure—are much higher, leading to much greater demand and expectations for access. The argument raised by Wirtz et al. [60] about the need for effective negotiation of prices therefore applies particularly to the medicines for hepatitis C, as there are still not enough alternative treatment options to allow for effective competition.
Setting prices and effective price negotiation is complex. Ramani and Urias [61] used economic game theory to evaluate when compulsory licenses can be effective in price negotiation in developing countries and emphasised the impact that having complete information (or not) about the product can have on negotiation outcomes. Purchasers usually do not have complete information on pharmaceutical products when they set prices. In particular, they usually do not have information about the costs of production or costs of research and development that have been claimed to contribute to the price set by manufacturers [61]. In the case of hepatitis C medicines, the confidential agreements on prices also make it difficult to compare prices accurately across countries, should a purchaser wish to define a “fair price”, for example, by adjusting prices by PPP. Differential or tiered pricing based on the national wealth of countries has been suggested as an approach to setting fair prices, but, to date, there has been no agreement on how to set the tiers [62].
The World Health Organization currently recommends that all patients with chronic HCV should be assessed for treatment [33], but the challenge is clearly how to provide treatment at a total cost that health systems and patients can afford. We had to exclude several countries from the analysis due to the unavailability of price data because, in most cases, the medicines were not publicly funded or not marketed at all. Moreover, as illustrated in our analysis, affordable prices could not be achieved in many OECD countries, even if they have price control systems, which suggests a need for an updated pricing system. While generic competition is likely to reduce prices in countries that are included in voluntary licensing agreements or that will issue compulsory licences, the impact of these strategies is unlikely to impact prices in OECD countries. Tiered pricing agreements are in use for these medicines, but are unlikely to be sufficient to increase access to the medicines for all countries. In order for countries to increase investment and minimise the burden of hepatitis C, governments and industry stakeholders will need to jointly develop and implement fairer pricing frameworks that deliver lower and more affordable prices.
|
10.1371/journal.pcbi.0030126 | A Computational Pipeline for High- Throughput Discovery of cis-Regulatory Noncoding RNA in Prokaryotes | Noncoding RNAs (ncRNAs) are important functional RNAs that do not code for proteins. We present a highly efficient computational pipeline for discovering cis-regulatory ncRNA motifs de novo. The pipeline differs from previous methods in that it is structure-oriented, does not require a multiple-sequence alignment as input, and is capable of detecting RNA motifs with low sequence conservation. We also integrate RNA motif prediction with RNA homolog search, which improves the quality of the RNA motifs significantly. Here, we report the results of applying this pipeline to Firmicute bacteria. Our top-ranking motifs include most known Firmicute elements found in the RNA family database (Rfam). Comparing our motif models with Rfam's hand-curated motif models, we achieve high accuracy in both membership prediction and base-pair–level secondary structure prediction (at least 75% average sensitivity and specificity on both tasks). Of the ncRNA candidates not in Rfam, we find compelling evidence that some of them are functional, and analyze several potential ribosomal protein leaders in depth.
| For decades, scientists believed that, with a few key exceptions, RNA played a secondary role in the cell. Recent discoveries have sharply revised this simple picture, revealing widespread, diverse, and surprisingly sophisticated roles for RNA. For example, many bacteria use RNA elements called “riboswitches” to switch various gene activities on or off in response to extremely sensitive detection of specific molecules. Discovery of new functional RNA elements remains a very challenging task, both computationally and experimentally. It is computationally difficult largely because of the importance of an RNA molecule's 3-D structure, and the fact that molecules with very different nucleotide sequences can fold into the same shape. In this paper, we propose a computational procedure, based on comparing the genomes of multiple bacteria, for discovery of novel RNAs. Unlike most previous approaches, ours does not require a letter-by-letter alignment of these diverse genomes, making it more applicable to RNA elements whose structure, but not nucleotide sequence, has been preserved through evolution. In an extensive test on the Firmicutes, a bacterial phylum containing well-studied organisms such as Bacillus subtilis and important pathogens such as anthrax, we recover most known noncoding RNA elements, as well as making many novel predictions.
| Recent discoveries of novel noncoding RNAs (ncRNAs) such as microRNAs and riboswitches suggest that ncRNAs have important and diverse functional and regulatory roles that impact gene transcription, translation, localization, replication, and degradation [1–3]. In the last few years, several groups have performed genome-scale computational ncRNA predictions based on comparative genomic analysis. In particular, Barrick et al. [4] used a pairwise, BLAST-based approach to discover novel riboswitch candidates in bacterial genomes, many of which now have been experimentally verified. Similar studies have been conducted in various bacterial groups [5–8]. More recent work has extended these searches to eukaryotes [9–13], discovering a large number of known microRNAs while producing thousands of novel ncRNA candidates. With some exceptions, such as [4] and [13], these approaches follow a similar paradigm, which is to search for conserved secondary structures on multiple-sequence alignments that are constructed based on sequence similarity alone. Typically, these schemes use measures such as mutual information between pairs of alignment columns to signal base-paired regions. However, the signals such methods seek, namely compensatory base-pair mutations, are exactly the signals that may cause sequence-based alignment methods to misalign, or alternatively refuse to align, homologous ncRNA sequences. Even local misalignments may weaken this key structural signal, making the methods sensitive to alignment quality, which is especially problematic on diverged sequences.
In this paper, we present a novel structure-oriented computational pipeline for genome-scale prediction of cis-regulatory ncRNAs. It exploits, but does not require, sequence conservation. The pipeline differs from previous methods in three respects. First, it searches in unaligned upstream sequences of homologous genes, instead of well-aligned regions constructed by sequence-based methods. Second, we predict RNA motifs in unaligned sequences using a tool called CMfinder [14], which is very sensitive to RNA motifs with low sequence conservation, and robust to inclusion of long flanking regions or unrelated sequences. Finally, we integrate RNA motif prediction with RNA homology search. For every predicted motif, we scan a genome database for more homologs, which are then used to refine the model. This iterative process improves the model and expands the motif families automatically.
In this study, we apply this pipeline to discover ncRNA elements in prokaryotes. We chose prokaryotes mainly because of the large number of fully sequenced genomes and the great sequence divergence among the species, which can be well-exploited by our approach.
Our approach has two key advantages. First, it is efficient and highly automated. Earlier steps are more computationally efficient than later steps, and we can apply filters between steps so that poor candidates are eliminated from subsequent analysis. Thus, even though we use some computationally expensive algorithms, the pipeline is scalable to larger problems. Besides providing RNA motif prediction, the pipeline also integrates gene context and functional analysis, which facilitates manual biological evaluation. Second, this pipeline is highly accurate in finding prokaryotic ncRNAs, especially RNA cis-regulatory elements.
To demonstrate the performance of this approach, we report our search results in Firmicutes, a Gram-positive bacterial division that includes Bacillus subtilis, a relatively well-studied model organism with many known ncRNAs. The method exhibits low false-positive rates on negative controls (permuted alignments), and low false-negative rates on known Firmicute ncRNAs. The RNA family database (Rfam) [15], a partially hand-curated database of noncoding RNAs, includes 13 ncRNA families categorized as cis-regulatory elements with representatives in B. subtilis. Of these, 11 are included among our top 50 predictions and a 12th appears somewhat lower in our ranking. Two other Rfam families are also represented among our top 50 predictions. In addition, both the secondary structure prediction and identified family members are in excellent agreement with Rfam annotation. For the 14 Rfam families mentioned above, we achieved 91% specificity and 84% sensitivity on average in identifying family members, and 77% specificity and 75% sensitivity in secondary structure prediction. Many promising novel ncRNA candidates were also discovered and are discussed below.
In outline, our pipeline consists of the following major steps. (See Figure 1, Materials and Methods, and the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline for more details.) First, we used the National Center for Biotechnology Information's (NCBI's) Conserved Domain Database (CDD) [16] to identify homologous gene sets. For each gene, we collected its 5′ upstream sequence. We call the set of 5′ sequences associated with one CDD group a dataset. cis-Regulatory elements are often conserved within such groups. Second, we applied FootPrinter [17], a DNA phylogenetic footprinting tool, to select datasets that are likely to host ncRNAs. In our experience, functional RNAs such as riboswitches often show low overall sequence conservation, but contain interspersed patches where conservation is high. FootPrinter is very effective at highlighting the latter regions. Third, we used CMfinder to infer RNA motifs in each unaligned sequence dataset. CMfinder is a structure-oriented local alignment tool that is robust to varying sequence conservation and length of extraneous flanking regions. We postprocessed motifs to identify distinct motifs corresponding to different RNA elements by removing poor and redundant motifs and clustering the rest based on overlap. Fourth, we used RaveNnA [18–20] to find additional motif instances by scanning the prokaryotic genome database. Riboswitches, for example, often regulate multiple operons that contribute to a single pathway, but no single CDD domain will be common to all of these operons. Thus, the search step was a powerful adjunct to the motif discovery process. These newly discovered motif members were incorporated into a refined motif model, again using CMfinder, and in some cases the search and motif refinement steps were repeated. Motif postprocessing was also repeated after the search/refinement steps. Both CMfinder and RaveNnA rely on the Infernal covariance model software package [21] for RNA motif modeling and search. Finally, we performed gene context analysis and literature searches (manually) for the top-ranking motifs.
We included 44 completely sequenced Firmicute species (see the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline) and 2,946 CDD groups in this study. For each of the three main steps—FootPrinter, CMfinder, and RaveNnA-based refinement—we produced scores to determine which candidates were worthy of continuing analysis. For evaluation purposes, we recorded the scores of candidates at each step, but eliminated none; in the future, we will use them as filters.
The initial CMfinder step produced 35,975 motifs in total. Motif postprocessing reduced this to 1,740 motifs grouped into 1,050 clusters. After RaveNnA-based refinement, more motifs were identified as redundant and removed. A total of 1,466 motifs remained, grouped into 1,060 clusters. (A few of the original clusters were subdivided based on divergent search results.) The full list of candidates is available in the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline.
To evaluate how many of our top candidates could have arisen by chance, we performed a randomized control experiment. We first computed CLUSTALW alignments of the 100 sequence datasets having the highest motif scores (before the RaveNnA scan). We then randomly permuted the alignments 50 times, maintaining the approximate gap pattern (see the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline). After degapping each permuted alignment (treating it as a set of unaligned sequences), we applied CMfinder, retaining the top-ranking motif from each randomized dataset. We used this collection of 5,000 motifs to estimate the background score distribution, and to infer p-values for predicted motifs in the original datasets. Results are shown in Figure 2. By this measure, all 100 top-scoring motifs have p-values less than 0.1, with the median at 0.016. In addition, 73 of the 100 candidates in the original dataset score higher than all motifs in the corresponding randomized datasets.
Note that this estimation of p-values is imperfect. In particular, with the scoring scheme we used, datasets containing phylogenetically close sequences tend to score well in comparison to more diverged sets, because permuting the CLUSTALW alignments preserves their sequence conservation. (Independently permuting individual sequences instead of alignments would be less realistic, since in practice cis-regulatory RNA motifs are often embedded in regions exhibiting some sequence conservation for other reasons.) Although imperfect, the significance of real motifs tends to be underestimated by this method.
To roughly assess the sensitivity with which the method discovers true ncRNAs, we looked at its recovery of known Rfam (version 7.0) families. We masked matches to Rfam's tRNA and rRNA models, since otherwise these widespread, strong motifs might hide nearby, weaker, but still interesting ncRNA structures. Other Rfam families were not masked and serve as a positive control for our methods. Table 1 shows the distribution of known Rfam families in our candidate list, together with their ranks after running FootPrinter, CMfinder, and RaveNnA. We used the refined motifs as the final output.
According to Rfam, B. subtilis contains members of 21 families, categorized into 13 cis-regulatory families, one intron element, and seven RNA gene families. We masked tRNAs and rRNAs (four of the seven gene families). Of the 17 remaining families, 13 appear within our top 50 candidates: 11 cis-regulatory families present in B. subtilis, together with two of the gene families (RNaseP_bact_b and SRP_bact). The four families not represented are two cis-regulatory elements (ykkC-yxkD and ydaO-yuaA), one RNA gene (tmRNA), and one intron element (Intron_gpI). The exclusion of Intron_gpI is not surprising, as we did not search intragenic regions. The ydaO–yuaA motif escaped detection because it is present in only three of the 68 sequences in its CDD group. The ykkC–yxkD and tmRNA motifs, although not among our top 50, would still have been ranked high enough to be discovered in a blind test. Note that, although our computational pipeline is oriented toward discovery of cis-regulatory elements, we sometimes find RNA genes such as RNaseP, SRP, and tmRNA because they happen to be conserved in synteny. We also found a partial tRNA motif, not masked since parts of the tRNA lie outside of the collected upstream sequences.
We can potentially filter the candidates at each step to scale this pipeline for larger genomes. In particular, we could have applied CMfinder to only the top half of the datasets according to FootPrinter, and performed genome scans on only the top 500 motifs, without missing any real Rfam families as listed in Table 1. On average, it takes FootPrinter less than 1 min, and CMfinder 10 min to process each dataset, while it takes RaveNnA 4.8 h to scan each motif. We could save considerable computation time by running expensive algorithms only on good candidates.
As shown in Table 1, the ranks for most known ncRNAs improve at each successive step of the pipeline, as more supporting evidence is found. Starting from FootPrinter motifs, CMfinder improves the alignment and identifies consensus secondary structure, while genome scans locate many more motif instances, typically providing still better alignments and additional clues to their functions.
To measure the quality of our automatically constructed motif models, we compared them with Rfam alignments for the same families. Rfam's covariance models are built from hand-curated “seed” alignments/structure annotations. These in turn are used to build Rfam's “full” alignments by automatically searching RFAMSEQ (http://www.sanger.ac.uk/Software/Rfam/ftp.shtml), a high-quality, nonredundant subset of EMBL (http://www.ebi.ac.uk/embl), and automatically aligning all hits.
For the 14 Rfam families in Table 1 for which we found good matching motifs, we selected the top two motifs from each family, and performed full-genome scans on RFAMSEQ, the same sequence database used to construct the Rfam full alignment. To reduce computation time, we did not scan eukaryote genomes, and the Rfam hits from these genomes were excluded from the following analysis. (This treatment affects only a few eukaryotic Cobalamin and Lysine hits, all believed to be Rfam errors or bacterial contamination in the genome sequences, plus a few THI hits, which are real.) For each motif, we selected scan hits at an E-value cutoff of 100, reconstructed the motif alignments using CMfinder, and removed the low-scoring instances (<20 bits). We compared these predicted motifs to corresponding Rfam full alignments, which serve as the gold standard in this test. Table 2 shows the accuracy of our motifs in membership prediction, motif coverage, and secondary structure prediction. Secondary structures were compared at the base-pair level, and only the base pairs with at least one end falling into the overlapped regions are counted. For both predicted motifs and Rfam full alignments, we removed noncanonical base pairs from each sequence. Of the two motifs chosen for each family, we report the one with better results.
For membership prediction, we achieved an average of 84% sensitivity and 91% specificity. The overlapped regions between predicted motif members and corresponding Rfam members account for 81% of the length of the predicted members, and 82% of the length of Rfam members. In the overlapped regions, the secondary structure prediction has 75% sensitivity and 77% specificity. These results suggest our predicted motif models are very accurate compared with Rfam models, which are learned from the hand-curated seed alignments.
For many riboswitch families, the main differences between our motif models and Rfam models are located in boundary regions. Our predicted motifs tend to include the transcription terminator (if present), which is a stable hairpin followed by a stretch of U's (e.g., Lysine, S_box, T-box). Although transcription terminators are functionally important, the Rfam riboswitch models do not include them. On the other hand, CMfinder tends to miss the closing helix of large multiloop structures (e.g., Cobalamin, ykoK). Most other differences are local perturbations such as small shifts or extra base pairs.
As shown in Table 2, we achieved more than 80% membership sensitivity for all families except yybP–ykoY, Glycine, and Cobalamin. The predicted yybP–ykoY motif differs from Rfam's motif mainly at the multiloop closing helix. Cobalamin and Glycine are two riboswitches with poor sequence conservation (46% and 51% average sequence identity, respectively). While our motifs from the initial full-genome scan may be too specific, sensitivity increases significantly with only a small loss in specificity after another iteration of RaveNnA scan and refinement (unpublished data).
For ykkC–yxkD and T-box, we predicted more members than Rfam. The predicted ykkC–ykxD motif includes the transcription terminator, which caused false positives in our full-genome scans. These false positives, however, all have much less significant E-values than the true positives, and hence are relatively easy to eliminate by inspection. In contrast, for T-box we believe most “false positives” (with respect to Rfam 7.0) are actually real. Out of 291 members not included in the Rfam full alignment, 127 are upstream of and on the same strand as aminoacyl-tRNA synthetase genes, where most T-box leaders are found, and the others are largely in poorly annotated regions.
We examined the best-scoring motif (see RNA motif discovery in Materials and Methods and the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline for details of the motif-scoring function) in each of the top 200 motif clusters. Of these 200 motifs, 116 were deemed unlikely to represent novel ncRNAs: they have covariance model scores less than 40 bits, single hairpin structures, and most were shorter than 30 nucleotides. (Many of these 116 are nevertheless biologically relevant. Many correspond to transcription terminators of upstream genes, and others contain known inverted repeat motifs targeted by DNA binding proteins.) Of 84 remaining motifs, 20 correspond to Rfam families, and 11 to hypothetical transposons. The remaining 53 are candidates for novel ncRNAs. Literature review suggests that many of these candidates are functional. We manually removed the redundant candidates with the same functional roles (for details, see Manual inspection and ribosomal protein leader analysis in Materials and Methods), and present the rest in Table 3.
In this study, we have presented a method for automatically finding cis-regulatory RNA motifs in prokaryotes. In a careful test with available sequenced Firmicutes, the method exhibited excellent rejection of negative controls (randomly permuted alignments) and excellent recovery of known, experimentally validated ncRNAs, including most riboswitches known in this bacterial group, as well as RNA elements such as 6S that have only recently been recognized there. Careful inspection and refinement of several novel motifs in ribosomal protein leaders provides compelling evidence that they are indeed conserved structures involved in regulation of these important operons.
In addition, our computational pipeline found dozens of other good RNA motifs that constitute strong candidates for novel functional elements, consistent with the increasing appreciation of the importance of RNA in all living organisms. Finally, our method is sufficiently scalable to be applied to all sequenced prokaryotes. We are in the process of doing so, and preliminary results include several novel riboswitch candidates.
We attribute the power of this pipeline to two key characteristics—a relaxation of the constraints on sequence conservation imposed by most previous methods, and integration of motif inference with genome-scale search. Our method performs motif inference on regions that are not defined by sequence conservation: we search unaligned sequences upstream of homologous genes, instead of multiple-sequence alignments constructed by sequence comparison tools. In addition, both the RNA motif–finding algorithm CMfinder and the RNA homology search algorithms RaveNnA/Infernal exploit structural information. Sequence conservation can be used as well, but is not required. Finally, automatic refinement of motifs to incorporate genome-scale search results has proven to be a powerful component of the pipeline (as in other contexts, such as PSI-BLAST [37]). The integration of these tools enables us to discover RNA motifs with low sequence conservation, and to expand the motif family with remote homologs. For example, the predicted motif for the Glycine Riboswitch has only 35% average pairwise sequence similarity. Remote RNA homologs with appropriate gene context are particularly important, as they are the strongest evidence, short of experiments, that a motif is functional, as well as providing clues to that function.
Future work will seek to strengthen this pipeline by improved exploitation of phylogeny and by an improved scoring system.
Phylogeny is crucial in all comparative genome analysis, without which the concept of conservation is meaningless. It is important in our work because the sequences upon which motif inference is performed are not evolutionarily equidistant, and the significance of conserved nucleotides and compensatory mutations are distance-dependent. Building on the classic phylogenetic likelihood model of Felsenstein [38], Pfold [39] and Evofold [12] use an RNA-oriented phylogenetic model to select from a given multiple-sequence alignment the regions that fit the structural model best. Unfortunately, in our application, neither an alignment nor an evolutionary tree is initially available, and, for our application, use of the corresponding species tree is inadequate in the common case when there are multiple sequences per species. Incorporating phylogeny into motif search is another challenge.
We would also like to improve our scoring scheme. As predicted motifs are subject to expensive manual evaluation and experiments, automatic candidate evaluation to guide resource investment is critical. Our current composite scoring system attempts to discriminate among potential RNA motifs by considering a set of features, including species distributions, structure stabilities, motif sizes, and local sequence conservation patterns. While we can easily recognize motifs that are significant in all these aspects, it is more difficult to order those that are only good by some, but not all, criteria. We have tried to combine the features automatically using machine-learning algorithms such as support vector and logistic regression. However, due to the heterogeneity of the features and limitations of available training data, the results were not as good as our handcrafted composite scoring function. One particular issue is that many of our top-scoring motifs are short single hairpins. They score well because they are widespread, structurally stable, and contain limited but clear sequence conservation. Although short motifs can be functionally important, many do not contain sufficient signal for genome scale homology scans, resulting in false positives that degrade the motif. Other complications include transposons, transcription terminators, DNA–protein binding sites, RNA-polymerase and RNA-ribosome binding sites, etc. The key challenge here is to design a metric that is correctly normalized across various known features and various types of ncRNAs with different sizes, structures, and phylogenetic divergence.
These opportunities for improvement notwithstanding, the approach described in this study has proven itself to be highly effective in discovering noncoding RNA elements in prokaryotes, and promises more discoveries to come.
We obtained genome sequences from 67 fully sequenced Firmicute species from the NCBI microbial database (RefSeq [40] release 14, 20 November 2005). We first collected amino acid sequences from all annotated protein-coding genes in these species, and categorized them based on NCBI's CDD (version 2.05) [16]. The CDD domain models are curated from various resources, including Pfam, SMART, and COG. In the NCBI microbial database, 92% of all functionally annotated proteins (i.e., with nonhypothetical description field) are assigned to at least one CDD group, as are 32% of “hypothetical” proteins. By definition, all members of a CDD group contain a conserved domain in their protein sequences. A group typically includes both orthologs and paralogs. We assigned proteins to a CDD group using “rpsblast” from the NCBI BLAST package (http://www.ncbi.nlm.nih.gov/BLAST), with an E-value cutoff threshold of 0.01. To reduce redundancy, we removed near-duplicate genomes from analysis. To do this, we created a vector for each complete genome, whose ith component was the number of predicted occurrences of the ith conserved domain in that genome. We normalized these vectors to have unit (Euclidean) length, and measured their similarity in terms of the projection of one CDD vector onto another (i.e., the dot product between them). Beginning with records assigned the lowest accession numbers, we then assembled a set of genomes by accepting each subsequent genome only when its similarity index with all selected datasets was less than 0.95. After removing redundancy in this way, 44 complete genomes remained for processing in subsequent steps.
We removed CDD groups that contained too few members (four or less), since motif discovery is unreliable on such small groups. We also removed 145 groups with too many members (70 or more), since motif discovery is expensive on such large groups.
For each gene in a CDD group, we collected a few hundred nucleotides upstream of its start codon, which typically includes both 5′ UTR and promoter sequences. The prevalence of operons in prokaryotic genomes complicates the extraction of the regulatory regions, as the desired regulatory region may be upstream of the entire operon rather than immediately upstream of the selected gene. To handle this complication in a conservative manner, we extracted the noncoding sequences upstream of the gene and upstream of its plausible operon using MicroFootPrinter [41]. Specifically, if the next coding region upstream is in the same orientation and fewer than 100 nucleotides upstream, this short intergenic sequence is included in our sequence dataset, and the same procedure is applied to the upstream gene. This process continues until interrupted either by a coding region in the opposite orientation or an intergenic region longer than 100 nucleotides. Finally, up to 600 nucleotides of the last intergenic region are included in the sequence dataset. After collecting the upstream sequences, we removed redundant sequences (95% sequence identity across 80% of the sequence according to BLAST), and masked regions that match tRNA or rRNA models in the Rfam database.
FootPrinter [17] identifies conserved sequence motifs in a set of unaligned homologous sequences using phylogenetic analysis. We scored each FootPrinter motif by the number of motif instances minus the corresponding parsimony score, and scored each dataset as the sum of its top 30 motif scores. The resulting scores are used to rank all datasets. This ranking is performed by MicroFootPrinter [41], a front end to FootPrinter [17].
We used CMfinder version 0.2 [14] for RNA motif prediction in unaligned sequences. For each dataset, we produced up to five single stem-loop motifs, five double stem-loop motifs, and used CMfinder heuristics to combine the motifs into more complicated structures if possible.
At various subsequent points, we ranked all CMfinder motifs using a heuristic scoring function that favors motifs with instances in diverged species, stable secondary structure, and local sequence conservation. We used local sequence conservation to discriminate trustworthy alignments with reliable anchors from purely structural motifs (e.g., alignments of single hairpins) that could easily arise by chance, while penalizing global sequence conservation, as highly similar sequences are more likely to be conserved by selection pressure on primary sequence than on structure. We refer to these scores as composite scores. The details of the scoring function are described in the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline.
Next, we filtered the motif set to remove poor motifs and combine redundant ones.
Operationally, a “motif” is a covariance model (CM), and a “motif instance” is a sequence that matches the CM with a score above a specified threshold. For each motif, we removed instances with CM score less than ten bits, and removed all but one copy of completely identical instances. Then, we ranked the motifs by composite scores, as outlined above and detailed in the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline. We further removed motifs with at most four instances and pairwise similarity greater than 0.95, and motifs with composite scores below 50. Afterwards, we selected up to four motifs for each dataset, selected in decreasing score order so that the lower ranking motifs do not overlap significantly with any higher ranking selected motif. By our definition, motif A overlapped significantly with another motif B if the number of nonoverlapping instances of A was less than 30% of the number of overlapping instances, and the average length of the nonoverlapping regions in the overlapped instances of A was less than half of the average length of the overlapped regions. Next, we removed redundant motifs from different datasets. We called motif A redundant with motif B if A overlapped significantly with B and the number of its predicted bases pairs not in B was less than 30% of the number of its base pairs in B. If A and B are redundant with each other, we chose the higher-ranking motif.
Finally, we clustered overlapping motifs as follows. We identified the overlap between motifs according to the genomic coordinates of their instances. One motif was grouped with another if at least half of its instances overlapped, and the overlapped regions are longer than half of the motif length. The motifs were clustered progressively, with high-ranking motifs processed first. We ranked clusters based on their highest-scoring motifs.
One of the key strengths of our method is its integration of motif discovery with motif search. Motif discovery is focused on groups of orthologs defined by common CDD membership, since such groups seem likely to be enriched for common cis-regulatory elements. However, many cis-regulatory elements such as riboswitches will be found near a variety of operons involved in a coherent pathway, which may not share a common CDD group. Hence, genome-scale search for additional motif instances is an important component of our approach. Additional instances allow us to construct more accurate motif models, as well as giving insight into potential biological roles for the elements.
Given RNA motifs produced by CMfinder, we searched for additional instances using Infernal CMs [21] accelerated with the ML-heuristic filter [20] implemented in RaveNnA 0.2f. For reasons of speed, two levels of search were used. The initial search database was derived from all 75 finished Firmicute genomes in RefSeq17 (30 April 2006) [40], a total of approximately 200 million nucleotides. Based on sequence annotations, we extracted only intergenic regions for searching, but extended each by 50 nucleotides in each direction to account for common errors in protein-coding gene annotations. The resulting database contained approximately 34 million nucleotides. This small database made it feasible to perform searches for all motifs (averaging 4.8 CPU h per motif), and reduced false positives when compared with the full-genome database. After motif refinement (incorporating hits from this “mini” scan), we performed “full” scans with selected motifs. Full scans examined the prokaryotic subset of the 8 GB RFAMSEQ dataset (version 7.0, March 2005), a total of approximately 900 MB. In particular, comparisons to Rfam (e.g., Table 2) were based on full scans, since Rfam full alignments are also derived from scans of RFAMSEQ. For model refinement, we ran CMfinder on all hits with RaveNnA E-values < 10. E-values were calculated as in [42]. The necessary extreme value distribution calculations dominate the run times for mini-scans, but not for full scans. The refined motif set is again postprocessed and ranked as described above.
To find which of our predicted motifs were already known, we compared them against the Rfam database. Specifically, we BLASTed our motif instances against Rfam full family members (produced by scanning Rfam covariance models on the RFAMSEQ genomic database; see [15]). For BLAST, we used a word size 12, and selected the hits with length greater than 30 nt, E-value < 10, and sequence identify exceeding 90%. These permissive BLAST thresholds resulted in a few isolated hits that we believe to be false positives. These motifs match fragments, each of about 30 bases, of the Rfam RNA-OUT, Intron-gpII, QaRNA, and RNaseP_bact_a families. In general, they are too short, weak, and/or isolated to be compelling, in sharp contrast to the matches reported in Table 1.
The genomic contexts of the refined motif instances were drawn using the Bio::Graphics modules of BioPerl [43]. For the ribosomal motifs, CMfinder structural alignments were trimmed to relevant regions and manually revised before conducting standard genome scans against the microbial subset of the RefSeq17 database. Hits with the correct genomic context were aligned according to the starting covariance model and manually revised once more to create final sequence alignments (available in the online supplement at http://bio.cs.washington.edu/supplements/yzizhen/pipeline). The Neural Network Promoter Prediction program [44] (version 2.2) was used to predict putative transcription start sites, and programs from the Vienna RNA package [45] were used to examine possible regulatory conformations.
Additional datasets and technical details are available at http://bio.cs.washington.edu/supplements/yzizhen/pipeline.
Five of the ribosomal protein leaders discussed in the Results section appear in Rfam release 8.0 (http://www.sanger.ac.uk/Software/Rfam), with the following accession numbers: L10 r-protein leader (RF00557), L13 r-protein leader (RF00555), L19 r-protein leader (RF00556), L20 (IF-3) r-protein leader (RF00558), L21 r-protein leader (RF00559).
|
10.1371/journal.pntd.0001897 | A Transcriptomic Analysis of Echinococcus granulosus Larval Stages: Implications for Parasite Biology and Host Adaptation | The cestode Echinococcus granulosus - the agent of cystic echinococcosis, a zoonosis affecting humans and domestic animals worldwide - is an excellent model for the study of host-parasite cross-talk that interfaces with two mammalian hosts. To develop the molecular analysis of these interactions, we carried out an EST survey of E. granulosus larval stages. We report the salient features of this study with a focus on genes reflecting physiological adaptations of different parasite stages.
We generated ∼10,000 ESTs from two sets of full-length enriched libraries (derived from oligo-capped and trans-spliced cDNAs) prepared with three parasite materials: hydatid cyst wall, larval worms (protoscoleces), and pepsin/H+-activated protoscoleces. The ESTs were clustered into 2700 distinct gene products. In the context of the biology of E. granulosus, our analyses reveal: (i) a diverse group of abundant long non-protein coding transcripts showing homology to a middle repetitive element (EgBRep) that could either be active molecular species or represent precursors of small RNAs (like piRNAs); (ii) an up-regulation of fermentative pathways in the tissue of the cyst wall; (iii) highly expressed thiol- and selenol-dependent antioxidant enzyme targets of thioredoxin glutathione reductase, the functional hub of redox metabolism in parasitic flatworms; (iv) candidate apomucins for the external layer of the tissue-dwelling hydatid cyst, a mucin-rich structure that is critical for survival in the intermediate host; (v) a set of tetraspanins, a protein family that appears to have expanded in the cestode lineage; and (vi) a set of platyhelminth-specific gene products that may offer targets for novel pan-platyhelminth drug development.
This survey has greatly increased the quality and the quantity of the molecular information on E. granulosus and constitutes a valuable resource for gene prediction on the parasite genome and for further genomic and proteomic analyses focused on cestodes and platyhelminths.
| Cestodes are a neglected group of platyhelminth parasites, despite causing chronic infections to humans and domestic animals worldwide. We used Echinococcus granulosus as a model to study the molecular basis of the host-parasite cross-talk during cestode infections. For this purpose, we carried out a survey of the genes expressed by parasite larval stages interfacing with definitive and intermediate hosts. Sequencing from several high quality cDNA libraries provided numerous insights into the expression of genes involved in important aspects of E. granulosus biology, e.g. its metabolism (energy production and antioxidant defences) and the synthesis of key parasite structures (notably, the one exposed to humans and livestock intermediate hosts). Our results also uncovered the existence of an intriguing set of abundant repeat-associated non-protein coding transcripts that may participate in the regulation of gene expression in all surveyed stages. The dataset now generated constitutes a valuable resource for gene prediction on the parasite genome and for further genomic and proteomic studies focused on cestodes and platyhelminths. In particular, the detailed characterization of a range of newly discovered genes will contribute to a better understanding of the biology of cestode infections and, therefore, to the development of products allowing their efficient control.
| Cestodes are a major group of helminths infecting humans and domesticated animals, of global sanitary and economic importance [1] and include the parasites responsible for echinococcosis [2] and cysticercosis [3]. While genomic initiatives are now well advanced for some of these organisms [4], and proteomic analyses have recently been carried out [5], [6], [7], our knowledge at the transcriptomic level remains limited. We selected Echinococcus granulosus as a suitable target for analysis of gene expression by key life cycle stages.
E. granulosus is the agent of cystic echinococcosis, a major zoonosis that affects humans and a wide range of domestic and wild animals worldwide [8], [9]. Control efforts have had little global impact and the infection remains highly endemic in the Southern Cone of Latin America (Argentina, Chile, Uruguay, Southern Brazil and Peru), as well as in large areas of Asia and Africa, and in patches of Europe and North America [10]. Although difficult to assess due to underreporting, the disease has a substantial global burden, which is estimated at over 1 million DALYs per year [11].
The E. granulosus life cycle involves two mammalian hosts. The intermediate hosts (ungulates and, accidentally, humans) ingest eggs that develop into a hydatid cyst containing larval worms or protoscoleces (PS), bathed in hydatid fluid that includes parasite as well as host proteins. The PS are clearly differentiated into distinct tissues (the rostellar pad, the neck, the suckers and the body; [12]), and the hydatid cyst is delimited by a cyst wall (CW), consisting of an inner germinal layer of metabolically active parasite cells and an outer protective acellular mucin-rich laminated layer [13], which appears to be evolutionarily optimized for eliciting non-inflammatory responses from the host immune system [14]. The cyst is usually surrounded by a host-derived collagen capsule, the adventitial layer. Infection in the definitive host (always a canid) arises from ingestion of PS encysted in the viscera of the intermediate hosts. PS are activated by contact with stomach acid and enzymes, which can be reproduced in the laboratory by exposure to pepsin at low pH. In the duodenum, they develop into adult tapeworms that can reside for long periods, indicating that PS establishment requires modulation of the host immune response [15], [16]. In addition, E. granulosus has a fascinating alternate reverse development, as PS escaping from a ruptured cyst in an intermediate host are able to differentiate asexually into secondary hydatid cysts (reviewed by [17]).
To study the molecular basis of the host-parasite interaction, and to gain understanding of E. granulosus developmental and metabolic aspects, we have analyzed the transcriptomes from the CW, the resting PS (i.e. as present in the hydatid cyst) and pepsin/H+-activated PS (PSP). We previously reported a new method to construct full-length cDNA libraries by an oligo-capping method [18]. Because some E. granulosus mRNAs bear a trans-spliced leader (SL) sequence [19], which blocks oligo-capping [18], in this study we have analyzed both oligo-capped and SL-bearing transcripts to ensure that we also captured genes that are processed by trans-splicing. Transcriptome analyses focusing on other parasitic platyhelminth species have been published, including the trematodes Schistosoma mansoni [20], Schistosoma japonicum [21], Clonorchis sinensis [22], Opisthorchis viverrini [23] and Fasciola hepatica [24], and the cestodes Mesocestoides corti (syn. vogae) [25], Echinococcus multilocularis [26], Moniezia expansa [27], and Taenia solium [28], [29], [30]. The recent availability of high throughput sequencing technologies has also stimulated transcriptome surveys of various adult liver flukes [31], and the cestode Taenia pisiformis [34].
We report the analysis of 9,452 ESTs from ∼2,700 distinct genes, generated from E. granulosus larval stages. These data represent about 20% of the estimated 11,000 protein-coding genes of the parasite [4]. In addition, they reveal the expression of remarkably abundant putatively non-protein-coding transcripts (ncRNAs) that could either be active by themselves as long ncRNAs or represent precursors of small RNAs. The full genome sequence of E. granulosus, now nearing completion [4], together with the transcriptomic data presented here will constitute invaluable resources to deepen our understanding of the biology of this parasite.
E. granulosus PS and CW (germinal and laminated layers) were recovered under aseptic conditions from hydatid cysts of the G1 genotype, present in the lungs of naturally infected bovines in Uruguay. Cysts were collected during the routine work of local abattoirs in Montevideo (Uruguay). The G1 genotype, the common ‘sheep strain’ which infects cattle in areas of intense sheep farming, has recently been reclassified into E. granulosus sensu stricto (that also includes G2 and G3; [35], [36]); it has a worldwide distribution and its presence coincides with high prevalence of human infection [9]. PS and CW were stored at −80°C in Trizol reagent (GibcoBRL) until RNA extraction. One fraction of freshly isolated PS was incubated with pepsin prior to treatment with Trizol (PSP). The processing of parasite materials and the construction of cDNA libraries were previously described in detail [18]. In brief, two sets of full-length enriched libraries were prepared using total RNA from the three materials (CW, PS and PSP). RNA from each source was reverse transcribed with a tagged oligo-dT. In the first set of libraries, full-length mRNAs were ligated to a 5′oligo, permitting PCR amplification of the intact mRNA population (oligo-capped (GR) libraries). In the second set, a 5′primer for the E. granulosus SL sequence [19]) was used (SL libraries).
The libraries were plated out and random colonies picked for EST sequencing. A small-scale analysis (5′first-pass sequencing) was initially carried out on AB3730 instruments (Applied Biosystems) in the GenePool Facility (Edinburgh), on about 250 randomly isolated clones from each library, as previously described [18]. Further sequencing from these libraries was performed at the Sanger Institute and the Centro de Biotecnologia in MegaBace 1000 instruments (Amersham Biosciences). An alkaline lysis method for plasmid DNA preparation in 96-well plates was used; plasmid DNA was subsequently purified through Millipore plates and resuspended in 30 µl of MilliQ water. 5′ and 3′ ESTs were carried out from each plasmid, using 500 ng of DNA and the DYEnamic ET Terminator Kit (Amersham Biosciences), according to the instructions of the manufacturer.
Sequence processing was performed using the PartiGene pipeline [37]. Raw sequence trace data was processed to remove low quality, vector, host (bovine), linking and poly(dA) sequences. For annotation purposes, each sequence was subject to a BLASTN search against the non-redundant DNA database [38] as well as a BLASTX search against the non-redundant protein database [39]. Sequences have been submitted to dbEST [40]. Sequences were collated and clustered on the basis of BLAST similarity to derive groups of sequences, which putatively derive from the same gene using the software package - CLOBB [41]. These groups were then used to derive a set of consensus sequences using the freely available software package PHRAP (P. Green unpublished data). It is worth noting that, while the CLOBB clustering tool attempts to minimize the generation of chimeric consensi, transcripts representing alternative splice forms may be clustered into separate groups whereas members of the same gene family can be merged into the same group [41]. This set of consensus sequences together with those groups containing only a single sequence (‘singletons’) form a non-redundant set of gene sequences, which we refer to as a partial genome. The corresponding E. granulosus dataset is available from PartiGeneDB (http://www.compsysbio.org/partigene/annotation/viewset.php). For comparative purposes, we also performed TBLASTX comparisons against: 1) a set of 688 eukaryotic partial genomes in our in-house partial genome database (PartiGeneDB - [42]); 2) a set of 3,178 non-redundant (clustered) sequences derived from 12,483 ESTs generated from E. multilocularis (K. Brehm and C. Fernández, personal communication); and 3) a set of 2,271 non-redundant (clustered) sequences derived from 3,947 ESTs generated from Fasciola hepatica (M. Berriman, personal communication).
Peptide predictions were performed using the prot4EST software [43]. Domain and signal peptide predictions were obtained using PFAM [44] and SignalP V3.0 [45], respectively. Similarity analyses comparing peptides among three different datasets were performed using the SimiTri comparison tool [46]. Alignments were initially created using ClustalW2 [47] and refined manually. Analyses of the presence of putative O-glycosylation sites, signals for GPI incorporation and transmembrane helices were carried out with the tools available at the ExPASy Proteomics Server (http://expasy.org/proteomics): NetOGlyc, PI predictor and TMHMM, respectively. Putative platyhelminth orthologs of E. granulosus cDNAs were identified using BLAST by applying the best-reciprocal-hits approach [48]. For the phylogenetic analysis of identified tetraspanins, an alignment was manually refined taking into account the consensus of 6-Cys-a and 8-Cys-a cysteine patterns (adapted from [49] and [50]) and used to construct a minimum evolution phylogenetic tree using MEGA 4 [51] with default parameters. Bootstrap values were expressed as percentage of 1000 replicates and were considered significant if >50%.
A total of 9,462 ESTs (7722 5′ESTs and 1740 3′ESTs) were generated from six full-length enriched E. granulosus cDNA libraries constructed from three sources of parasite material: CW, PS and PSP. These represent key stages in the parasite life cycle that interface with either the intermediate host (mainly the CW, during the chronic phase of infection) or the definitive host (mainly PSP, at the onset of infection). The boundaries between stages are not absolute, and each preparation should be considered as ‘highly enriched’ in transcripts from the corresponding stage. For example, the CW from a healthy cyst usually contains some PS, and pepsin/H+ treatment does not activate all PS in a sample because their development inside the cyst is not synchronous.
Following strategies targeted at cloning cDNAs with an intact 5′ end, we constructed two sets of libraries, either by exploiting the 5′ trans-spliced leader sequence (SL libraries) [52] or by using an oligo-capping method based on the GeneRacer protocol (GR libraries) to select full length cDNAs [53]. The two library construction methods produced sequences of similar length (Table 1). After processing, the dataset gave 2,700 putative genes comprised of 1,328 clusters containing more than one sequence and 1,372 ‘singletons’ (see E. granulosus dataset at PartiGeneDB: http://www.compsysbio.org/partigene/annotation/viewset.php) (Table 1). A total of 166 putative genes (23 clusters and 143 singletons) were derived from 3′ESTs only. Taking into account the library construction strategies and that a majority of ESTs were carried out from the 5′end, this number provides an (over)estimated maximum of the transcripts that could correspond to non-overlapping regions of the same gene.
The distribution of the clusters according to the parasite stage and also the type of cDNA library in which they were found are summarized in Figure 1. The GR and SL libraries were largely non-overlapping as expected from previous work [18], with only ∼10.5% (140/1328) of clusters comprising reads from both types (Figure 1A and B). The lack of overlap between GR and SL libraries is due to the fact that the GR oligo rarely ligates to the 5′ SL, likely because of some structural feature of the Echinococcus SL (perhaps the formation of a short hairpin loop, as was recently proposed [54]).
In both GR and SL library datasets, the proportion of clusters associated with only one stage (‘stage-specific clusters’) was considerable (Figure 1C). For example, 43% of hydatid cyst wall GR clusters (106/244) were not found in other stages, and 26% of hydatid cyst wall SL clusters (103/399) were similarly stage-specific. In addition, 44% (332/747) of clusters involving PSP in GR and SL libraries, were absent from the untreated PS sample. The high level of stage-specific expression may reflect the sharply contrasting environments and developmental programs associated with the different stages. On the other hand, as we have not sampled the transcriptome to exhaustion, some of these differences are more likely due to limited sampling rather than to differential gene expression. In fact, a much greater overlap between libraries was noted when considering clusters derived from five or more sequences (Figure 1D; see also next section).
Table 2 presents the most highly represented transcripts from each analyzed stage (CW, PS and PSP). Surprisingly, the most highly abundant transcripts in the three parasite stages (EGC00310 and EGC03058) were non-protein coding RNAs (ncRNAs) showing similarity to the E. granulosus repetitive DNA element, EgBRep [55]. As described in more detail below, these molecules are closely related and can be regarded as a single cluster with micro-variation. Interestingly, a separate cluster showing similarity to EgBRep was largely PS specific and, in contrast to the previous ones, derived from trans-spliced cDNAs (EGC02791).
All other highly expressed transcripts coded for proteins, most of which showed similarity to sequences from other platyhelminths. The CW expressed two stage-specific transcripts at high levels: a novel sequence coding for a putative apomucin (EGC00317) and a member of the tetraspanin family (EGC00290). Interestingly, a further tetraspanin-containing transcript (EGC00446) was restricted to the PS and PSP stages (see below). The remaining highly expressed clusters corresponded to transcripts represented in the three stages but showing some stage bias in the number of ESTs. It is noteworthy that the majority (12/16) corresponded to trans-spliced cDNAs, including enzymes participating in energy metabolism (notably, EGC00369, fructose biphosphate aldolase, highly abundant in the CW) and antioxidant systems (EGC00370, thioredoxin-like, abundant in the three stages). The cDNAs that were not trans-spliced comprised three ribosomal proteins, prominent in PSP (EGC00474, EGC00350 and EGC00467); and a putative splicing factor, highly expressed in the CW (EGC00843).
Four SL-bearing transcripts encoding hypothetical proteins were amongst the most highly expressed; two of them in all three stages (EGC00548 and EGC00373) and two in PS (EGC00658; EGC00524). Given that high levels of expression are often indicative of essential roles, these represent interesting targets for further investigation.
Consideration of all clusters (see Table S1) reinforced these observations; in fact, clusters representing highly expressed transcripts (≥20 ESTs) included: non-protein coding RNAs (EGC02905; EGC00351; EGC00637 and EGC01002), abundant in GR libraries; and mRNAs coding for lactate dehydrogenase (EGC00284), another enzyme from the glycolytic pathway, that predominated in CW; and several ribosomal proteins (EGC00595; EGC00605; EGC00634; EGC01107) in PSP. In addition, a protein containing a dynein light chain domain (EGC00319), immunolocalized to the PS tegument and the germinal layer (EgTeg; [56]) and detected in cyst fluid, PS and germinal layer [6], was highly expressed in all stages, mainly in PSP and CW (see also next section).
From the 2,700 clusters identified, we were able to derive 2,584 peptide predictions which were each scanned for putative PFAM domains [44]. Overall, 1,034 domains, representing 193 unique domains, were identified in 808 peptides, as detailed in Table S1. Figure 2 shows the most abundant domains identified within the dataset. We compared the abundance of each PFAM domain relative to EST datasets obtained from ten additional platyhelminths and five other lophotrochozoans. Even though care must be taken while interpreting the data because all sets are partial, this type of comparisons provides a first glimpse into species differences (see e.g. [57], [58]).
In fact, despite the datasets differing in size and the diversity of stages used (see legend to Figure 2 for details), some interesting trends emerged. Four of the top five domains were consistently abundant across the Lophotrochozoa: WD domain (PF00400); RNA recognition motif (PF00076); ankyrin repeat (PF00023) and EF hand (PF00036), as were also the Ras family (PF00071); mitochondrial carrier protein (PF00153); and tetratricopeptide repeat (PF00515).
Relative to other species, the protein kinase domain (PF00069) was relatively poor within both Echinococcus species. Conversely, the tetraspanin domain (PF00335) was expanded in platyhelminths; E. granulosus proteins identified as containing this domain are analyzed further below. In addition, both trematode and cestode lineages showed expansion in the dynein light chain domain (PF01221), whereas the annexin (PF00191) and Like-Sm ribonucleoprotein (LSM; PF01423) domains appeared expanded only in the cestode lineage. Two of these domains (dynein light chain and annexin) are associated with cellular organization and the third one (LSM) with RNA metabolism.
Thirteen predicted polypeptides (mostly from PS and PSP libraries) contained the dynein light chain domain, involved in intracellular motility of vesicles and organelles along microtubules [59]. Six predicted proteins contained up to four annexin domains; some being highly represented in the CW (EGC00693) or the PSP (EGC00359) stages. The annexins (or lipocortins) are eukaryotic calcium-dependent phospholipid-binding proteins implicated in multiple functions, including exocytosis and endocytosis, signal transduction, and extracellular matrix organization [60].
Thirteen predicted polypeptides encoded by transcripts isolated from all E. granulosus stages contained the LSM domain present in an RNA-binding protein superfamily involved in pre-mRNA splicing and mRNA processing [61]. Interestingly, a homologue in Schmidtea mediterranea (Smed-SmB) is essential for the proliferation of planarian stem cells [62]. Finally, a domain related to bacterial transferase hexapeptide (PF00132), present in a number of transferase protein families [63], appeared expanded in the E. granulosus dataset, entirely within the SL library-derived ESTs.
Each of the 2,584 peptide predictions (1,848 of which had an initiation methionine) were parsed through the SignalP web server [45], to determine the presence of a putative secretory or anchor sequence. In total 254 peptides (9.8%) were predicted to possess a secretory leader signal (similar to a previous study focusing on T. solium larvae [30]), while an additional 157 (6.1%) were predicted to contain a signal anchor. There was no obvious bias to either the GR and SL, or to specific stage libraries (Table S1).
Previously, in a transcriptomic study of the parasitic nematode Nippostrongylus brasiliensis, we noted that signal sequence-bearing proteins showed reduced evolutionary conservation [64]. This observation was confirmed and extended in a subsequent study: parasitic nematodes were found to have a greater proportion of novel, secreted proteins than free-living ones [65]. Here, we examined the conservation of proteins predicted to be secreted within the E. granulosus dataset. Based on TBLASTX similarity to partial genomes derived from 688 different eukaryotes, we identified genes/clusters that were unique to E. granulosus (15.8%; 14.7% of predicted peptides), specific to Echinococcus (30%; 27.7% of predicted peptides), specific to platyhelminths (44.5%) or specific to metazoa (55.2%; Figure 3). However, of peptides with a predicted secretory leader sequence, 18.1% were unique to E. granulosus and 35.8% were specific to Echinococcus. While the former difference is not statistically significant, the latter, being about 30% higher than in the overall dataset, is (p<0.005, Chi-squared test). For signal anchor sequences, the proportions were: 15.3% and 24.2% respectively. While errors in prediction accuracy related to both the SignalP software [45] and truncated sequences may erroneously classify some peptides as containing a secretory sequence, there is no reason to expect that such errors would occur disproportionately amongst the various groups. These results therefore suggest that secreted proteins in Echinococcus are less evolutionarily conserved than non-secreted proteins. However, these differences in conservation are much less dramatic than previously reported for N. brasiliensis, in which 48.9% of signal positive peptides could be described as genus-specific compared to 26.8% for the dataset overall [64].
As shown in Figure 2, E. granulosus is a parasitic cestode and is grouped within the phylum Platyhelminths, along with Trematodes (e.g. Schistosoma) and Tricladids (e.g. Schmidtea and Dugesia) [66]. Platyhelminths are related to Annelida and Mollusca within the Lophotrochozoa [67], [68]. To investigate the similarity relationships of the genes within our dataset to these various taxonomic groupings, we employed the tool SimiTri [46], that allows simultaneous display and analysis of relative similarity relationships of one dataset to three different databases, to visualize the data from the taxonomic split shown in Figure 3.
SimiTri analysis showed that E. granulosus sequences were, as expected, more closely related to E. multilocularis and T. solium than to either Tricladids or Trematodes (Figure 4A). In addition, very few genes were found to be more similar to a Tricladid species than to a Trematode. This could reflect the closer phylogenetic relationship between Cestodes and Trematodes, which are usually grouped in the Neodermata clade [69]. However, these results may be biased from the larger number of Trematode sequences (74,794) used in this analysis relative to Tricladid sequences (22,327). To examine the impact of sequence coverage, we compared the BLAST score distribution of the E. granulosus sequences to randomly selected sets of 22,327 Trematode sequences (Figure S1). This analysis suggests that the higher number of Trematode sequences, rather than the closer relationship between Cestodes and Trematodes, was responsible for the larger number of E. granulosus hits to Trematodes compared with Tricladids.
Interestingly, Figure 4B shows a relatively low level of enrichment of E. granulosus sequences with closer similarity to other Lophotrochozoan (Mollusca and Annelida) sequences than to other Eukaryotes. However, the low level of enrichment for the former may again simply represent a smaller dataset of comparator sequences. Finally, Figure 4C shows the relationships to three other major clades of metazoans – Deuterostomia, Nematoda and Arthropoda. The majority of genes showed greater similarity to arthropod and/or deuterostome sequences than to nematode sequences. Given the supporting evidence for the grouping of Nematoda and Arthropoda (Ecdysozoa; [68], [70], [71]), this latter result while potentially indicating the highly diverged nature of nematode genes compared with the other two phyla, nonetheless highlights the limitations of using BLAST sequence similarity scores to infer phylogenetic relationships. See [21], [22], [24], [25], [30] for further discussion on similarity between cestode and trematode datasets and other metazoans.
From the BLAST analyses, we were also able to identify a set of 391 E. granulosus genes that shared sequence similarity only with platyhelminths. Table 3 shows the 34 putative genes that had significant sequence similarity only to four or more other platyhelminth EST datasets. Of these, 19 showed sequence similarity neither to a gene or a protein of known function nor to an identifiable protein domain; of these, five were predicted to be secreted. Only three genes were found to possess a characterized protein domain while 15 showed significant sequence similarity to previously identified or predicted platyhelminth genes with functional annotation. Due to the ubiquity of these gene products within platyhelminths, and although we await their full characterization, they represent a rich source for the identification of potentially novel pan-platyhelminth drug targets.
The SL libraries differed from GR-based libraries in a number of aspects, including a lower level of stage-specificity (Figure 1B). Interestingly, a higher overlap of clusters from SL libraries was observed between CW and PSP, the two stages showing comparatively higher metabolic activity, than between PS and either PSP or CW. In addition, as previously noted, a majority of abundant clusters originated from SL libraries (see Table 2). As only a fraction of the transcriptome is processed by trans-splicing (estimated to be 25–30% in E. multilocularis [19], [72]), our equivalent sampling from libraries derived through the two methods (46% GR sequences vs 54% SL sequences; see Table 1) could explain this bias. However, taking into account that ESTs from either type of library were equally redundant, the previous observations may indicate that a set of trans-spliced transcripts is indeed highly expressed in all surveyed stages.
Altogether, 187 clusters, representing 21 ESTs from GR-based libraries and 1,428 ESTs from SL-based libraries, were found to possess a full SL sequence at the 5′end (Table S1). Ligation of the GR oligo to the 5′ spliced leader (SL) was observed in the case of highly expressed transcripts (e.g. EGC00373 and EGC00435 in Table 2). In addition, oligo-capped transcripts lacking SL were found in clusters corresponding to genes that are usually trans-spliced (e.g. EGC00369 and EGC00647 in Table 2). These transcripts could correspond to molecules not yet trans-spliced in vivo; or to genes that can be expressed with or without the SL [19]. Regarding the latter possibility, it is noteworthy that high-throughput sequencing of the SL trans-spliced transcriptome of the tunicate Ciona intestinalis revealed that the conventional dichotomy of ‘trans-spliced’ vs ‘non-trans-spliced’ genes should be supplanted by a view recognizing frequently and infrequently trans-spliced genes categories [73].
The set of clusters possessing a full SL sequence allowed us to further characterize E. granulosus SL bearing transcripts. Because a conserved and unique feature of flatworm SLs is the presence of a 3′end AUG able to serve as an initiation methionine in vivo [74], we analyzed whether the SL ATG was in frame with the major ORF of the cDNAs and, furthermore, what proportion of these was full-length. Of the 187 SL-bearing clusters, 143 were predicted to be full length, using the ATG in the SL as the putative start codon (8 of these are listed in Table 2 together with 6 where the SL ATG is not in frame with the predicted ORF). It is likely that not all E. granulosus trans-spliced transcripts actually use the SL AUG in vivo, as alternative AUGs were often found within a few codons of the SL AUG. This was the case, e.g. in 4/8 cDNAs listed in Table 2 (an additional ATG was present within 5 codons 3′of the SL); however, in the remaining 4 cDNAs, the SL AUG would be required as an initiation methionine if the N terminus was to fully correspond to those of phylogenetically conserved orthologous proteins. Thus, our data provide additional evidence that the SL AUG could serve as an initiation methionine in platyhelminths, as indicated by earlier studies in this phylum [19], [74], [75], [76]. Moreover, we searched for E. granulosus orthologs of 35 S. japonicum genes known to be both expressed by trans-splicing and using the SL AUG as an initiation methionine [74]. Putative orthologs (BLAST bit score ≥100; or >40% identity over at least 90% coverage) were identified for 16, 15 of which were derived from SL libraries; of these, 10 would use the SL AUG as an initiation methionine, indicating that the use of trans-splicing and initiation from the SL AUG is itself phylogenetically conserved in the Neodermata.
We then examined the potential functional relationships between the products encoded by different trans-spliced mRNAs. No particular functions or processes were found to be enriched within trans-spliced cDNAs, in agreement with previous reports in other flatworms [19], [75], [76], [77], including a recent study that identified a large set of trans-spliced genes in S. mansoni using high-throughput sequencing (11% out of ∼11,000; [78]). In contrast, and as was described for tunicates [73], [79], [80], genes encoding ribosomal proteins tended not to be trans-spliced (see Table S1).
Although polypeptides could be predicted from 95.7% of the clusters, the remaining 116 clusters appeared to be non-protein coding. Quite strikingly, a majority (66) of these – accounting for ∼700 ESTs mostly from GR libraries of the three stages – contained segments displaying high identity (≥90%) with fragments of EgBRep, a previously described middle repetitive DNA element from E. granulosus, showing structural similarities to mobile elements [55]. Some of these clusters were relatively abundant (notably, EGC00310 and EGC03058; see Table 2; and also EGC02905, EGC02701, EGC00351, EGC00367 and EGC01002, all with ≥20 ESTs; see Table S1). Collectively, the ESTs within these clusters represented >10% of sequences from each stage.
The assembled sequences of clusters EGC00310 and EGC03058 corresponded to full-length transcripts of ∼900 nt, putatively capped and polyadenylated (as shown by the presence of the GR oligo at the 5′end and poly(dA) at the 3′end in non-trimmed sequences). These transcripts matched the minus strand of EgBRep over ∼150 nt at both the 5′ and 3′ends (Figures 5A and 5B). Moreover, multiple reads mapping between these conserved flanking sequences showed microdiversity in the central tract, reaching a global identity of about 90%. Manual assembly of the EgBRep-containing ESTs, avoiding artificial collapse of contigs by the automated algorithm (see Figure 5C), identified two clusters, named Cluster A (512 ESTs, including all but 4 of the ESTs from the original clusters EGC00310 and EGC03058) and Cluster B (187 ESTs) (see Figure 5B and Table S2). Interestingly, some EgBRep-containing sequences were trans-spliced (notably, those in EGC02791; see Tables 2 and S1). These were almost exclusively from the PS library and corresponded to trans-spliced polyadenylated transcripts of ∼225 nt that included the 150 nt 3′end fragment similar to EgBRep (see Figure 5B and ClusB.contig10 in Table S2).
Comparison of these consensus sequences to the current version of the E. granulosus genome (available at http://www.sanger.ac.uk/cgi-bin/blast/submitblast/Echinococcus) identified scaffolds showing regions of high identity (90–100%) with the manually assembled contigs, and revealed that some of them are likely to derive from transcripts processed by cis-splicing (e.g. ClusB.contig8 has 2 exons, and ClusB.contig7 has 3 exons). For every EgBRep-containing contig, several highly similar fragments (>80% identity) were present in the draft genome.
Transcripts with similarity to EgBRep were also identified in E. multilocularis ESTs from an oligo-capped metacestode library, including presumed orthologs of the abundant E. granulosus transcripts derived from EGC00310 and EGC03058, with an overall similarity between Echinococcus spp. of 92% (see e.g. clusters EMC00034 and EMC00190 in PartiGeneDB). Moreover, abundant, putatively non-protein coding cDNAs, showing scattered segments of 85–100% identity with the E. granulosus EgBRep-containing cDNAs, were present in the T. solium transcriptome (∼6,100 clusters available at PartiGeneDB; see e.g. TSE00132, TSE00439 and TSE00790).
The occurrence of these EgBRep-containing cDNAs in all surveyed stages is a major feature of the larval transcriptome of E. granulosus. Structurally, these transcripts correspond to a class of long (>200 nt) non-protein coding RNAs (ncRNAs), first described during the large scale sequencing of mouse full-length cDNA libraries [81], that resemble mRNAs (being capped, polyadenylated and often spliced), yet lacking clear open reading frames. Recent genome-wide studies have identified large numbers of long ncRNAs in human and model organisms [82], [83], [84], [85], [86], [87] and shown that some of them overlap with repeats [82], [83], [85], [87], and that short conserved regions nested in rapidly evolving sequences are present in long ncRNAs conserved between species (see e.g. [82], [85], [87]). In addition, some C. elegans primary long ncRNAs have been found to be trans-spliced [87]. Long ncRNAs have been implicated in the regulation of gene expression through a variety of mechanisms (reviewed by [88], [89]) and were found to participate in stem cell pluripotency and differentiation [90]. In addition, an appreciable portion can be processed to yield small RNAs ([84]; reviewed by [89]).
Because EgBRep-containing transcripts are associated with repeats, they could be precursors of piRNAs, a class of strikingly diverse small RNAs implicated in transposon silencing in the metazoan germ-line (reviewed by [91]). piRNAs are likely generated via processing of long single-stranded precursors (primary piRNAs), transcribed by RNA polymerase II from discrete genomic loci (piRNA clusters), some of which are highly enriched in transposons and other repeats (reviewed by 91,92). Notably, a long ncRNA associated with an insect transposable element has been proposed to be the precursor of rasiRNAs [93], a class of piRNAs first identified in Drosophila melanogaster [94]
In recent years, the piRNA pathway has emerged as a distinctive trait of planarian somatic stem cells (neoblasts) and piRNAs were found to predominate among small RNAs in the neoblasts of S. mediterranea [95], [96]. Neoblasts are the only mitotically active cells in planarians; they are responsible for their extraordinary regenerative capacity and are known to also give rise to germ-line stem cells (reviewed by [97]). In the Neodermata, and in cestodes in particular, there is evidence that similar mechanisms of self-renewal exist ([98], [99]; reviewed by [54]). It remains to be determined, therefore, whether EgBRep-containing long ncRNAs are themselves active molecular species or represent precursors of small RNAs; in the latter case, they could be precursors of piRNAs in proliferating cells from each of the parasite materials sampled in our study.
Genes in several key energy production pathways were differentially expressed in the surveyed stages, with fermentation predominating in CW, and gluconeogenesis being up-regulated in CW and PSP (Table 4). The data are consistent with the previously reported existence of a complete tricarboxylic acid (TCA) cycle in E. granulosus [100], [101]. Genes encoding components of respiratory complexes I, III and IV were also identified, indicating that aerobic respiration can take place in the surveyed stages (Table 4, Figure 6).
Some enzymes belonging to key fermentation pathways coupled to glycolysis were also found (Figure 6). In particular, cytosolic fermentation to lactate appeared to be an important metabolic route in the germinal layer: lactate dehydrogenase (LDH) was highly expressed in the CW. In addition, transcripts for phosphoenol pyruvate carboxykinase (PEPCK) and cytosolic malate dehydrogenase (cMDH) were also present (mainly in CW libraries), indicating the existence of a route for mitochondrial fermentation via malate dismutation (Figure 6), which is an unusual feature of helminth metabolism. The existence of these fermentative pathways is consistent with the fact that lactate and succinate were described as the major end-products of carbohydrate metabolism [102].
In addition, enzymes for gluconeogenesis (fructose-1,6-bisphosphatase; and also PEPCK), glycogenolysis and glycogenesis were also found (Table 4), in agreement with the accepted view that glucose is the major respiratory substrate and glycogen the main energy store molecule in flatworms [102].
Considered globally, the germinal layer appears to possess a high metabolic activity (see Table 4), involving, in particular, fermentative pathways. The synthesis of the laminated layer towards the outside of the cyst and the generation of brood capsules containing PS towards the inside are major metabolic demands for the germinal layer, of both energy and intermediate metabolites. It is possible that the oxygen supply within the hydatid cyst may be limited by the thick laminated layer. In this respect, it is worth noting that in vitro growth of E. multilocularis metacestode has been reported to be more active under microaerobic conditions, suggesting metabolic adaptations to low oxygen [103], which may include glycolysis through generation of lactate, and use of the PEPCK-succinate pathway. Alternatively, the up-regulation of lactate fermentation (and malate dismutation) could be due to ‘the Warburg effect’ observed in cancer and all proliferating cells [104], [105]. Indeed, proliferative tissues convert most glucose to lactate through ‘aerobic glycolysis’, regardless of whether oxygen is present; lactate fermentation and other anaerobic pathways are thought to facilitate the uptake and incorporation of nutrients into the biomass (reviewed by [104], [105]; see also Figure 6). Interestingly, glutamine synthetase, which is also highly expressed in proliferating tissues, was observed to be an abundant transcript in the CW (and PS; see EGC00519 in Table S1). In addition to the essential role of glutamine in protein and nucleotide synthesis, this amino acid is an anabolic substrate. Glutamine can be converted into pyruvate via TCA and glutaminolysis providing biosynthetic carbons for the production of macromolecules [106], [107].
Parasites must cope with oxidants and reactive oxygen species (ROS) derived from their own aerobic metabolism and also from host activated cells such as phagocytes. Several redox-based antioxidant enzymes were present in all surveyed stages, and many of them were highly expressed (Table 5). Peroxiredoxins (Prx, formerly known as thioredoxin peroxidases), glutathione peroxidase (GPx), thioredoxin (Trx), selenoprotein W, glutaredoxin (Grx) and methionine sulfoxide reductase (Msr) were among the 7% most highly expressed genes. A cytosolic Prx was particularly abundant in the CW, while expression of Gpx, Msr-b (stereospecific for the Met-S-sulfoxide), selenoprotein W and the Trx-related EGC00370 increased upon pepsin/H+ PS activation. Although Cu/Zn superoxide dismutase(s) are known to be expressed in both PS and CW [108] we did not identify corresponding clusters in our data. We also failed to identify any clusters corresponding to catalase transcripts, confirming previous reports of absence of catalase activity in E. granulosus and other flatworms (reviewed by [109]). Globally, the data indicate that a broad range of antioxidant defences are dependent on the enzyme thioredoxin glutathione reductase (TGR), which functions as a metabolic hub for transferring electrons to glutathione (GSH), Trx, Grx and from these latter to their targets, such as Prx, Msr, GPx, etc (reviewed by [109], [110]). Although TGR was absent from the dataset (which may be due to the fact that it is encoded by a long mRNA, of 2.8 kb), all known direct and indirect targets of this enzyme were present.
Many eukaryotic selenoproteins are important antioxidant enzymes with higher turnover rate than their Cys homologs. E. granulosus TGR is known to be a selenoenzyme [111] and the GPx and selenoprotein W transcripts we detected also contain an in-frame UGA codon and a SECIS (Selenocysteine insertion sequence) element [112]. However, the Msr-b is a Cys-containing protein and not a selenoprotein, as is the case of one of the isoforms present in mammals [112].
In addition to acting as direct and indirect antioxidant, GSH also serves a detoxification role through glutathione S-transferases (GSTs). These enzymes are primarily involved in detoxification of electrophiles, but many of them possess additional or distinct functions, including the neutralization of oxidative stress (through e.g. removal of lipid peroxides, inactivation of secondarily oxidized products and regeneration of S-thiolated proteins), as well as the catalysis of metabolic reactions not involved in detoxification (e.g. biosynthesis of leukotrienes and prostaglandins) (reviewed by [113], [114]). Four distinct GSTs belonging to different families and classes were present in our dataset. Three belong to the family of cytosolic GSTs: two are of sigma class and one corresponds to the previously characterized mu-class enzyme [115]. The last one belongs to the microsomal GST family. Although the precise functions of these GSTs remain to be determined, sigma-class GSTs have been mostly implicated in prostaglandin synthesis [113], [114].
Several clusters coding for apomucins were identified in the larval transcriptome on the basis of a high Ser/Thr content offering multiple potential O-glycosylation sites consistent with mucin synthesis. A set of 4 apomucins expressed by the CW were not found in PS and PSP, whereas a second set (16 clusters) were present in all assayed materials (Figure 7 and Table 6).
The CW apomucins have a distinct structure. Three (EGC00317, EGC02904 and EGC04254) were the most highly expressed protein-coding transcripts of the germinal layer altogether (4% of ESTs from the CWGR library, with EGC00317 accounting for 2.6%; Table 2). These feature no tandem repeats, contain a very high proportion of putative O-glycosylation sites with interspersed basic residues and a common C-terminal sequence that is predicted to correspond to a signal for the addition of glycosylphosphatidylinositol (GPI) anchors. Two of them (EGC02904 and EGC04254) may be splice or allelic variants of each other (they differ mainly by a 40 amino acid insertion in the mucin core), and carry unpaired Cys residues in their N-terminal extension. The fourth CW apomucin (EGC05092) has the same N-terminus as the proteins predicted from EGC02904 and EGC04254 but it has a distinct mucin core with two different tandemly repeated units of 10 amino acids. All four apomucins have a marked predominance of Thr over Ser residues, suggestive of secreted mucins.
Interestingly, a putative ortholog of EGC00317 was identified among E. multilocularis ESTs from an oligo-capped metacestode library (see EMC00019 at PartiGeneDB and Figure 7A). The overall identity between the predicted Echinococcus spp. apomucins was 84%; it was high (>95%) over the signal peptide and C-terminal sequence, but surprisingly low for putative orthologs of these organisms over the rest of the molecule (∼63%).
This family of apomucins could form the backbones of the mucins from the fibrilar component of the laminated layer, a unique Echinococcus structure whose synthesis is known to be a major metabolic activity of the germinal layer, as was recently proposed in a comprehensive review of this structure [13]. The high level of expression of these apomucins and the existence of an ortholog in the transcriptome of E. multilocularis metacestodes support this inference. In addition, Thr is known to be the most abundant amino acid of laminated layer preparations (reviewed by [13]), consistent with the preponderance of this residue in the predicted mature apomucins (Table 6). Finally, in agreement with intense mucin biosynthesis, a number of CW clusters encode enzymes and transporters involved in the assembly of O-glycans (Table 7). In particular, probably reflecting the marked predominance of galactose in the major glycans purified from the laminated layer [116], several transcripts correspond to proteins participating in galactose metabolism, the synthesis of UDP-galactose and its translocation across Golgi membranes.
The second set of mucin-encoding transcripts (EGC02902 and related clusters in Figure 7B and Table 6) include a very short acidic N-terminus followed by a varying number of tandemly repeated units of 28 amino acids. These repeats each contain two acidic residues and about 15 Ser/Thr (Ser/Thr ratio ∼0.8), all of which would be glycosylated. The C-terminal extension ends with a stretch predicted to be a transmembrane helix, indicating that they are cell-surface proteins. These mucins could thus be constituents of the mucin coat known to cover the tegument of larval and adult worms [17]. The presence of transcripts from these genes in the CW libraries could derive from apomucin expression in the germinal layer or from developing PS in the tissue of the CW.
Fourteen clusters encoded members of the tetraspanin family (TSP, Figure 2) and some of them were among the most abundant in the dataset (notably, EGC00290 and EGC00446; Table 2). TSPs are a large family of highly expressed type II membrane proteins (200–350 amino acids) with a characteristic topology (four transmembrane domains; small and large outer loops, short N- and C-terminal tails). They have conserved disulfide bridges in the large extracellular loop (LEL) that are the basis of a structural classification ([117]; reviewed by [118]).
Eleven E. granulosus TSPs (EgTSPs; Table 8) showed substantial similarity to TSPs from E. multilocularis (Em-TSPs; [119]) and T. solium (TsT-24; [120]); (see Figure S2). Two EgTSPs (EGC00709 and EG04933) were most similar to schistosome TSPs. EgTSP EGC04745 was not classified with other flatworm TSPs and was most similar to an insect TSP. Some transcripts likely encode variants of the same TSP (proteins predicted from the two contigs in EGC00097 share 94% identity, and EGC00817 and EGC03391 share 93% identity), as has been observed in schistosomes [121].
Phylogenetic analysis of the EgTSPs identified three clades (Figure 8A). Group A includes two close paralogs (the variants EGC00817 and EGC03391, and EGC00129, with 67% identity), and two more distant proteins. Group B comprises three proteins, including another pair of close paralogs (the variants from EGC00097 and EGC00849, with >70% identity). A third pair of close paralogs forms a separate group (Group C; EGC00290 and EGC00446, with 48% identity), while the remaining two EgTSPs (EGC00709 and EGC04745) appear quite distant from the rest, especially the one with no flatworm homolog.
Alignment of the LEL variable region of EgTSPs highlighted their Cys patterns and, in some cases, allowed assigning them to specific groups (Figure 8B). Most EgTSPs have 6 Cys in their LEL and conform to the 6-a pattern [49], [50]. In addition, some Group A EgTSPs show structural features present in CD63-like TSPs [49]; interestingly, these EgTSPs also contain a putative tyrosine-based sorting signal (YXXΦ, where Φ is a bulky, hydrophobic residue), which is known to be involved in CD63 intracellular trafficking (reviewed by [122]; see Figure S2). The other EgTSPs from Group A have only 4 Cys. It is likely that, as described for other animal TSPs, Cys 4 and 5 were secondarily lost in these proteins [49], [50]. Group B and Group C EgTSPs and the one predicted from EGC04745 also have a 6-Cys-a pattern but they lack other structural features of CD63-like TSPs and their LELs are longer. Finally, EGC00709 encodes a TSP with 8-Cys-a pattern and conforming to the TSPAN15-like group [49]. CD63- and TSPAN15-like EgTSPs have been identified in all metazoan groups ([49], [123]; see also [121]).
A majority of EgTSPs were expressed in the CW, some of them at high levels (in particular, EGC00290 from Group C, EGC00299 from Group B, EGC00817 and EGC00129 from Group A). EGC00446 (Group C) and EGC00643 (Group B) included ESTs derived solely from PS and PSP libraries (Figure 8A and Table 8). A similar level of developmentally regulated transcription was recently reported for schistosome TSPs [121].
Most of the EgTSPs identified in our dataset represent cestode expansions of the family. Indeed, excepting two proteins, they are considerably distant even from trematode TSPs. This observation supports the hypothesis that gene duplication and rapid divergence have been major driving forces in the evolution of TSPs, where lineages are phylum-specific and many genes appear to be species-specific [50], [121], [124]. Interestingly, distinct members from the identified groups would be up-regulated in particular stages. TSPs regulate migration, fusion and signaling by acting as organizers of multimolecular membrane complexes involving the plasma membrane, intracellular vesicular compartments and exosomes (reviewed by [118] and [125]). Novel TSPs may thus have evolved to fulfill the highly diverse requirements of distinct parasite stages. In this context, it is worth noting that TSPs have been assayed as vaccine antigens for schistosomiasis [126], [127] and primary alveolar echinococcosis [119] in mouse models. In both systems, some level of protection was observed upon immunization with particular TSPs. Mammalian TSPs involved in highly specific functions are also amenable to targeting using antibodies, with considerable therapeutic potential against various pathologies (reviewed by [128]).
Three clusters sharing sequence similarity with E. granulosus antigen B (AgB) were identified within our dataset: EGC00327, EGC00450 and EGC03328. AgB is a highly abundant lipoprotein present in hydatid fluid [129]. It is the most relevant antigen for hydatid disease diagnosis (see e.g. [130]) and has been associated with a number of immunomodulatory functions in the host [131]. AgB has been extensively characterized at the protein [5], [132], [133] and gene levels (see e.g. [4], [134]); and its physiological lipid ligands have recently been described [135]. EGC00450 and EGC03328 with 21 and 14 ESTs respectively, derived exclusively from PSGR and PSPGR libraries. They corresponded to virtually identical AgB3 variants that differ only in the length of the acidic stretch. The third cluster, EGC00327 with 8 CWGR ESTs, corresponded to AgB4. These findings indicate a clear bias in the expression of AgB3 and AgB4 subunits in the different parasite materials.
Remarkably, no ESTs encoding AgB1 or AgB2 were found in our dataset. These subunits were originally cloned from PS [136], [137], and the corresponding cDNAs have subsequently been detected by several authors, mainly in PS (see e.g. [138], [139], [140]).
Two studies, on E. granulosus [134] and E. multilocularis [141], have reported developmentally regulated expression of AgB subunits in the Echinococcus life cycle, using real-time PCR and semi-quantitative PCR, respectively. Both included material from the germinal layer and the adult stage; but resting PS were only assayed in E. multilocularis [141] and pepsin/H+-activated PS only in E. granulosus [134]. The two studies found that AgB1, B2, B3, and B4 were expressed in the CW. AgB4 was expressed at lower levels than the other subunits, and was most highly expressed in CW. AgB1 and B3 predominated in PS [141], whereas AgB3 was highly dominant in PSP [134] and adult worms [134], [141] (the latter also expressed some AgB5 [134], [141]). If we assume that expression in PS is similar between Echinococcus spp., our data on AgB3 and AgB4 are consistent with these reports. In contrast, the absence of cDNAs corresponding to AgB1, B2 and B3 in the CW library, and to AgB1 in the PS library appear to contradict the previous observations. We hypothesized that the discrepancy could derive from the oligo-capping procedure, which is known to exclude transcripts whose 5′UTRs do not efficiently ligate to the oligo-cap [18]. To explore this possibility, we cloned cDNAs from AgB1–AgB4 obtained by RACE or RLM-RACE: no difference was detected in cloning efficiencies for the transcripts of the different genes. The analysis of the 5′UTR from oligo-capped cDNAs showed the presence of different numbers of GT repeats in AgB1–AgB4 subunits, which did not appear to interfere in the cloning procedure. AgB1 was the most expressed gene in the germinal layer and AgB3 in PS, while AgB2 was the least expressed in both stages (A. Arend and A. Zaha, unpublished). Consequently, we have no explanation as to why AgB1 encoding ESTs were absent from our dataset.
Although cestodes are a major group of parasites of humans and animals, extensive genomic coverage has only recently begun for these organisms [4]. Key advances have been made with transcriptomics for several platyhelminths, including mainly parasitic trematodes (see e.g. [20], [21], [22]) and the planarians S. mediterranea [142], [143], [144]; and Dugesia japonica [145], to which we can now add our gene discovery project on the dog tapeworm E. granulosus. This has fulfilled our objectives of greatly expanding the information available on genes expressed by larval parasites, and of identifying a series of candidate molecules involved in the host-parasite cross-talk in hydatid infections.
The new data we present in this report provide insights on many important biological features of this fascinating parasitic organism. Firstly, E. granulosus follows an elaborate developmental program through its life cycle that relies on the activity of somatic stem cells (reviewed by [54]). The highly expressed long ncRNAs we have identified may be involved in the regulation of gene expression through that program in response to environmental cues in the host. In addition, we have identified a number of genes reflecting specificities of particular stages including those whose expression is up-regulated by pepsin-acid activation. Regarding these latter, a major finding was the identification of a family of Kunitz-type serine protease inhibitors associated mostly with pepsin/H+-treated PS, which we have previously described [146]. Another major finding relates to the metabolic activity needed to maintain the intermediate host interface. Indeed, we found clear signs of enhanced energy production in the germinal layer and identified several genes that could form the mucin backbones of the laminated layer, as well as enzymes involved in their glycosylation.
Secondly, we have identified numerous new potential genes for investigation, either because they are highly expressed by the parasitic larvae and are novel in sequence, or because by sequence similarity to genes of known function they are attractive candidates for drug targeting. The generation of effective new pharmaceuticals is critically important for both Echinococcus species (and also for T. solium), which cannot be controlled by current agents and which therefore can develop life-threatening infections [1].
Thirdly, the dataset richly illustrates the dynamics of multigene family evolution in platyhelminths, both with respect to selective expansion of particular families and with regards to the subset bearing predicted signal peptides. At this stage, before the completion of the genome, gene family expansion at the transcriptomic level could represent either or both gene multiplication and diversification, or elevated expression of a similar repertoire of gene variants. In either instance, certain gene families are clearly of emphasized importance in E. granulosus.
Finally, because ESTs were derived from full-length enriched cDNA libraries prepared from carefully selected parasite materials, our data will constitute a high quality complement of the full genome sequence of the parasite, now nearing completion [4]. Indeed, preliminary sequence comparisons found that 94% of our predicted consensus sequences could be mapped to the current draft genome of E. granulosus (>90% identity over >80% consensus sequence length – data not shown).
The E. granulosus ESTs generated in this work were deposited in dbEST with the following accession numbers: BI243991-BI244549; BQ172910-BQ173849; BU582013; CN648894-CN653840; CV223690-CV223699; CV678041-CV681224; CV678546; CV678796.
|
10.1371/journal.ppat.1000064 | A Transgenic Drosophila Model Demonstrates That the Helicobacter pylori CagA Protein Functions as a Eukaryotic Gab Adaptor | Infection with the human gastric pathogen Helicobacter pylori is associated with a spectrum of diseases including gastritis, peptic ulcers, gastric adenocarcinoma, and gastric mucosa–associated lymphoid tissue lymphoma. The cytotoxin-associated gene A (CagA) protein of H. pylori, which is translocated into host cells via a type IV secretion system, is a major risk factor for disease development. Experiments in gastric tissue culture cells have shown that once translocated, CagA activates the phosphatase SHP-2, which is a component of receptor tyrosine kinase (RTK) pathways whose over-activation is associated with cancer formation. Based on CagA's ability to activate SHP-2, it has been proposed that CagA functions as a prokaryotic mimic of the eukaryotic Grb2-associated binder (Gab) adaptor protein, which normally activates SHP-2. We have developed a transgenic Drosophila model to test this hypothesis by investigating whether CagA can function in a well-characterized Gab-dependent process: the specification of photoreceptors cells in the Drosophila eye. We demonstrate that CagA expression is sufficient to rescue photoreceptor development in the absence of the Drosophila Gab homologue, Daughter of Sevenless (DOS). Furthermore, CagA's ability to promote photoreceptor development requires the SHP-2 phosphatase Corkscrew (CSW). These results provide the first demonstration that CagA functions as a Gab protein within the tissue of an organism and provide insight into CagA's oncogenic potential. Since many translocated bacterial proteins target highly conserved eukaryotic cellular processes, such as the RTK signaling pathway, the transgenic Drosophila model should be of general use for testing the in vivo function of bacterial effector proteins and for identifying the host genes through which they function.
| Like many pathogens, the human gastric bacterium Helicobacter pylori orchestrates infection through the activity of proteins that it translocates into host cells. The H. pylori translocated protein, CagA, which shares no homology to any other proteins, is a significant risk factor for H. pylori–associated diseases including gastric cancer. Experiments in tissue culture cells have shown that CagA can activate SHP-2 phosphatase, a component of the receptor tyrosine kinase signaling pathway. Based on this activity, CagA has been proposed to function as a mimic of Gab proteins that serve as adaptors in this signaling pathway. We have developed a transgenic Drosophila model to test this hypothesis in the tissues of an organism. We demonstrate that CagA can substitute for Gab and restore developmental defects caused by the loss of the Drosophila Gab, including promoting photoreceptor specification in the developing eye. Furthermore, we show that CagA functions similarly to Gab because it requires the Drosophila SHP-2 to exert its effect on photoreceptor development. Our transgenic Drosophila model provides new insight into CagA's activity in tissues and will allow us to identify host factors through which CagA functions to manipulate cellular signaling pathways and promote disease.
| The human pathogen, Helicobacter pylori, infects the stomachs of at least half the world's population and chronic infection is associated with the development of diseases such as gastritis, peptic ulcers and gastric cancer [1]. A major virulence determinant of H. pylori is the cytotoxin associated gene A (CagA) which is translocated into host cells via a type four secretion system (reviewed in [2]). Inside host cells, CagA is phosphorylated by Src family kinases on tyrosines contained in repeated five-amino acid motifs (EPIYA) in CagA's carboxyl terminus. Phosphorylated CagA disrupts receptor tyrosine kinase (RTK) signaling pathways by directly activating Src homology 2 (SH2) domain containing tyrosine phosphatase (SHP-2) (reviewed in [3]). Normally SHP-2 is activated by the scaffolding adaptor Grb2-associated binder (Gab) proteins, thereby amplifying RTK signaling pathways to control cell growth, differentiation and survival (reviewed in [4]). The Gab proteins occupy a pivotal position in RTK signaling pathways by interacting directly with RTKs such as the c-Met receptor of the Hepatocyte growth factor/Scatter factor (HGF/SF) as well as downstream cytoplasmic proteins including SHP-2, v-crk sarcoma virus CT10 oncogene homolog (avian)-like (Crk(L)), and Growth factor receptor-bound protein 2 (Grb2) (reviewed in [5],[6],[7]). Although CagA shares no sequence similarity with Gab proteins, CagA has been shown to activate SHP-2 in tissue culture cells, resulting in cell elongation [8],[9]. Similarly, in tissue culture cells CagA has been found to associate with c-Met, Crk(L) and Grb2 [10],[11],[12]. Based on these interactions, CagA has been hypothesized to mimic Gab proteins and to function as an oncogene by over-activating RTK signaling [13]. The significance of CagA's interactions with RTK signaling pathway proteins, however, has only been explored in tissue culture cells.
We have developed transgenic Drosophila with inducible CagA expression as a model to understand CagA's mechanisms of action in complex epithelial tissues. In order to test the hypothesis that CagA can function as a Gab substitute, we investigated CagA activity in a well-characterized Gab-dependent process, the specification of photoreceptors in the Drosophila eye [14],[15],[16]. The Drosophila compound eye, whose crystalline array of facets or ommatidia are exquisitely sensitive to perturbations in cell specification, has been used as a powerful system for the discovery and genetic analysis of RTK signaling components [17],[18]. Drosophila RTK signaling proteins are highly conserved with their mammalian orthologues and oncogenic mutations in these proteins, such as those that constitutively activate RTK receptors or their downstream effectors, function similarly in both Drosophila and mammalian cells [19]. The Drosophila model also offers elegant tools for genetic manipulations including the UAS/GAL4 system [20] for expression of transgenes in a tissue specific manner, the FLP/FRT system for the generation of somatic mutant clones [21], and null mutations in most RTK signaling pathway members, which allow us to probe the in vivo requirements for CagA's activation of RTK signaling pathways. Finally, Drosophila are amenable to forward genetic approaches that will facilitate the discovery of host factors required for CagA function in eukaryotic cells [22].
RTK signaling is required for multiple steps of Drosophila photoreceptor development. The Drosophila epidermal growth factor receptor (EGFR) is necessary for cell proliferation in the early eye imaginal disc, cell survival in the differentiating region of the disc behind the morphogenetic furrow, and recruitment of all photoreceptors except R8 [23]. A second RTK, Sevenless (SEV) is required exclusively for the R7 photoreceptor to adopt the appropriate fate, as opposed to becoming a nonneuronal cone cell [24] (reviewed in [25]). The Drosophila Gab adaptor, Daughter of Sevenless (DOS) is required for full signaling through both the EGFR and SEV pathways [16]. Clones of eye imaginal cells lacking DOS activity fail to proliferate and produce few photoreceptors, similar to clones lacking EGFR [16],[26],[27]. The EGFR pathway is required additionally for multiple aspects of Drosophila development [28].
Here we show that CagA can substitute for the Drosophila Gab adaptor, DOS, and rescue phenotypes associated with loss of dos, including larval lethality and photoreceptor differentiation. We further demonstrate that CagA functions through the Drosophila SHP-2 homologue, Corkscrew (CSW) similar to Gab. Our work demonstrates the power of using a genetically tractable system like Drosophila to dissect the mechanism of action of a prokaryotic protein that modulates a conserved eukaryotic signaling pathway.
To determine if the Drosophila system would be useful for dissecting the molecular mechanism of CagA-induced activation of RTK signaling, we examined whether CagA exhibited similar properties when expressed in Drosophila tissue to those previously observed in mammalian tissue culture cells. We used P-element mediated transgenesis to generate Drosophila with a transgene encoding an N-terminal hemagglutinin (HA) tagged CagA under control of the yeast GAL4 upstream activating sequence (UAS-CagA). Additionally, we generated transgenic flies with a mutated version of CagA lacking the EPIYA tyrosine phosphorylation motifs (UAS-CagAEPISA). These transgenic flies were crossed to flies that expressed the GAL4 transcription factor under tissue-specific or inducible promoters to express CagA in specific cells and at specific times during development. In the experiments described here, the GMR-GAL4 line was used to express CagA in all cells of the developing imaginal eye disc after the morphogenetic furrow.
Western analysis of anti-HA affinity purified proteins from heads of adult UAS-CagA/GMR-GAL4 flies showed that CagA was expressed (α-HA) and phosphorylated (α-P-Tyr, Figure 1A). Similar to CagA's distribution in tissue culture cells [8],[29], we showed in the Drosophila eye disc CagA was localized predominantly to the cell cortex (Figure 1C). Examination of the cellular morphology of the pupal retina revealed that CagA expression caused disorganization of the epithelium. The wild type retinal epithelium is organized into regular cell clusters, each containing a single R7 and R8 photoreceptor (Figure 1D). In retina expressing CagA, the normal cell shapes and neighbor relationships were perturbed (Figure 1E), similar to CagA-dependent epithelial disorganization observed in mammalian tissue culture monolayers [29],[30]. When we examined the eyes of adult flies expressing a single copy of CagA with GMR-GAL4, we observed a perturbation of the normal crystalline array of the ommatidia (compare wild type, Figure 1F, with CagA expression, Figure 1G). Expression of two copies of the UAS-CagA transgene dramatically enhanced the eye phenotype, indicating that the developmental pathways disrupted were sensitive to the amount of CagA expressed (Figure 1H). Expressing one copy of the CagA mutant lacking the tyrosine phosphorylation sites (CagAEPISA) did not perturb the crystalline array of the adult eye to the extent caused by wild type CagA (Figure 1I) even though the CagAEPISA protein was expressed at similar levels as CagA (Figure 1B). Dose dependent perturbations of Drosophila eye patterning, as observed with CagA expression, have been used as the basis for genetic screens for modifiers of the rough eye phenotype to elucidate several signaling pathways, including RTK pathways. [17],[31]
To test the hypothesis that CagA functions as a prokaryotic mimic of eukaryotic Gab proteins, we asked whether CagA expression could rescue phenotypes caused by the loss of the Drosophila Gab, DOS. DOS functions downstream of multiple RTKs during development, and homozygous dos loss-of-function mutants rarely develop into pupae and never survive to adulthood [16]. Rescue of dos mutants' lethality has been used as an in vivo assay to determine the function of specific domains of DOS [26]. We therefore determined the percentage of dos homozygous mutants that survived to the pupal stage of development with or without CagA expressed ubiquitously with temporal precision using the heat shock inducible Hsp-GAL4. The frequency of dos homozygous mutants was scored as a percentage of expected pupae that should develop if the dos mutants showed no lethality defect. As expected, a low percentage (33%) of homozygous dos mutant pupae expressing only Hsp-GAL4 were observed (Figure 2A). When CagA was expressed, we observed a significant increase to 89% of the pupae developing that lacked dos (Figure 2A). These results indicate that CagA can substitute for essential functions of DOS during Drosophila development.
To specifically test whether CagA could substitute for Gab in photoreceptor development, we generated mitotic dos/dos clones within the eye using the FLP/FRT recombinase system [27],[32]. In these experiments the dos mutation was recombined onto a chromosome arm containing a centromere proximal FRT recombination site and maintained in trans to a chromosome containing the same FRT site as well as a GFP transgene. By expressing FLP recombinase in the developing eye we induced mitotic recombination between FRT sites, which generated clones of homozygous cells (+/+ and dos/dos) in an otherwise heterozygous background (dos/+). The dos/dos mutant cells were distinguished by their lack of GFP, and the photoreceptors were visualized by staining for the photoreceptor-specific protein ELAV. As previously reported [16],[26] the dos/dos clones rarely contained photoreceptors and were composed of very few cells (Figure 2B–E), due to the dual requirements for EGFR signaling in cell survival and photoreceptor specification [23]. As expected, expression of DOS with GMR-GAL4 in dos/dos cells resulted in much larger clones with increased numbers of photoreceptors (Figure 2B, F–H). Expression of CagA in dos/dos cells was able to rescue clone size and photoreceptor development similarly to expression of DOS with the same driver (Figure 2B, I–K). Two independent dos mutants gave similar results (Figure 2 and data not shown). These data demonstrate that CagA can substitute for DOS during the development of photoreceptors.
We predicted that if CagA functions similarly to Gab, then CagA would require the downstream signaling molecule SHP-2/CSW to promote photoreceptor development. As a downstream component of RTK pathways, CSW is required for photoreceptor development [17]. In contrast to wild type larval eye discs, in which thousands of photoreceptors are specified (Figure 3A), in larval eye discs of csw null mutants only a few photoreceptors develop along the morphogenetic furrow (Figure 3B, E) as described previously [33]. The residual photoreceptors in the csw eye discs were mostly R8 cells (data not shown), the only photoreceptor class that does not require RTK signaling for its specification [23]. A significant increase in photoreceptor number could be achieved in the csw mutant eye discs by expression of UAS-CSW with GMR-GAL4 (Figure 3C) or Hsp-GAL4 (data not shown). However, expression of CagA from multiple different transgenic lines using either GMR-GAL4 or Hsp-GAL4 failed to increase the number of photoreceptors in two different csw null mutants (Figure 3D, E, data not shown). These results argue that CagA, like DOS, requires SHP-2/CSW to promote photoreceptor development.
We used a transgenic Drosophila system to test the hypothesis that H. pylori's virulence factor CagA can substitute for the Gab adaptor in RTK signaling pathways. This system is ideal for these studies because RTK signaling pathway components can be genetically manipulated, resulting in interpretable phenotypic consequences for tissue development. First, we have demonstrated that CagA in Drosophila tissue is phosphorylated, that it associates with the cell cortex, and that its expression causes epithelial disorganization as in mammalian tissue culture cells. Second, we have provided genetic evidence that CagA can substitute for Gab by demonstrating that CagA expression restores larval viability and photoreceptor development in mutants lacking the Drosophila Gab, DOS. Our inability to rescue dos mutants to adulthood with CagA expression may be due to differences in RTK activation or to non-overlapping functions of Gab and CagA. Indeed too much CagA expression (using an actin-GAL4 driver) is lethal to flies (unpublished results), which is not the case for ubiquitous expression of DOS [26]. Third, our genetic epistasis analysis with mutants lacking csw has shown that CagA functions through the Drosophila SHP-2 homologue, similar to results from tissue culture experiments [8],[9].
RTK signaling is essential for several fundamental biological processes and erroneous signaling can promote tumor formation [19]. Gain-of-function mutations of SHP-2 have been established as oncogenic in numerous leukemia types as well as other diseases like Noonan's Syndrome [4],[34],[35]. Over-expression of the Gab scaffolding adaptor proteins is associated with the development of several types of cancers, including breast cancer [6],[7] and gastric cancer [36]. The specific cancers that develop as a result of these mutations reflect tissue sensitivities to increased Gab and SHP-2. In the case of H. pylori infection, CagA provides a tissue specific activation of RTK signaling that can precipitate events leading to gastric carcinogenesis [37], as suggested by a recent report of CagA-expressing transgenic mice [38].
Our approach of examining the cellular effects of CagA expression in Drosophila tissue takes advantage of the fact that bacterial proteins frequently target essential, highly conserved cell-signaling pathways. Drosophila has been employed traditionally as a model organism for dissecting signaling pathways in development, but in recent years it has also proven useful in understanding host-pathogen interactions (reviewed in [39],[40]), and in one instance has been used as a heterologous system for expression of the bacterial toxins, anthrax lethal and edema factors [41]. Here we have exploited Drosophila eye development to demonstrate CagA's capacity to function as a RTK adaptor. Future studies using this transgenic Drosophila model will allow us to better understand the cellular and tissue-wide consequences of CagA's disruption of eukaryotic signaling pathways and to identify candidate host factors through which CagA functions.
CagA cDNA was amplified from genomic DNA from H. pylori G27. The CagAEPISA (lacking EPIYA tyrosine phosphorylation motifs) cDNA was amplified from a plasmid provided by Manuel Amieva (originally from Markus Stein [42]). CagAEPISA lacks the tyrosines in the four 5-amino acid motifs, EPIYA, which are phosphorylated by host kinases (point mutations at nucleotide 2684 [A→C] and 2740 [A→C] and a deletion at nucleotide 2878 to 3082). CagA and CagAEPISA were cloned into a modified pUAST vector with an N-terminal hemagglutinin (HA) tag (provided by Chris Q. Doe). Transgenic lines were generated by injecting Qiagen-purified plasmid DNA into y,w1118 embryos. Several independent transformant lines were established for each construct.
Genetic null alleles of csw (cswC114 and csw13-87) and dos (dos1.46 and dos2.46) were obtained from Michael Simon. The UAS-DOS strain was from Thomas Raabe and the UAS-CSW strain (UAS-flgcsw[WTCIM]) from Lizabeth Perkins. UAS-CagA and UAS-CagAEPISA (lacking EPIYA tyrosine phosphorylation motifs) transgenes were expressed in the eye using P{w[+mC] = GAL4-ninaE.GMR}12 (GMR-GAL4, Bloomington Stock Center (BSC) # 1104). P{GAL4-Hsp70.PB}2 (Hsp-GAL4, BSC # 2077) was used for heat-shock inducible expression of transgenes.
Fly heads were fixed overnight at 4°C in 2% gluteraldehyde in 0.1 M sodium cacodylate buffer (pH 7.2) and dehydrated through an ethanol series (30%, 50%, 70%, 80%, 90% 95%, three times in absolute ethanol) at room temperature for 10 minutes in each solution. Samples were critically point dried, sputter coated with gold and viewed using a JEOL 6400 SEM.
Eye imaginal discs were dissected from third instar wandering larvae, fixed for 30 minutes (4% formaldehyde, 0.1 M PIPES (pH 6.9), 0.3% Triton X-100, 2 mM EGTA, 1 mM MgSO4). Discs were washed (0.3% Triton X-100 in phosphate buffered saline, PBS) and blocked for one hour (1% BSA, 0.3% Triton X-100 in PBS). Primary antibodies included rat anti-ELAV 1∶10 (05HB 7E8A10, from Chris Q. Doe), rat anti-HA 1∶100 (Roche) and chicken anti-GFP 1∶2,000 (Chemicon). Secondary antibodies included anti- rat conjugated Rhodamine Red 1∶200 (Jackson ImmunoResearch), anti-rat conjugated AlexaFluor 488 1∶200 (Molecular Probes), anti-mouse conjugated Cy3 1∶200 (Jackson ImmunoResearch) and anti-chicken conjugated Cy2 1∶100 (Jackson ImmunoResearch). Phalloidin conjugated to Tetramethyl Rhodamine Iso-Thiocyanate (TRITC, Sigma Aldrich, 1∶500) was used to stain F-actin. Imaginal discs were visualized using a Nikon TE2000 U with C1 Digital Eclipse confocal microscope.
Wandering third instar larvae were placed at 25°C and approximately 50 hours later the pupal retinas were dissected (50% pupal stage). Retinas were dissected in PBS, fixed for 20 minutes (4% paraformaldehyde in PBS) and washed three times in PBT (0.5% Triton X-100 in PBS). Retinas were blocked at least 15 minutes in 10% normal goat serum in PBT. Antibodies were diluted in the blocking solution. Primary antibodies included mouse MAb 24B10 which stains all photoreceptors and their axons [43] (Developmental Studies Hybridoma Bank, 1∶200), rabbit anti-SAL, which stains R7 and R8 nuclei (also called SPALT, provide by Reinhard Schuh [44], 1∶100), guinea pig anti-SENSELESS, which stains R8 nuclei (proved by Hugo Bellen [45], 1∶1000). Secondary antibodies from Molecular Probes included AlexaFluor 555 conjugated anti-mouse, AlexaFluor 488 conjugated anti-rabbit and AlexaFluor 633 conjugated anti-guinea pig, which were all used at 1∶250. Pupal retinas were visualized using a Leica TCS SP5 confocal microscope.
Fly heads were collected by flash freezing adult flies in liquid nitrogen, shaking the flies in a conical tube, and then separating the heads from the bodies using a mesh sieve. Heads (∼1.5 mL) were homogenized in ice cold lysis buffer (50 mM Hepes, 150 mM NaCl, 1 mM EDTA, 1 mM Na3VO4, 0.5% Triton X-100 and Complete protease inhibitors [Roche]) and then centrifuged at 16,000 G for 5 minutes. Supernatant from the lysate solution (1.5 mL) was added to 50 µL anti-HA Affinity Matrix (Roche) which was incubated overnight at 4°C with gentle agitation. The anti-HA affinity matrix was washed 4 times with ice-cold lysis buffer. CagA was eluted from the matrix by boiling in 100 uL sample loading buffer and separated using manufactures protocols for 7% NuPAGE® Novex Tris-Acetate gels, transferred to polyvinylidene difluoride membranes, blocked overnight at 4°C (200 mM Tris pH 7.5, 100 mM NaCl, 0.1% Tween-20 and 3% BSA (Fisher)), probed using appropriate antibodies and detected using enhanced chemiluminescene (ECL plus, Amersham Biosciences). Mouse anti-HA was used at 1∶1,000 (Babco). Mouse anti-phospho tyrosine was used at 1∶2,000 (Cell Signal Technologies). Horseradish peroxidase-conjugated sheep anti-mouse (Amersham Biosciences) was used at 1∶5,000.
Hsp70-GAL4 balanced over CyO, P{Ubi-GFP} with dos2.42 over TM3, P{Act-GFP}, Ser were crossed to dos1.46/TM3, P{Act-GFP}, Ser (negative control) or UAS-CagA; dos1.46/TM3,P{Act-GFP}, Ser. Progeny were raised at 30°C and pupae were examined for GFP florescence using a Stemi SV 11 Apo Zeiss microscope. The number of non-GFP expressing progeny was scored as a percentage of the total number of pupae that developed per bottle and averaged across bottles of the same genotype. At least 12 bottles were scored per cross with between 150–450 pupae examined per bottle.
The FLP/FRT recombinase system was used to induce somatic clones in the eye [27]. Males y w,ey-FLP 3.5/Y; GMR-GAL4; FRT2, dos1.46/CyO-TM6B were crossed to P{ey-FLP.N}6, ry506 (BSC #5577); P{Ubi-GFP.nls}3L1 P{Ubi-GFP.nls}3L2 P{FRT(whs)}2A (BSC #5825) (negative control) or UAS-DOS; P{Ubi-GFP.nls}3L1 P{Ubi-GFP.nls}3L2 P{FRT(whs)}2A (positive control). Male GMR-GAL4, UAS-CagA; FRT2, dos1.46/CyO-TM6B were crossed to P{ey-FLP.N}6, ry506; P{Ubi-GFP.nls}3L1 P{Ubi-GFP.nls}3L2 P{FRT(whs)}2A. Imaginal eye discs were stained with anti-ELAV and anti-GFP antibodies.
Two genetic null alleles of csw were used to examine if CagA could rescue loss of csw. The cswC114 or csw13-87 alleles were balanced over FM7, P{Act-GFP} with GMR-GAL4 balanced over CyO, P{Ubi-GFP} on the second chromosome. These females were then crossed to males y1w1118, P{Ubi-GFP.nls}X1 P{FRT(whs)}9-2 (BSC # 5832)/Y (negative control), y1w1118, P{Ubi-GFP.nls}X1 P{FRT(whs)}9-2/Y; UAS-CSW (positive control) or y1w1118, P{Ubi-GFP.nls}X1 P{FRT(whs)}9-2/Y; UAS-CagA. Eye imaginal discs were dissected from male larvae. |
10.1371/journal.pbio.3000226 | The architecture of cell differentiation in choanoflagellates and sponge choanocytes | Although collar cells are conserved across animals and their closest relatives, the choanoflagellates, little is known about their ancestry, their subcellular architecture, or how they differentiate. The choanoflagellate Salpingoeca rosetta expresses genes necessary for animal development and can alternate between unicellular and multicellular states, making it a powerful model for investigating the origin of animal multicellularity and mechanisms underlying cell differentiation. To compare the subcellular architecture of solitary collar cells in S. rosetta with that of multicellular ‘rosette’ colonies and collar cells in sponges, we reconstructed entire cells in 3D through transmission electron microscopy on serial ultrathin sections. Structural analysis of our 3D reconstructions revealed important differences between single and colonial choanoflagellate cells, with colonial cells exhibiting a more amoeboid morphology consistent with higher levels of macropinocytotic activity. Comparison of multiple reconstructed rosette colonies highlighted the variable nature of cell sizes, cell–cell contact networks, and colony arrangement. Importantly, we uncovered the presence of elongated cells in some rosette colonies that likely represent a distinct and differentiated cell type, pointing toward spatial cell differentiation. Intercellular bridges within choanoflagellate colonies displayed a variety of morphologies and connected some but not all neighbouring cells. Reconstruction of sponge choanocytes revealed ultrastructural commonalities but also differences in major organelle composition in comparison to choanoflagellates. Together, our comparative reconstructions uncover the architecture of cell differentiation in choanoflagellates and sponge choanocytes and constitute an important step in reconstructing the cell biology of the last common ancestor of animals.
| Choanoflagellates are microscopic aquatic organisms that can alternate between single-celled and multicellular states, and sequencing of their genomes has revealed that choanoflagellates are the closest single-celled relatives of animals. Moreover, choanoflagellates are a form of ‘collar cell’—a cell type crowned by an array of finger-like microvilli and a single, whip-like flagellum. This cell type is also found throughout the animal kingdom; therefore, studying the structure of the choanoflagellate collar cell can shed light on how this cell type and animal multicellularity might have evolved. We used electron microscopy to reconstruct in 3D the total subcellular composition of single-celled and multicellular choanoflagellates as well as the collar cells from a marine sponge, which represents an early-branching animal lineage. We found differences between single-celled and multicellular choanoflagellates in structures associated with cellular energetics, membrane trafficking, and cell morphology. Likewise, we describe a complex system of cell–cell connections associated with multicellular choanoflagellates. Finally, comparison of choanoflagellates and sponge collar cells revealed subcellular differences associated with feeding and cellular energetics. Taken together, this study is an important step forward in reconstructing the biology of the last common ancestor of the animals.
| Collar cells were likely one of the first animal cell types [1–3]. Defined as apicobasally polarised cells crowned with an actin-rich microvillar collar surrounding an apical flagellum [4], they are conserved across almost all animal phyla (Fig 1A) as well as in their closest living relatives, the choanoflagellates [1]. In choanoflagellates and sponges, the undulation of the apical flagellum draws bacteria and other particulate material to the collar, where it can be phagocytosed for food. In many other animals, collar cells function as sensory epidermal cells, nephridial cells, and various inner epithelial cells [1].
Multicellularity evolved multiple times independently in eukaryotes [1,6]. Choanoflagellates are uniquely suited for investigating characteristics of the last common multicellular ancestor of animals and the origin of animal-specific innovations. Several independent phylogenomic analyses [7–9] have placed them as the closest branching lineage to the animals. It is thought that the transition from a free-swimming facultatively unicellular collar cell to one in an obligately multicellular animal condition emerged along the animal stem lineage [2]. While it has been hypothesised that the common ancestor of animals may have exhibited a complex, polymorphic life cycle [10,11], parsimony suggests that at least one of these life stages would have possessed choanoflagellate-like collar cells [1]. Investigation of the choanoflagellate cell plan therefore has the potential to shed light on the evolution of one of the most ancient animal cell types.
The colony-forming choanoflagellate S. rosetta [12] has emerged as a promising model organism to investigate the properties of the progenitor of the animals [13]. This species exhibits a complex life cycle, transitioning through both single and colonial collar cell types [12,14] (Fig 1B). The development of rosette colonies can be induced by rosette-inducing factor (RIF) (Fig 1B), which is a sulfonolipid from the bacterium Algoriphagus machipongonensis [15]. Most importantly, choanoflagellate colonies form by cell division, and cells within rosette colonies are held together by cytoplasmic bridges, filopodia, and extracellular matrix (ECM) [12]. Cell types of S. rosetta have been previously well investigated using molecular tools [16–18], which have revealed that choanoflagellates possess a suite of genes essential for animal multicellularity and development.
However, our structural understanding of how choanoflagellate cells like S. rosetta organise themselves into colonies—and how these compare to early-branching animal collar cells—remains unquantified relative to molecular investigations. Given the importance of cell differentiation for the origin of animals, we hypothesised that choanoflagellate colonial cells would not simply represent a cluster of single cells but would be morphologically differentiated from single cells. Our previous studies show that the proteins Flotillin and Homer colocalise in the nucleus of all single choanoflagellate cells, but not in all colonial cells providing preliminary evidence of cell differentiation within choanoflagellate rosette colonies [16]. In contrast, the nearly indistinguishable transcriptomes of single cells and colonies [17] speak against cell differentiation.
In this study, we used serial ultrathin transmission electron microsocopy (ssTEM) sectioning to reconstruct the microanatomy of unicellular and colonial S. rosetta cells to identify structural differences between collar cells in a single versus a multicellular choanoflagellate condition. To place our choanoflagellate reconstructions into the context of collar cells from an early-branching animal phylum, we reconstructed a section of a sponge choanocyte chamber from the homoscleromorph sponge Oscarella carmela [19] (Box 1). Our characterisation of the microanatomy of choanoflagellates and sponge choanocytes sheds light on collar cell differentiation, has implications for the origin and evolution of animal cell types, and is an important step in reconstructing the putative biology of the last common ancestor of the animals.
Three randomly selected single cells and three randomly selected colonial cells from a single colony were chosen for the reconstruction of entire choanoflagellate cells and subcellular structures (Fig 1, S1–S3 Figs, S1–S6 Movies). Both single and colonial S. rosetta cells exhibited a prominent, central nucleus enveloped by a mitochondrial reticulum and basal food vacuoles—as well as intracellular glycogen reserves—consistent with the coarse choanoflagellate cellular architecture reported in previous studies [20,21] (reviewed in [13,22]) (Fig 1, S1–S3 Figs, S1–S6 Movies). However, with the increased resolution of electron microscopy, we detected three morphologically distinct populations of intracellular vesicles with distinct subcellular localisations (Fig 1G, S1I–S1L Fig): 1) large vesicles (extremely electron-lucent, 226 ± 53 nm in diameter) (S1J, S1J’, and S1J” Fig); 2) Golgi-associated vesicles (electron-dense inclusions, 50 ± 10 nm in diameter) (S1I, S1I’, and S1I” Fig); and 3) apical vesicles (electron-lucent, 103 ± 21 nm in diameter) (S1K, S1K’ and S1K” Fig). Extracellular vesicles were also observed to be associated with two of the single cells (electron-lucent, 173 ± 36 nm in diameter) and appeared to bud from the microvillar membrane (S1L, S1L’ and S1L” Fig). Choanoflagellate cells subjected to fluorescent labelling were congruent with 3D ssTEM reconstructions in terms of organelle localisation (Fig 1B and 1C), providing evidence that the 3D models presented herein are biologically representative.
Our 3D ssTEM reconstructions allowed for detailed volumetric and numerical comparisons among single and colonial S. rosetta cells (Fig 2, S2 Fig, S1 and S2 Tables). Overall, the general deposition of major organelles was unchanged in both cell types (Fig 1E–1L, Fig 2A and 2B, S2A–S2C Fig). In addition, single and colonial cells devote a similar proportion of cell volume to most of their major organelles (nucleus: single cells 12.92% ± 0.58% versus colonial cells 11.56% ± 0.27%; nucleolus: 1.85% ± 0.33% versus 2.2% ± 0.22%; mitochondria: 5.08% ± 1.14% versus 6.63% ± 0.42%; food vacuoles: 9.22% ± 2.75% versus 6.85% ± 0.87%; and glycogen storage: 8.71% ± 2.36% versus 7.50% ± 1.12%) (Fig 2, S2 Fig, S1 and S2 Tables).
We did, however, uncover some interesting ultrastructural differences between single and colonial cells (Fig 2C). Colonial cells devoted a higher proportion of cell volume to endoplasmic reticulum (ER) (single: 3.27% ± 0.35% versus colonial: 6.86% ± 0.39%). This contrast was coupled to a differential ER morphology across cell types. The ER of colonial cells frequently displayed wide, flat sheets (Fig 3E), which were not observed in the reconstructed single cells. Single cells exhibited a higher number of Golgi-associated vesicles (single: 166.3 ± 32.7 versus colonial: 72.3 ± 26.5) and individual mitochondria than colonial cells (single: 25.3 ± 5.8 versus colonial: 4.3 ± 4.2) (Fig 2C, S2 Table) despite lacking volumetric differences between cell types.
Finally, we found that colonial cells are characterised by a more amoeboid morphology than single cells (Fig 3A). Colonial cells exhibited a higher relative proportion of endocytotic vacuoles by volume (single: 0.07 ± 0.07 versus colonial: 0.32 ± 0.12)—a phenomenon coupled to a higher overall number of endocytotic vacuoles (single: 1 ± 1 versus colonial: 5 ± 2) and pseudopodial projections per cell (single: 1 ± 1 versus colonial 8 ± 2) (Fig 2C, S1 and S2 Tables). Many of the pseudopodial projections and endocytotic vacuoles bore the morphology of lamellipod ruffles and macropinosomes (Fig 3A), suggesting that colonial cells are typified by high macropinocytotic activity.
While high-magnification 3D ssTEM enabled the high-resolution reconstruction of individual colonial cells, their context and interactions with neighbouring cells were lost. To address this, we reconstructed the subcellular structures of a seven-cell rosette colony (complete rosette, RC1) from 80-nm sections taken at lower magnification (Fig 3A–3D, S7 Movie) as well as the gross morphology of four larger rosettes (RC2–5) from 150-nm sections to provide a more representative survey (Fig 3E–3P). We found that individual cells in rosette colonies vary widely in volume (Fig 3M and 3N), although no pattern was detected in the volumetric cellular arrangement along the rosette z-axis (Fig 3M). In addition, mean cell size was comparable among different rosettes, including those that contained different numbers of cells (S4B Fig). However, we did find a positive correlation between cell number and the number of intercellular bridges per cell across rosette colonies (S4B Fig).
Importantly, we uncovered the presence of unusually shaped cells in two of the five S. rosetta rosette colonies (carrot-shaped cell 5 in RC3 and chili-shaped cell 5 in RC4, both labelled orange with an asterisk) (Fig 3M). These unusual cells were both found at the same location along the rosette z-axis, exhibited an elongated morphology distinct from other colonial cells (Fig 3O and 3P and S8 and S9 Movies), and were small in volume. Cells 5 from RC3 and RC4 were 9.87 μm3 and 13.35 μm3, respectively (Fig 3N)—the mean volume of the cells in RC3 and RC4 was 27.38 μm3 and 27.25 μm3, respectively (Fig 3N). While each of these unusual cells possessed a flagellum, a collar, connections to neighbouring cells via intercellular bridges, and had a similar proportion of cell volume dedicated to most of their major organelles as observed in other colonial cells, these cells devoted a larger volumetric percentage of the cell body to the nucleus (29.8% and 30.78%, respectively, versus the mean colonial proportion of 13.76% ± 0.49%).
Our 3D ssTEM reconstructions of rosette colonies also revealed the distribution of intercellular bridges and the connections formed between individual cells (Fig 3M). We found intercellular bridges in all analysed rosette colonies (RC1–5), totalling 36 bridges. There was no detectable pattern regarding bridge networking across rosette colonies. Bridges were distributed from the cell equator to either of the poles along the cellular z-axis, and the average bridge was 0.75 ± 0.38 μm in length (S4C Fig). Prior studies [12,17] of S. rosetta bridges suggested that bridges are typically short (0.15 μm), connecting two adjacent cells and containing parallel plates of electron-dense material. In contrast, the bridges detected in this study exhibited striking morphological diversity (Fig 3M, 3Q–3U), with lengths ranging from 0.21–1.72 μm. The majority of bridges consisted of a protracted cytoplasmic connection between two cells, and in many cases, the septum was localised asymmetrically along the bridge (S4C Fig). Most surprisingly, some bridges were not connected to any neighbouring cells at all, but rather the septum was situated on the end of a thin, elongated cellular protrusion (Fig 3S). In addition, we observed asymmetric bridge width and degraded electron-dense structures proximal to bridge remnants being incorporated into the cell body of a contiguous cell (Fig 3T and 3U). These data suggest that intercellular bridges could be disconnected from neighbouring cells and that the electron-dense septum may be inherited.
Both choanoflagellates and sponge collar cells influence local hydrodynamics by beating their single flagellum to draw in bacteria that are captured by the apical collar complex [23], however sponge choanocytes are part of an obligately multicellular organism (Fig 4A). Sponge choanocytes therefore provide an excellent representative of an early-branching animal collar cell against which to compare choanoflagellate cell architectures. Our 3D ssTEM reconstructions allowed for the reconstruction of five choanocytes and for the volumetric and numerical comparison of choanocyte and choanoflagellate subcellular structures (Fig 4B–4E, S5 and S6 Figs, S10 Movie). We detected little ultrastructural variability between the five choanocytes (S5 Fig, S3 and S4 Tables). All five cells exhibited a prominent basal nucleus, small and unreticulated mitochondria, food vacuoles scattered around the entire cell, and an apical Golgi apparatus (Fig 4B–4D, S5 and S6 Figs)—consistent with the coarse choanocyte cellular architecture reported in previous studies [23,24] (reviewed in [1,25]).
Furthermore, our data showed many ultrastructural commonalities between sponge choanocytes and choanoflagellates. For example, the number of microvilli that surround the apical flagellum in single and colonial choanoflagellates is comparable to the number of microvilli in sponge choanocytes (single: 32 ± 2 versus colonial: 35.3 ± 4.9 versus choanocytes: 30.6 ± 4.1) (S6A Fig). We also found that the number of food vacuoles and the number and volumetric proportion of the Golgi apparatus are similar in all three cell types (S6A Fig). Although sponge choanocytes did not appear to exhibit the same macropinocytotic activity as colonial choanoflagellates throughout the cell (some micropinocytotic inclusions are present toward the cell apex [S6D and S6E Fig]), basal sections of choanocytes were heavily amoeboid (S6B and S6C Fig). These amoeboid protrusions may not only be for mechanical anchorage into the mesohyl but may play a role in phagocytosis, as we observed bacteria in the mesohyl to be engulfed by basal pseudopodia (S6F and S6G Fig). Thus, both choanocytes and colonial choanoflagellates are typified by high-amoeboid cell activity.
We also observed some ultrastructural differences between choanocytes and choanoflagellates. In contrast with cells from choanoflagellate rosettes, sponge choanocytes lack filopodia and intercellular bridges. Choanocytes also do not possess glycogen reserves and devote significantly less of their cell volume (9.25% ± 0.39%) than choanoflagellates (single: 12.92% ± 0.58% and colonial: 11.56% ± 0.27%) to the nucleus and less to mitochondria (2.5% ± 0.3% versus single: 5.08% ± 1.14% and colonial: 6.63% ± 0.42%) (S6A Fig). However, choanocytes devote significantly more of their volume to food vacuoles (20.7% ± 1.01%) than choanoflagellates (single: 9.22% ± 2.75% and colonial: 6.85% ± 0.87%) (Fig 4E). High-resolution reconstructions of the choanocyte and choanoflagellate apical pole (Fig 4F and 4G, S11 and S12 Movies) showed differences in terms of vesicle type and localisation, Golgi positioning, and collar arrangement (conical in choanoflagellates while cylindrical in choanocytes, as previously noted [23]). The flagellar basal body has previously been meticulously characterised in both choanocytes and choanoflagellates, and some differences have been reported between the two by other authors [26–31]. These findings are reiterated by our reconstructions and observations (Fig 4F and 4G).
Our comparison between single and colonial choanoflagellate cells provides new insights into ultrastructural commonalities and differences associated with the conversion of solitary to colonial cells. Our study also revealed morphologically distinct populations of vesicles in choanoflagellates. Golgi-associated vesicles (S1I, S1I’ and S1I” Fig), due to their tight association with the Golgi apparatus, likely represent standard Golgi-trafficking vesicles carrying cargo between the different Golgi cisternae [4]. Apical vesicles (S1K, S1K’ and S1K” Fig), due to their close proximity to the plasma membrane, are probably secretory vesicles involved in exocytosis of ECM material [4], which bud off the trans-Golgi network and fuse with the plasma membrane. The localisation of neurosecretory soluble N-ethylmaleimide-sensitive-factor attachment receptor (SNARE) proteins to the apical pole of the choanoflagellate Monosiga brevicollis supports this hypothesis [32,33]. The large vesicles (S1J, S1J’ and S1J” Fig) may not represent true vesicles but rather nascent food vacuoles, congruent with what is already known about phagocytosis in choanoflagellates [34]. The finding of extracellular vesicles (S1L, S1L’ and S1L” Fig) associated with the S. rosetta microvillar collar is, to our knowledge, a novel finding in choanoflagellates. Extracellular vesicles in animal cells play diverse roles in cell physiology, such as antigen presentation (reviewed by [35]), morphogenesis [36,37], and disseminating pathogenic proteins [38,39]. Association of extracellular vesicles with apical microvilli, as reported here in S. rosetta, bears a striking similarity to animal enterocytes [40]. Extracellular vesicles released from enterocyte microvilli are enriched in intestinal alkaline phosphatase and are thought to be antibacterial in nature [40,41]. It is therefore conceivable that choanoflagellate extracellular vesicles too contain hydrolytic enzymes to catalyse the degradation of bacteria in the collar—the site of prey capture [34].
Moreover, our findings reveal that colonial cells likely represent distinct and differentiated cell types relative to single cells. The ultrastructural differences between single and colonial cells in ER, Golgi-associated vesicles, and the amoeboid and pinocytotic nature of colonial cells hint toward a demand on endomembrane reorganisation and intracellular trafficking (one possibility could be the increased uptake of RIF-1 to keep colonies intact). ER and mitochondrial morphology change dynamically, and stark changes have been observed in other eukaryotic cells due to changes in cell cycle [42] and cytoskeletal activity [43,44]. Mitochondria and the ER also show an intimate association [45], and the contrast in the number of individual mitochondria in different choanoflagellate cells was particularly striking. The reduced numbers of mitochondria in colonial cells indicate a lower energy consumption than in single cells. The demand on energy in single cells, which have to swim to find new food sources, may well be higher than in colonial cells, which do not have to swim, as they tumble to stay in food-rich environments. In animal cell types, fusion/fission dynamics have been previously associated with cellular stress [46] and substrate availability [47], but it is of most interest for choanoflagellates in the context of aerobic metabolism. For example, the fresh water choanoflagellate Desmarella moniliformis exhibits a shift in mitochondrial profile prior to encystment and metabolic dormancy [48], and choanoflagellates have been uncovered from hypoxic waters [49]. The role of oxygen in the origin and evolution of the animals has long been discussed [50] and is currently met with controversy [51,52]. Coupled to a previous report of positive aerotaxis in S. rosetta rosette colonies [53], our finding emphasises the need to better understand variation in aerobic metabolism between single and colonial choanoflagellates.
Particularly surprising was the finding of extensive macropinocytotic activity in colonial cells. Macropinocytosis—defined by the formation of phase-lucent vacuoles >0.2 μm in diameter from wave-like, plasma membrane ruffles [54]—is conserved from the Amoebozoa [55] to animal cell types [56]. It is parsimonious to infer that the macropinocytotic activity of S. rosetta colonial cells represents a trophic adaptation considering that previous biophysical studies have reported more favourable feeding hydrodynamics in rosette colonies [57], although a more recent study does not confirm these findings [53]. Even in macropinosomes with no observable cargo, dissolved proteins [58] and ATP [56] from extracellular fluid have been previously reported to be metabolically exploited by animal macropinocytotic cell types. However, this nonselectivity, coupled with the large volume of engulfed fluid, makes macropinocytosis an efficient cellular process to sample the extracellular milieu. It is therefore tempting to speculate that macropinocytosis may also play a role in detecting environmental chemical signals in colonial S. rosetta cells.
The reconstruction of multiple choanoflagellate rosette colonies reveals the asymmetric and disconnected morphology of intercellular bridges and provides important clues to choanoflagellate colony formation and potentially the evolution of animal multicellularity. Bridges displaying electron-dense septa reminiscent of those found in S. rosetta have been previously identified in other colony-forming choanoflagellate species [59], and it has been hypothesised that these structures represent stable channels for intercellular communication [17]. Our data suggest that bridges can be disconnected and that the electron-dense septum may be asymmetrically inherited. In this way, choanoflagellate bridges may resemble the mitotic midbody in animal cells [60]. Relatively recent molecular studies have suggested that inheritance of the mitotic midbody may be associated with diverse developmental roles [61–63] in the recipient cell. While homology between the electron-dense septum in choanoflagellates and metazoan midbodies cannot be determined from these data, asymmetric inheritance of this structure could play an analogous role in the development of colonial cells. It may still be that S. rosetta bridges play a role in cell–cell communication, albeit transiently. However, the exit of colonial cells from the rosette (as previously reported [12]) must involve bridge disconnection, and a proper understanding of the fate of the septum could augment our understanding of choanoflagellate cell differentiation and destiny in colony development.
The discovery of the highly derived ‘carrot’ and ‘chili’ cell types was not expected (Fig 3O and 3P). The morphological similarity, the enlarged nuclear volume, and the position on the colony z-axis shared between the two cells suggests that they represent a distinct S. rosetta cell type in rosette colonies. These data hint that cell differentiation within colonies may be more complex than previously realised and provide potential evidence for division of labour in choanoflagellate colonies. Previously proposed models of animal evolution via a colonial intermediate place emphasis on cellular differentiation and division of labour as key innovations toward obligate animal multicellularity [2,64]. We cannot exclude the possibility that the carrot- and chili-shaped cells are results of cells preparing to divide or cells shortly after cell division, but we think this is highly unlikely. There is no precedence for it in the literature, and our own live-cell observations of choanoflagellate cell divisions do not support this either. ‘Chili’ and ‘carrot’ cells in choanoflagellate colonies might be either caused by programmatic cellular differentiation or stochastic developmental noise. Cells in rosettes, which have different numbers of intracellular bridges and adjacent cells, may sense (through macropinocytosis) and respond (through apical vesicles) to the local environment of cells, thus making stochastic, cell-autonomous differentiation more likely than deterministic cell differentiation. We cannot rule out either of these causes at the moment but believe further research into the cell biology of these putative novel cell types is desperately needed, and single-cell transcriptomic data and live-cell imaging of choanoflagellate rosette development could shed more light on cell type variation in rosette colonies.
The 3D cellular architecture of sponge choanocytes allowed for the detailed comparison of their architecture with choanoflagellates. Although we observed many ultrastructural similarities between choanoflagellates and sponge choanocytes, there were noteworthy differences in terms of food vacuole, mitochondria, and glycogen composition. This is likely due to the different physiological niches occupied by the two cell types. As free-living protists, choanoflagellates must maintain energetically costly motility and may devote a larger proportion of their cytoplasm to mitochondrial reticula and glycogen stores at the expense of food vacuoles. Choanocytes are but one cell type in a sessile multicellular organism that exhibits cellular differentiation and strict division of labour. As such, choanocytes represent a specialised feeding cell type (which devotes a significantly higher amount of its cell volume to food vacuoles) that operates a vastly different physiology to the independent ancestral collar cell. These ultrastructural differences are good identifying features marking the differential biology of generalist versus specialist collar cells.
Recently, morphological and functional differences between choanocytes and choanoflagellates have been taken as evidence that collar cells have evolved by convergent evolution for feeding on bacteria [23,65]. Evolution is expected to lead to differences among homologous structures (an excellent example are vertebrate limbs that are all very different—some are wings, some are legs, some are fins but are still homologous) and thus it is not surprising to observe (ultra)structural differences between choanoflagellates and choanocytes.
While we recognise the limitations of our findings due to the morphological descriptive nature of this study and the small sample size, the comparative 3D reconstruction of collar cells from two different phyla, choanoflagellates and sponges, allowed for an unbiased view of their cellular architecture and for the reconstruction of key properties of the enigmatic ancestral collar cell. Our data reveal distinct ultrastructural features in single and colonial choanoflagellates and demonstrate that cells within rosette colonies vary significantly in their cell size and shape. The newly identified ‘carrot’ and ‘chili’ cells reveal that cells within choanoflagellate colonies do not simply consist of an assemblage of equivalent single cells, but some may represent a distinctly differentiated cell type displaying ultrastructural modifications. Likewise, our data suggest that sponge choanocytes are not simply an incremental variation on the choanoflagellate cell plan but are specialised feeding cells, as indicated by their high volumetric proportion of food vacuoles. Together, our data show a remarkable variety of collar cell architecture and suggest cell type differentiation may have been present in the stem lineage leading to the animals.
Colony-free S. rosetta cultures (ATCC 50818) were grown with coisolated prey bacteria in 0.22 μm filtered choanoflagellate growth medium [66] diluted at a ratio of 1:4 with autoclaved seawater. Cultures were maintained at 18°C and split 1.5:10 once a week. Colony-enriched S. rosetta cultures (PX1) were likewise maintained but monoxenically cultured with the prey bacterium A. machipongonensis [67] to induce rosette formation.
To support the annotation of organelles from ssTEM sections, the microanatomy of S. rosetta cells was chemically characterised by fluorescent vital staining. Cells were pelleted by gentle centrifugation (500x g for 10 min at 4°C) in a Heraeus Megafuge 40R (ThermoFisher Scientific) and resuspended in a small volume of culture medium. Concentrated cell suspension (500 μl) was applied to glass-bottom dishes, coated with poly-L-lysine solution (P8920, Sigma-Aldrich), and left for 10–30 min until cells were sufficiently adhered. PX1 cultures were concentrated into 100 μl of culture medium to promote the adherence of rosette colonies.
Adhered cells were incubated in 500 μl of fluorescent vital dye diluted in 0.22 μm filtered seawater. Cells were incubated with 4.9 μM Hoechst 33342 Dye for 30 min (to label nuclei), 1 μM LysoTracker Yellow HCK-123 for 1.5 h (to label food vacuoles), and 250 nM MitoTracker Red CM-H2Xros for 30 min (to label mitochondria). All vital dyes were from ThermoFisher Scientific (H3570, L12491, T35356, and M7513, respectively). Fluorescent-DIC microscopy was conducted under a 100x oil-immersion objective lens using a Leica DMi8 epifluorescent microscope (Leica, Germany). Vital dyes were viewed by excitation at 395 nm and emission at 435–485 nm (Hoechst 33342 Dye), 470 nm and emission at 500–550 nm (LysoTracker Yellow HCK-123 and FM 1–43 Dye), and 575 nm and 575–615 nm (MitoTracker Red CM-H2Xros). Micrographs were recorded with an ORCA-Flash4.0 digital camera (Hamamatsu Photonics, Japan). All cells were imaged live. No-dye controls using only the dye solvent dimethyl sulfoxide (DMSO) (D4540, Sigma-Aldrich) were run for each wavelength to identify and control for levels of background fluorescence. Chemical fixation during vital staining and TEM sectioning was avoided where possible in this study to reduce fixation artefacts.
To visualise cell bodies, flagella, filopodia, and collar-adherent cells were fixed for 5 min with 1 ml 6% acetone and for 15 min with 1 ml 4% formaldehyde. Acetone and formaldehyde were diluted in artificial seawater, pH 8.0. Cells were washed gently four times with 1 ml washing buffer (100 mM PIPES at pH 6.9, 1 mM EGTA, and 0.1 mM MgSO4) and incubated for 30 min in 1 ml blocking buffer (washing buffer with 1% BSA and 0.3% Triton X-100). Cells were incubated with primary antibodies against tubulin (E7, 1:400; Developmental Studies Hybridoma Bank), diluted in 0.15 ml blocking buffer for 1 h, washed four times with 1 ml of blocking buffer, and incubated for 1 h in the dark with fluorescent secondary antibodies (1:100 in blocking buffer, Alexa Fluor 488 goat anti-mouse). Coverslips were washed three times with washing buffer, incubated with Alexa Fluor 568 Phalloidin for 15 min, and washed again three times with washing buffer. Coverslips were mounted onto slides with Fluorescent Mounting Media (4 ml; Prolong Gold Antifade with DAPI, Invitrogen). Images were taken with a 100x oil-immersion objective on a Leica DMI6000 B inverted compound microscope and Leica DFC350 FX camera. Images presented as z-stack maximum intensity projections.
ssTEM sections were imported as z-stacks into the Fiji [71] plugin TrakEM2 [72] and automatically aligned using default parameters, except for increasing steps per octave scale to 5 and reducing maximal alignment error to 50 px. Alignments were manually curated and adjusted if deemed unsatisfactory. Organelles and subcellular compartments were manually segmented and 3D reconstructed by automatically merging traced features along the z-axis. Meshes were then preliminarily smoothed in TrakEM2 and exported into the open-source 3D software Blender 2.77 [73]. Heavy smoothing of the cell body in TrakEM2 sacrifices fine structures associated with cellular projections or does not remove all distinct z-layers, which exist as reconstruction artefacts. Therefore, cell bodies were manually smoothed using the F Smooth Sculpt Tool in Blender of final distinct z-layers for presentation purposes only (S3 Fig). All organelles were subjected to the same smoothing parameters across individual cells. All analysis was conducted using unsmoothed, unprocessed meshes. Organelle volumes were automatically quantified by the TrakEM2 software and enumerated in Blender 2.77 by separating meshes in their total loose parts.
The microvillar collar and flagellum were excluded from volumetric analysis, as their total, representative length could not be imaged at this magnification. Cytosolic volume was calculated by subtracting total organelle volume from cell body volume and is inclusive of cytosol, ribosomes, and unresolved smaller structures excluded from 3D reconstruction. Endocytotic vacuoles were distinguished from food vacuoles by connection to the extracellular medium in ssTEMs or by localisation to a cell protrusion. Cells in rosette colonies are numbered in order of their appearance along the image stack z-axis. Rosette colony diameters were calculated by measuring the largest distance of the z-axis midsection. Bridge length was measured in one dimension along the bridge midsection. Mean vesicle diameters were calculated from 20 measurements (or as many as possible if the vesicle type was rare) from single cells.
Univariate differences in the volume and number of subcellular structures between the two cell types were evaluated using two-sample t tests. Shapiro–Wilk and Levene’s tests were used to assess normality and homogeneity of variance, respectively. Statistical comparisons were conducted using data scaled against total cell volume. Correlations between colony cell number, cell volume, and bridges per cell were assessed using Pearson correlation tests. All statistical analyses were conducted using R v 3.3.1 [74] implemented in RStudio v 0.99.903 [75].
|
10.1371/journal.pgen.1003465 | Genetic Requirements for Signaling from an Autoactive Plant NB-LRR Intracellular Innate Immune Receptor | Plants react to pathogen attack via recognition of, and response to, pathogen-specific molecules at the cell surface and inside the cell. Pathogen effectors (virulence factors) are monitored by intracellular nucleotide-binding leucine-rich repeat (NB-LRR) sensor proteins in plants and mammals. Here, we study the genetic requirements for defense responses of an autoactive mutant of ADR1-L2, an Arabidopsis coiled-coil (CC)-NB-LRR protein. ADR1-L2 functions upstream of salicylic acid (SA) accumulation in several defense contexts, and it can act in this context as a “helper” to transduce specific microbial activation signals from “sensor” NB-LRRs. This helper activity does not require an intact P-loop. ADR1-L2 and another of two closely related members of this small NB-LRR family are also required for propagation of unregulated runaway cell death (rcd) in an lsd1 mutant. We demonstrate here that, in this particular context, ADR1-L2 function is P-loop dependent. We generated an autoactive missense mutation, ADR1-L2D484V, in a small homology motif termed MHD. Expression of ADR1-L2D848V leads to dwarfed plants that exhibit increased disease resistance and constitutively high SA levels. The morphological phenotype also requires an intact P-loop, suggesting that these ADR1-L2D484V phenotypes reflect canonical activation of this NB-LRR protein. We used ADR1-L2D484V to define genetic requirements for signaling. Signaling from ADR1-L2D484V does not require NADPH oxidase and is negatively regulated by EDS1 and AtMC1. Transcriptional regulation of ADR1-L2D484V is correlated with its phenotypic outputs; these outputs are both SA–dependent and –independent. The genetic requirements for ADR1-L2D484V activity resemble those that regulate an SA–gradient-dependent signal amplification of defense and cell death signaling initially observed in the absence of LSD1. Importantly, ADR1-L2D484V autoactivation signaling is controlled by both EDS1 and SA in separable, but linked pathways. These data allows us to propose a genetic model that provides insight into an SA–dependent feedback regulation loop, which, surprisingly, includes ADR1-L2.
| Plants possess an active, inducible disease resistance system, and induction of these responses depends in part on plant resistance proteins. Present understanding of these resistance proteins likens them to molecular switches that bind nucleotides to activate disease resistance responses. Previously it was shown that Activated Disease Resistance 1-like 2 (ADR1-L2), a plant disease resistance protein, is important in the immune response, but can function in the contexts analysed independently of what is currently considered the canonical nucleotide switch activation. Here, we show that, in addition to these previously reported functions, ADR1-L2 also works as a typical, activated disease resistance protein. We use an autoactive mutant form of the protein and show that it promotes disease resistance. We find that ADR1-L2 works in an EDS1-dependent feedback loop with salicylic acid, a hormone known to be essential for plant disease resistance. This work allows us to broaden the understanding of how plant disease resistance proteins function to generate defense against pathogens.
| Plants encounter a wide variety of pathogens. To defend against infection, plants rely on their organ surfaces as pre-formed barriers to infection. Plants have also evolved an active, two-layered immune system [1]. The first branch utilizes transmembrane receptors (PRRs, or pattern recognition receptors) which detect microbe-associated molecular patterns (MAMPs) of various pathogens [2]. MAMP detection elicits a rapid, relatively low-amplitude host transcriptional response resulting in MAMP-triggered immunity (MTI) which is sufficient to halt growth of many microbes [1], [3]. Successful pathogens can suppress or delay MTI via delivery of effector molecules into host cells. Effectors are typically virulence proteins [4]. Gram-negative bacterial pathogens deliver effectors via injection into the plant cell by the Type III Secretion System (TTSS). Plants respond to effectors with the second tier of recognition, which is dependent on highly polymorphic intracellular disease resistance (R) proteins of the NB-LRR family. NB-LRRs are specifically activated by the presence and/or action of effectors to trigger robust defense responses termed Effector-Triggered Immunity (ETI), which can include localized hypersensitive cell death [1].
NB-LRR proteins are members of the signal transduction ATPases with numerous domains (STAND) superfamily, which also includes animal innate immune sensors of the nucleotide-binding domain and leucine-rich repeat-containing (NLR) class [5], [6]. STAND proteins are ATPases that function as molecular switches: in the “off” position they bind ADP, and in the “on” position they bind ATP, activating nucleotide hydrolysis and triggering downstream defense responses. This model is proposed for plant NB-LRRs, though there is very little experimental data pertinent to it [7]. Two essential, conserved homology regions necessary for proper plant NB-LRR activity are the P-loop (Walker-A) and the thus far plant-specific ‘MHD motif’ located in the ARC2 sub-domain of the extended NB-ARC domain. Mutations in the P-loop typically lead to loss of function [8], [9]. Conversely, mutation of the Asp (D) in the MHD motif often leads to autoactivity of the NB-LRR protein [10]–[15], resulting in either lethality or a severely dwarfed morphology. These pleiotropic phenotypes are thought to be the consequence of ectopic accumulation of SA, a key defense hormone whose synthesis from chorismate is controlled by the isochorismate synthase gene (ICS1/SID2) [16], and consequent defense activation [11], [13], [15]. Additionally, several NB-LRRs, in both plants and animals, work in pairs: in these cases, one can function as an effector-specific ‘sensor’, and the other as a ‘helper’ protein. This may allow or drive the formation of higher-order protein complexes necessary for properly regulated defense activation [17]–[20].
ADR1-L2 (Activated Disease Resistance 1-like 2) is one of a small family of NB-LRR proteins that includes ADR1 and ADR1-L1 [21]. We recently demonstrated that ADR1-L2 functions downstream of the production of reactive oxygen intermediates (ROI), and upstream of SA accumulation, in basal defense (defined as the response that limits the growth and proliferation of genetically virulent pathogens). ADR1-L2 also functions in MAMP-triggered SA accumulation, and as a ‘helper’ protein during some, but not all ETI responses driven by effector-mediated activation of specific sensor NB-LRR proteins [22].
Surprisingly, none of the ADR1-L2 functions above required an intact P-loop [22]. In addition to these ‘non-canonical’ activities, we suggested that ADR1-L2 might have as yet undefined P-loop dependent, ‘canonical’ functions that, in the absence of the specific effector required for activation, are difficult to define. ADR1-L2 would not be the first NB-LRR protein to have multiple, independent functions. The mouse NLR protein NLRC4 has two separate functions as a ‘helper’ protein in the recognition of both the MAMP flagellin and PrgJ, a component of the Salmonella TTSS. These activities are downstream of the activation of two different sensor NLRs: NAIP5 is necessary for flagellin perception, and NAIP2 is required for PrgJ recognition [17], [20]. Importantly, NLRC4 ‘helper’ activity is also P-loop independent [17], [20].
Canonical, effector-driven NB-LRR activation typically leads to an NADPH oxidase-dependent ROI burst [23]. The adr1 family triple mutant (adr1 adr1-L1 adr1-L2) exhibited normal ROI production after successful pathogen recognition [22]. Thus, the ADR1-L2 helper function noted above is downstream or independent of this oxidative burst. However, adr1 triple mutants failed to accumulate wild-type levels of SA in this context [22]. Another protein that functions downstream of the effector-driven oxidative burst and both regulates and responds to SA accumulation is Lesion Simulating Disease resistance 1 (LSD1) [23], [24]. Loss of LSD1 leads to improper regulation of runaway cell death, or rcd [24] that eventually engulfs the affected leaf. The Arabidopsis NADPH oxidase AtRbohD, which is required for effector-driven oxidative burst, is not required for lsd1-mediated cell death [23]. On the other hand, lsd1 rcd is both induced by, and requires, SA [24], [25]. lsd1 rcd is also regulated by Enhanced Disease Susceptibility 1 (EDS1) and a type I metacaspase, AtMC1; eds1 lsd1 and atmc1 lsd1 plants do not exhibit rcd [26], [27]. EDS1 is a defense response regulator, required for both basal defense and Toll/interleukin-1 (TIR)-NB-LRR mediated ETI [28]. EDS1 and SA act in a regulatory feedback loop, with SA up-regulating EDS1 expression and EDS1 functioning as a potentiator of SA-mediated signaling [29], [30]. AtMC1 is a positive regulator of ETI-mediated cell death [27].
To define the genetic requirements of putative canonical functions of ADR1-L2 in the absence of an effector known to activate it, we created an autoactive MHD mutant, ADR1-L2D484V. This allele displayed the dwarfed morphology that is the hallmark of MHD mutants [11], [13], [15]. We demonstrate that this autoactivity is P-loop dependent, downstream of AtRbohD-mediated ROI production, partially dependent on SA synthesis, and negatively regulated by EDS1 and AtMC1. We then present and validate a model for the interaction of EDS1, LSD1, and ADR1-L2, showing that these proteins function in both SA-dependent and SA-independent feedback regulatory loops that are interconnected.
ADR1-L2 is a CC-NB-LRR that is a positive regulator of lsd1 rcd [22]. It is part of a small family of NB-LRRs that includes ADR1 and ADR1-L1 [21], [22]. We generated adr1 lsd1-2 and adr1-L1 lsd1-2 double mutants and sprayed them with the SA analog benzothiadiazole (BTH) [31] to test whether adr1 and adr1-L1 also suppress the initiation and propagation of lsd1 rcd. Col-0 wild-type plants were unaffected by BTH treatment, whereas lsd1-2 plants sprayed with BTH showed typical rcd [24]. As reported, the adr1-L2 lsd1-2 double mutants fully suppressed lsd1 rcd [22]. adr1-L1 also fully suppressed lsd1-2 rcd, while adr1 had only a slight effect (Figure 1A, 1B). We quantified this phenotype by monitoring cellular ion leakage via changes in media conductivity, an established proxy for membrane damage associated with cell death [32]. Col-0 plants did not exhibit significant changes in media conductivity, but lsd1-2 plants showed increasing conductivity, with the highest reading at 92 hours post-BTH treatment. adr1-L1 lsd1-2 and adr1-L2 lsd1-2 both exhibited complete ion leakage suppression, while adr1 lsd1-2 exhibited a marginal effect (Figure 1C). Thus, ADR1-L1 and ADR1-L2 are each required for lsd1 rcd.
We noted that adr1-L1 and adr1-L2 exhibited non-allelic non-complementation (NANC), a rare genetic condition where plants that are heterozygous at both loci phenotypically resemble either homozygous single mutant. Thus, plants homozygous for lsd1-2 and heterozygous for both ADR1-L1 and ADR1-L2 exhibited full suppression of lsd1 rcd (Figure 1D). We also found that adr1-L2 was fully recessive, whereas adr1-L1 appeared to be semi-dominant (Figure 1D). NANC frequently indicates that the two genes act closely together or that the two proteins physically interact or are a part of the same protein complex, and that their overall dose is important for their shared function [33]. Because all three ADR1 proteins share significant amino acid identity, we speculated that lowering of the overall ADR1 dose might be sufficient to suppress lsd1 rcd. Thus, the weak adr1 rcd suppression phenotype might simply reflect low expression of ADR1 relative to ADR1-L1 and ADR1-L2. Quantitative RT-PCR analysis of gene specific mRNA levels confirmed that ADR1 is expressed at lower levels than ADR1-L1 and ADR1-L2 under our growth conditions, consistent with this model (Figure 1E).
ADR1-L2 is a positive regulator of lsd1-mediated cell death. This could be due either to (i) a requirement for ADR1-L2 activation in cells destined to die, followed by its continued activation in neighboring cells, as the SA-dependent signal for rcd spreads in the absence of LSD1 [23], [34]; or (ii) a requirement for ADR1-L2 activation in cells initially triggered to die, with this activation contributing to the spread of an ADR1-L2-independent cell death signal beyond the primary cell death site. To distinguish between these two hypotheses, we generated an estradiol-driven (Est) conditional expression system, which induces local target gene expression [35]. adr1-L2 lsd1-2 plants expressing an estradiol-induced, HA epitope-tagged ADR1-L2 transgene were constructed (Materials and Methods). Expression of ADR1-L2 was activated by local application of estradiol on only part of a leaf, thus creating an artificial chimera containing both adr1-L2 lsd1-2 and ADR1-L2 lsd1-2 sectors (Figure 2A). ADR1-L2 expression was limited to the area of estradiol application as measured via Western blot (Figure 2B). BTH treatment was then used to induce lsd1-mediated rcd. We observed that cell death was limited to the zone of estradiol treatment and did not expand into the adr1-L2 lsd1-2 sector (Figure 2C). This result supports our first hypothesis: ADR1-L2 expression is continuously required in cells undergoing lsd1-mediated rcd.
We previously noted that ADR1-L2 is required for SA accumulation following effector and MAMP recognition, and that this does not require an intact P-loop motif [22]. However, these results do not preclude additional, canonical P-loop-dependent functions for ADR1-L2. Thus, we tested whether or not the positive regulatory function of ADR1-L2 in lsd1 rcd is P-loop dependent. We generated adr1-L2 lsd1-2 plants expressing ADR1-L2AAA, a mutated allele of ADR1-L2 which carries alanine (A) substitutions in the three consecutive conserved residues within the P-loop motif which are essential for nucleotide binding [22]. Interestingly, ADR1-L2AAA fails to complement for lsd1 rcd following BTH treatment (Figure 3A), even though this construct retains wild type BTH-induced ADR1-L2 protein accumulation (Figure 3B). Despite repeated attempts, we could not recover adr1-L2 plants over-expressing ADR1-L2, presumably due to lethality of ectopic over-expression as noted for other sensor NB-LRR proteins (data not shown). Together these results suggest that ADR1-L2 activation in lsd1 rcd proceeds in a canonical, P-loop dependent manner.
Mutations of the aspartic acid (D) in the conserved MHD motif in plant NB-LRRs typically lead to autoactivity [10]–[14]. Mechanistically, this is thought to reflect either a preference for ATP binding or a lack of ATPase activity, either of which would favor the “on” state, according to current models of NB-LRR activation [7], [19]. Thus, a similar mutation in the MHD motif of ADR1-L2 should result in a permanent ‘on’ state, resulting in ectopic autoactivity. In the cases where it has been examined, NB-LRR autoactivity via MHD mutation has been shown to require an intact P-loop [10]–[14]. Thus, given the P-loop dependent function of ADR1-L2 in lsd1 rcd, we speculated that ADR1-L2 activity in additional defense contexts might also require an intact P-loop.
We generated adr1-L2 plants expressing ADR1-L2 with a Val (V) for Asp (D) substitution at amino acid 484 (Figure 4A; hereafter ADR1-L2D484V). As expected, ADR1-L2D484V transgenics exhibited a dwarfed, cpr (Constitutive PR expression)-like phenotype [36] with short hypocotyls, pointed leaves (Figure 4B), and a bushy appearance after bolting. In contrast, adr1-L2 plants expressing wild-type ADR1-L2 appeared morphologically similar to wild-type Col-0 plants (Figure 4B). Both transgenes were expressed from the native ADR1-L2 promoter, with C-terminal HA epitope tags (Figure 4C). We note that the majority of ADR1-L2D484V transgenic lines accumulated higher protein levels than those expressing the wild-type ADR1-L2 allele. We selected ADR1-L2 and ADR1-L2D484V lines expressing similar levels of protein to show that the cpr-like phenotype is not simply a result of higher protein levels in the autoactive mutant (Figure 4C); the differences in morphology persist. Additional ADR1-L2D484V lines expressing less ADR1-L2D484V protein were also recovered; these did not exhibit strong cpr-like phenotypes, suggesting that there is a threshold amount of ADR1-L2D484V required for the associated phenotypes (data not shown).
The ADR1 family members work additively to limit pathogen growth, with adr1 triple mutant plants exhibiting increased susceptibility to virulent pathogens [22]. We therefore tested the ability of autoactive ADR1-L2D484V to confer enhanced basal defense against otherwise virulent pathogens. ADR1-L2D484V plants displayed increased resistance to both Hyaloperonospora arabidopsidis (Hpa) Emco5 and Pseudomonas syringae pv tomato (Pto) DC3000 (Figure 4D, 4E). Trypan blue staining of cotyledons after inoculation with Hpa Emco5 revealed predominantly free hyphal growth in the wild-type Col-0 control and adr1-L2, which was enhanced in the fully susceptible control, eds1 (Figure 4F). ADR1-L2D484V plants, on the other hand, exhibited only localized hypersensitive cell death (HR) as well as a basal level of cell death (Figure 4F, top row) not seen in the other genotypes. Thus, ADR1-L2D484V constitutively triggers downstream signaling and increased immune function.
We examined the dependence of the ADR1-L2D484V cpr-like phenotype on the P-loop. The triple missense P-loop dead mutation, ADR1-L2AAA [22], and the autoactive ADR1-L2D484V mutation were combined in cis (Figure 4A) and transformed into adr1-L2 plants. ADR1-L2AAA D484V plants did not exhibit the cpr-like phenotype (Figure 5A) despite the fact that they expressed levels of ADR1-L2AAA D484V protein that are similar to ADR1-L2D484V levels sufficient to cause the dwarfed phenotype (Figure 5B). Thus, an intact P-loop domain is required for ADR1-L2D484V autoactivity. We infer that ADR1-L2D484V is an activated version of this NB-LRR which can be used to study the canonical, P-loop dependent functions of ADR1-L2.
ADR1-L2 was identified as a positive regulator of lsd1 rcd ([34], above). LSD1 and ADR1-L2 both function downstream of the NADPH oxidase-dependent ROI burst driven by NB-LRR sensor activation, but upstream of SA accumulation [22], [25], [26]. Additionally, ADR1-L2 is locally required for lsd1-mediated rcd and its function in this context is P-loop dependent (Figure 2, Figure 5). Thus, we hypothesized that genetic components known to regulate lsd1 rcd might also be required for ADR1-L2D484V activity. We generated double mutants between ADR1-L2D484V and the lsd1 suppressors sid2, eds1, and atmc1 to define genetic interactions required for the ADR1-L2D484V phenotypes. We also generated ADR1-L2D484V atrbohD double mutants to define whether an oxidative burst is required for the ADR1-L2D484V phenotypes. We examined these double mutants for ADR1-L2D484V protein accumulation, alterations in the ADR1-L2D484V cpr-like morphology, enhanced resistance to the virulent Hpa isolate Emco5, and steady-state SA levels.
AtRbohD is generally required for effector-driven, NB-LRR-dependent superoxide production, but not for lsd1 rcd [23]. In fact, lsd1-2 atrbohD plants exhibit increased rcd compared to lsd1-2 single mutants, a phenotype that depends on SA accumulation [25]. This result suggests that the NADPH oxidase can down-regulate the spread of cell death as SA-dependent signals emanate from an infection site [23]. atrbohD ADR1-L2D484V plants morphologically resembled the ADR1-L2D484V parent (Figure 6A, Figure S1) and expressed a similar level of ADR1-L2D484V protein (Figure 6B). Like the ADR1-L2D484V parent, atrbohD ADR1-L2D484V plants were significantly more resistant to Hpa Emco5 (Figure 6C), and had extremely high steady-state levels of SA (Figure 6D). We conclude that ADR1-L2D484V autoactivity, unlike effector-driven NB-LRR activation, is downstream, or independent, of AtRbohD.
SA is required for lsd1 rcd [25] and mediates basal defense in plants [37]. Additionally, SA levels are reduced in adr1-family triple mutant plants, corresponding to diminished basal defense and an increase in disease susceptibility [22]. Thus, it seemed likely that the increased basal defense in ADR1-L2D484V plants could be due to the massive increase in SA observed in this line (Figure 6D). We tested this hypothesis using the sid2 mutant, which is unable to synthesize SA due to a mutation in the biosynthetic isochorismate synthase gene, ICS1 [16]. sid2 ADR1-L2D484V plants morphologically resembled the ADR1-L2D484V parent (Figure 6A, Figure S1) and accumulated similar amounts of ADR1-L2D484V protein (Figure 6B). sid2 ADR1-L2D484V plants exhibited enhanced basal defense to Hpa Emco5, though not to the same extent as ADR1-L2D484V (Figure 6C). As expected, sid2 ADR1-L2D484V plants did not accumulate SA (Figure 6D). These observations indicate that the defense cpr-like phenotypes of ADR1-L2D484V consist of both SA-dependent and SA-independent components, whereas the cpr-like growth phenotype is SA-independent.
EDS1 is required for lsd1-mediated rcd [26] and is an essential regulator of both basal defense against virulent pathogens [38], [39] and TIR-NB-LRR dependent ETI [40]–[42]. Exogenous SA rescues eds1 basal defense phenotypes, suggesting that EDS1 acts upstream of ICS1, at least for the phenotypes assayed [42], [43]. Importantly, eds1 ADR1-L2D484V plants were significantly more dwarfed than ADR1-L2D484V (Figure 6A, Figure S1), though these two lines expressed similar levels of ADR1-L2D484V protein (Figure 6B). eds1 ADR1-L2D484V double mutants were completely resistant to Hpa Emco5 (Figure 6C), and had steady-state SA levels that were higher than the ADR1-L2D484V single mutant (Figure 6D). These surprising results demonstrate that EDS1 is a negative regulator of the SA-accumulation observed in ADR1-L2D484V.
AtMC1 is a metacaspase required for lsd1 rcd; AtMC1 also contributes significantly to ETI-dependent HR [27]. atmc1 ADR1-L2D484V plants were extremely dwarfed (Figure 6A, Figure S1). However, these plants were not sterile; they produced small amounts of seed and had a very long life cycle compared to wild-type Col-0 or ADR1-L2D484V plants (data not shown). They also accumulated more ADR1-L2D484V protein than the ADR1-L2D484V parent (Figure 6B). Cotyledons of the atmc1 ADR1-L2D484V plants were similar in size to those of ADR1-L2D484V plants, and we were thus able to perform Hpa infection assays. We determined that atmc1 ADR1-L2D484V cotyledons are completely resistant to Hpa Emco5 (Figure 6C). Due to the extremely small size of the atmc1 ADR1-L2D484V double mutant, we were unable to perform SA analysis on this line. Collectively, these data indicate that AtMC1 negatively regulates ADR1-L2D484V protein accumulation, and likely subsequent SA accumulation leading to a hyper-cpr phenotype.
ADR1-L2 is required for lsd1-mediated rcd [22]. We therefore examined whether ADR1-L2D484V affects the lsd1 phenotype. We crossed lsd1-2 and ADR1-L2D484V plants, and in the F3 generation homozygous ADR1-L2D484V plants were selected via Basta resistance markers on the transgene (Materials and Methods). ADR1-L2D484V homozygotes were genotyped for lsd1-2; none were lsd1-2 homozygous (Table S1). Additionally, we carried lsd1-2 homozygous, ADR1-L2D484V heterozygous plants forward an additional generation, and again used the Basta resistance marker to identify homozygous ADR1-L2D484V plants. None were recovered. Next, we attempted to transform lsd1-2 mutant plants with the same ADR1-L2D484V construct used in the adr1-L2 transformation. No lines were recovered that expressed detectable levels of ADR1-L2D484V protein, and no plants that were recovered displayed the dwarfed phenotype (data not shown). We concluded that lsd1-2 ADR1-L2D484V is lethal.
We therefore looked for genetic determinants required for lsd1 ADR1-L2D484V lethality. As stated above, eds1 and atmc1 are both suppressors of lsd1 rcd. We therefore crossed atmc1 lsd1-2 or eds1 lsd1-2 plants, which express wild-type growth, to ADR1-L2D484V. atmc1 lsd1-2 ADR1-L2D484V plants could not be recovered (data not shown), indicating that AtMC1 is not required for lethality of lsd1-2 ADR1-L2D484V. However, we did recover eds1 lsd1-2 ADR1-L2D484V plants. These plants surprisingly exhibited wild-type morphology (Figure 7A), resembling eds1 lsd1 [26]. The suppression of the ADR1-L2D484V cpr-like phenotype is likely due to a much lower level of steady state ADR1-L2D484V accumulation in the eds1 lsd1-2 ADR1-L2D484V plants compared to parental plants (Figure 7B). Despite examining many eds1 lsd1-2 ADR1-L2D484V plants from 4 independent progenies, no plant with ADR1-L2D484V parental expression levels was recovered. Additionally, eds1 lsd1-2 ADR1-L2D484V plants did not accumulate the high levels of SA observed in ADR1-L2D484V (Figure 7C).
In light of the surprising result that eds1 lsd1-2 ADR1-L2D484V plants are essentially wild-type, we re-confirmed the genotypes and phenotypes of eds1 ADR1-L2D484V and eds1 lsd1-2 ADR1-L2D484V. For this, we used a line that was homozygous for eds1 and ADR1-L2D484V but heterozygous for LSD1 and expressed the wild-type morphology. In the next generation, both dwarfed and wild-type size plants were identified (Figure S2A). These plants were genotyped for LSD1, and all dwarfed plants were found to be LSD1 homozygotes (Figure S2B, 20 of 70 plants were LSD1 homozygotes). Wild-type size plants were either LSD1/lsd1 heterozygotes (34 of 70 plants) or lsd1 mutants (16 of 70 plants), suggesting that the dominant wild-type phenotype in this context is the result of LSD1 haploinsufficiency. We therefore conclude that the difference in the phenotypes between eds1 lsd1-2 ADR1-L2D484V (wild-type) and both eds1 ADR1-L2D484V (nearly lethal) and lsd1 ADR1-L2D484V (lethal) is genuine. Further, in the presence of autoactive ADR1-L2D484V, the combined absence of EDS1 and the loss, or reduction, of LSD1 leads to down-regulation of ADR1-L2D484V protein accumulation and restoration of wild-type morphology.
We addressed whether the lowered accumulation of ADR1-L2D484V protein in eds1 lsd1-2 ADR1-L2D484V was due to transcriptional regulation. We performed quantitative RT-PCR, and discovered that the ADR1-L2D484V transcript levels in eds1 lsd1-2 ADR1-L2D484V plants were lower than in ADR1-L2D484V (Figure 7D), generally consistent with the diminution of ADR1-L2D484V protein in eds1 lsd1-2 ADR1-L2D484V (Figure 7B). LSD1 and EDS1 are known to work together in an SA regulatory feedback loop [26]. Given that eds1 lsd1-2 ADR1-L2D484V plants are morphologically normal, express lower levels of SA than ADR1-L2D484V, and accumulate lower levels of ADR1-L2 transcript and protein than ADR1-L2D484V (Figure 7), and that ADR1-L2 accumulation is up-regulated by BTH application (Figure 4C), we speculate that this loop also regulates ADR1-L2 expression. In support of this hypothesis, we also noted that ADR1-L2D484V transcript accumulated to significantly higher levels than the endogenous ADR1-L2 transcript in wild-type Col-0 plants (Figure 7D), indicating that plants expressing the activated ADR1-L2 allele constitutively up-regulate ADR1-L2 transcription.
The phenotypic suppression of lsd1 lethality and of eds1 ADR1-L2D484V morphological defects in eds1 lsd1 ADR1-L2D484V plants suggests that ADR1-L2D484V autoactivity signals via two parallel pathways leading to SA accumulation, one EDS1- and one LSD1-dependent. These converge through mutual negative regulation exerted by EDS1 on the LSD1-dependent pathway and vice versa. LSD1 dampens an SA regulatory feed-forward loop that requires EDS1 [26]. EDS1 dampens an LSD1-dependent SA-accumulation (Figure 7). Thus it is plausible that eds1 lsd1 ADR1-L2D484V resembles a wild-type plant because the SA levels cannot be feed-forward amplified. To test this hypothesis, we generated sid2 eds1 ADR1-L2D484V plants by crossing sid2 ADR1-L2D484V to eds1 ADR1-L2D484V. Similar to eds1 lsd1 ADR1-L2D484V, these plants exhibited complete suppression of the nearly lethal eds1 ADR1-L2D484V phenotype (Figure 8A). Additionally, the steady state accumulation of the transgene was lowered compared to either parental line (Figure 8B). We noted that the reduced protein accumulation was not caused by transgene silencing, as F2 progeny from sid2 ADR1-L2D484V×eds1 ADR1-L2D484V segregated the SID2 eds1 ADR1-L2D484V morphological phenotype (Figure 8A). Quantitative RT-PCR on ADR1-L2 transcript suggested that, similar to eds1 lsd1 ADR1-L2D484V, the reduced transgene accumulation is transcriptional (Figure 8C). As noted above, an additional hallmark of ADR1-L2D484V autoactivity is enhanced immune function. We thus tested whether the enhanced basal defense response of ADR1-L2D484V is affected in the eds1 sid2 mutant background. Strikingly, sid2 eds1 ADR1-L2D484V plants were extremely susceptible to Hpa Emco5, more so than either single sid2 or eds1 mutants (Figure 8D). A model consistent with these observations and previous publications is presented in Figure 9 and discussed below.
The autoactive phenotypes of ADR1-L2D484V plants require ADR1-L2D484V protein accumulation above a threshold. This indicates that the expression level of wild-type ADR1-L2 may also be under exquisite control. The co-chaperone RAR1, while not necessary for the function of all NB-LRRs, is required for the steady state accumulation of all NB-LRRs tested to date [44]–[47]. We thus crossed adr1-L2 pADR1-L2:ADR1-L2-HA to rar1-21 [46]. Plants genotyped as homozygous rar1-21 and homozygous RAR1 exhibited similar levels of ADR1-L2-HA protein (Figure S3A), indicating that RAR1 is not required for ADR1-L2 accumulation. The rar1 genotype was confirmed by Western blot for RAR1 protein (Figure S3B). ADR1-L2 expression can be up-regulated with BTH [22]. We therefore also tested whether RAR1 is required for the high levels of ADR1-L2 accumulating after BTH treatment. BTH induced ADR1-L2 protein in rar-21 ADR1-L2-HA plants accumulated to levels at least as high as those in RAR1 ADR1-L2-HA plants (Figure S3A). Therefore, RAR1 is dispensable for both steady-state ADR1-L2 accumulation, in contrast to other assayed NB-LRR proteins [44]–[47], and for its BTH-induced up-regulation.
We recently demonstrated that the plant NB-LRR immune receptor ADR1-L2 can have non-canonical, P-loop independent ‘helper’ functions in plant defense [22]. Here, we sought first to define canonical, P-loop dependent function(s) for ADR1-L2, and then to understand the genetic requirements for these functions. We demonstrated that wild-type ADR1-L2 is required locally at the site of BTH-driven cell death activation in the lsd1 cell death control mutant. This activity requires an intact P-loop and is thus canonical. In this context, ADR1-L2 genetically interacts with ADR1-L1 to control runaway cell death, as shown by NANC, further suggesting that members of the ADR1 family function together in cell death signaling. ADR1-L2 does not require RAR1 for either its steady state accumulation, nor for its induced accumulation following BTH treatment. This is the first report of either steady state or inducible NB-LRR accumulation that is not RAR1-dependent. This result may differentiate ‘helper’ NB-LRRs from ‘sensor’ NB-LRRs. We propose that levels of the former might be dictated by the signaling partners with which they function in specific stoichiometries, while the latter, acting as effector-sensors, are threshold-regulated by the NB-LRR co-chaperone complex [48].
Given the canonical P-loop-dependent function of ADR1-L2 as a positive regulator of lsd1 cell death, we inferred that ADR1-L2, like other NB-LRRs studied to date, retains the ability to undergo a nucleotide-dependent conformational switch to regulate its activation. Thus, we sought a context in which we could analyze canonical ADR1-L2 P-loop dependent functions, despite the absence of an effector to trigger it. We created an autoactive allele, ADR1-L2D484V. ADR1-L2D484V plants exhibit the dwarfed morphology and constitutively active defense responses observed in other autoactive NB-LRR mutants. We showed that this autoactivity requires an intact P-loop. We then used this allele as a proxy for canonical activation of ADR1-L2 in a series of epistasis experiments. We present a model consistent with our new findings and previous genetic analyses [22], [23], [25], [26], [30] (Figure 9).
Canonical, P-loop dependent, ‘sensor’ NB-LRR functions typically drive both the AtrbohD NADPH oxidase-dependent oxidative burst following effector perception and SID2-dependent SA accumulation [23]. By contrast, ADR1-L2D484V autoactivity is downstream, or independent, of AtrbohD, yet still drives SID2-dependent SA accumulation. This is consistent with the previously defined, P-loop-independent ‘helper’ activity of ADR1-L2 [22].
Plants expressing ADR1-L2D484V exhibited increased disease resistance and very high steady-state levels of SA. sid2 ADR1-L2D484V plants expressed, as expected, very low levels of SA, but these plants did not completely revert to wild-type morphology, and they maintained an increased level of enhanced disease resistance. Thus, there must be SA-independent regulation of activated ADR1-L2. Redundant functions of EDS1 and SA in plant defense mediated by ‘sensor’ NB-LRR functions have been reported [30]. In that work, sid2 or eds1 mutants were insufficient to disrupt CC-NB-LRR-mediated disease resistance, while combined loss of both gene products led to loss of resistance [30]. Our results support this model, since the constitutive activation of ADR1-L2D484V results in both SA-dependent and SA-independent phenotypes (Figure 9). Given these data, as well as the fact that eds1 lsd1 ADR1-L2D484V phenocopies sid2 eds1 ADR1-L2D484V we conclude that the SA-independent pathway we describe here requires EDS1 (Figure 9, left).
One of our most surprising observations is the phenotypic rescue of both the lethal lsd1 ADR1-L2D484V phenotype and the nearly lethal eds1 ADR1-L2D484V phenotype in eds1 lsd1 ADR1-L2D484V plants. It is important to recall that either adr1-L2 or eds1 suppresses lsd1 rcd [22], [26]. Recall also that the P-loop independent function of ADR1-L2 as a ‘helper’ is downstream of AtRbohD, but upstream of SA accumulation [22]. This is in agreement with the autoactive ADR1-L2D484V phenotype, which bypasses AtRbohD but still drives enhanced SA levels. Notably, loss of LSD1 in the eds1 ADR1-L2D484V context functionally resembles loss of SID2. Since SID2-dependent SA accumulation is regulated by LSD1, we conclude that both SA and EDS1 are required for ADR1-L2D484V autoactivity. Loss of either genetic component destroys the fine-tuned equilibrium between EDS1-dependent and SA-dependent processes in this autoactivity.
P-loop-dependent activation of ADR1-L2 results in SID2-dependent SA accumulation via two separate pathways (Figure 9). In the first pathway, ADR1-L2D484V constitutively signals to EDS1, which in turn positively regulates SID2, increasing SA levels. ADR1-L2D484V also triggers additional SA production in a parallel pathway that requires LSD1 and is antagonized by EDS1. In support of our model, SA regulates EDS1 transcription [29], which in turn regulates SID2 [49]. Once activated, ADR1-L2 causes cell death, which drives more AtRbohD-dependent ROI [34] and SA accumulation in surrounding cells [34], [50]. In both pathways, SA is part of a feedback loop that further potentiates the P-loop dependent activity of ADR1-L2, as indicated by the fact that ADR1-L2 is BTH inducible. Thus, ADR1-L2 is also both upstream and downstream of SA accumulation (Figure 9).
Our data are consistent with ADR1-L2 transcriptional regulation by both SA-dependent and -independent pathways (Figure 9). In an otherwise wild-type plant expressing activated ADR1-L2, the antagonism between EDS1 and LSD1 maintains SA production below toxic levels. In an lsd1 plant, the level of SA surpasses this level via ectopic ADR1-L2 activation and consequent SA production. This increased SA in turn drives higher ADR1-L2 expression, and the cycle repeats. This is exacerbated, and lethal, in lsd1 ADR1-L2D484V. eds1 and sid2 suppress lsd1 because feed forward regulation of the SA accumulation cycle is blocked. The surprising eds1 lsd1 ADR1-L2D484V and sid2 eds1 ADR1-L2D484V phenotypes are consistent with the low level of SA in these lines being insufficient to up-regulate ADR1-L2 expression. Thus, even though there is chronic signaling feeding the cycle in ADR1-L2D484V, the EDS1-dependent, SA-independent pathway is interrupted in eds1 lsd1 ADR1-L2D484V and sid2 eds1 ADR1-L2D484V. How LSD1 and EDS1 negatively regulate each other has yet to be determined, although our data suggest that LSD1 might regulate EDS1 function through transcriptional control, as EDS1 transcription levels are increased in an lsd1 mutant (Figure S4). In support of this hypothesis, a role for LSD1 as a cytosolic retention factor for the AtbZIP10 transcription factor [51] may provide a mechanism for LSD1 control of EDS1 expression.
Our model (Figure 9) supports a scenario in which in wild-type, P-loop dependent NB-LRR activation leads to local increased levels of SA via an AtRbohD-dependent ROI burst and SID2-dependent SA accumulation. The spread of this SA accumulation is spatially down-regulated through a combined action of EDS1 and LSD1 at increasing distance from the infection site. As stated above, our model also implies that SA functions both up- and down-stream of ADR1-L2. This may readily reconciled with our previous finding that ADR1-L2 helper function is required for SA accumulation and cell death, since ARD1-L2 is SA-up-regulated [22].
Overall, we present a general approach to characterize canonical, P-loop dependent functions of NB-LRR proteins in the absence of a specific effector. We applied this to a recently characterized ‘helper’ NB-LRR protein, ADR1-L2. We identified genetic components that regulate its P-loop-dependent, canonical functions, and found that they, in turn, are regulated by suppressors of the lsd1 rcd phenotype. Our work suggests that the genetic requirements for ‘helper’ NB-LRR function may differ from the effector-driven activation of canonical ‘sensor’ NB-LRRs. Given that ADR1-L2, unlike other NB-LRRs, is required for lsd1 rcd, we note that our results may be mainly relevant to the dissection of the functions of ADR1-L2 and its paralogues, rather than being broadly applicable to understanding of ‘sensor’ NB-LRRs. Nevertheless, in agreement with previous reports on ‘sensor’ NB-LRR function [30], we conclude that the P-loop-dependent autoactivity of ADR1-L2 relies on signaling pathways that differ in their requirement for SA accumulation, but which are both regulated by EDS1. Thus, though the requirements for ‘sensor’ and ‘helper’ NB-LRR functions may be separable, they could still share some overlapping features.
A significant challenge remains to address the sub-cellular localization of these regulatory circuits [52]. Resting state NB-LRRs are localized to diverse sub-cellular compartments, and dynamic re-localization may accompany effector-driven activation of some [19]. Defining any dynamics of protein localization associated with the differential ADR1-L2 canonical and non-canonical functions will be ultimately important for understanding the genetic network that we describe.
All Arabidopsis lines are in the Columbia (Col-0) ecotype. adr1-1 [22], adr1-L1-1 [22], adr1-L2-4 [22], eds1-2 [49], sid2-1, atrbohD [23], lsd1-2 [24], atmc1 [27], and rar1-21 [46] are described elsewhere; primers used to genotype these lines are in Table S2. For generation of adr1-L2 plants expressing ADR1-L2-HA, ADR1-L2D484V-HA, and ADR1-L2AAA D484V-HA lines, the C-terminal HA-tagged coding sequence of wild-type ADR1-L2 or the mutated alleles were fused to its native promoter (500 bp) and cloned in the pBAR (Basta resistant) Gateway vector [53]. For generation of adr1-L2 lsd1-2 plants expressing an estradiol inducible ADR1-L2-HA, the coding sequence of ADR1-L2 was cloned into a modified pMDC7 (hygromicin resistant) Gateway vector carrying a C-terminal HA tag. Arabidopsis transgenics were generated using Agrobacterium (GV3101)-mediated floral dip transformation [54]. Basta selection of transgenic plants was performed by spraying 10-day-old seedlings. Plants were grown under short day conditions (9 hrs light, 21°C; 15 hrs dark, 18°C).
Leaves from 4-week-old plants were harvested and total proteins were extracted by grinding frozen tissue in a buffer containing 20 mM Tris-HCl (pH 7.0), 150 mM NaCl, 1 mM EDTA (pH 8.0), 1% Triton X-100, 0.1% SDS, 10 mM DTT, and plant protein protease inhibitor cocktail (Sigma-Aldrich). Samples were centrifuged at 14,000 rpm for 15 min at 4°C to pellet debris. Proteins were separated on 7.5% (ADR1-HA) or 12% (RAR1) SDS-PAGE gels and were transferred to polyvinylidene difluoride membrane. Western blots were performed using standard methods. Anti-HA (Santa Cruz Biotechnology) antibody was used at a 1∶3000 dilution; anti-RAR1 (custom anti-RAR1 polyclonal antibody was made against the full length RAR1 with C-terminus GST tag by Cocalico Biologicals, Inc.) was used at a 1∶2000 dilution. Signals were detected by enhanced chemiluminescence using ECL Plus (Amersham Biosciences). For BTH induction experiments (300 µM), plants were collected 24 hpi.
SA and SAG measurements were performed as described [55]. Briefly, 100 mg of leaves were collected from 4-week-old plants and frozen in liquid nitrogen. Samples were ground and tissue was homogenized in 200 µl 0.1M acetate buffer pH 5.6. Samples were centrifuged for 15 min at 16,000 g at 4°C. 100 µl of supernatant was transferred to a new tube for free SA measurement, and 10 µl was incubated with 1 µl 0.5 U/µl β-glucosidase for 90 min at 37°C for total SA measurement. After incubation, plant extracts were diluted 5-fold with 44 µl acetate buffer for free SA measurement. 60 µl of LB, 5 µl of plant extract (treated or not with β-glucosidase), and 50 µl of Acinetobacter sp. ADPWH-lux (OD = 0.4) were added to each well of a black 96-well plate (BD Falcon). The plate was incubated at 37°C for 60 min and luminescence was read with Spectra Max L (Molecular Devices) microplate reader. For the standard curve, 1 µl of a known amount of SA (Sigma; from 0 to 1000 µg/ml) was diluted 10-fold in sid2-1 plant extract, and 5 µl of each standard (undiluted for free SA measurement, or 5-fold diluted for total SA) was added to the wells of the plate containing 60 µl of LB and 50 µl of Acinetobacter. SA standards were read in parallel with the experimental samples. For BTH induction experiments (300 µM), plants were collected 24 hpi.
Ten-day-old seedlings were spray-inoculated with 50,000 spores/ml of Hyaloperonospora arabidopsidis isolate Emco5. Pots were covered with a lid to increase humidity during inoculation and pathogen growth. Sporangiophores were counted at 4 dpi as described [56]. Pto DC3000(EV) was resuspended in 10 mM MgCl2 to a final concentration of 2.5×105 cfu/ml (OD600 = 0.0005). Twenty-day-old seedlings were dipped in the bacterial solution and growth was assessed as described [57].
4-week-old plants were sprayed with 300 µM BTH, or 10-day-old plants were inoculated with Hpa Emco5 as described above. Leaves were harvested and stained with lactophenol Trypan Blue (TB) to visualize dead cells as described [58]. For the conductivity measurements, 4-week-old plants were sprayed with 300 µM BTH. Plants were harvested and 4 leaf discs (7 mm) were cored and then floated in water for 30 min. These leaf discs were transferred to tubes containing 6 ml distilled water. Conductivity of the solution (μSiemens/cm) was determined with an Orion Conductivity Meter at the indicated time points [59].
The central portion of the right halves of leaves from 4-week-old transgenic adr1-L2 lsd1-2 plants expressing an estradiol inducible allele of ADR1-L2 were hand-infiltrated with Est (20 µM) using a needleless syringe. 300 µM BTH was sprayed on the whole plant 24 h post-Est application. 20 µM Est was then hand-infiltrated on the same portion of the leaves 2 dpi to ensure expression of ADR1-L2. Leaves were collected 5 dpi from the first Est infiltration.
Leaves from 4-week-old plants were collected, frozen into liquid nitrogen and ground into powder with a mortar and pestle. RNA was extracted using TRIzol (Invitrogen), DNased (Ambion Turbo DNase), and cleaned up with Qiagen RNeasy Mini kit. Reverse transcription was performed (Ambion RETROscript) using 1 µg total RNA, and cDNA was analyzed with SYBR green (Applied Biosystem) using an Applied Biosystems ViiA7. Primers used are listed in Table S2.
Pots of sibling plants fixed for eds1 and segregating lsd1-2 (LSD1 heterzygotes) were Basta sprayed to check for segregation of ADR1-L2D484V. Those found to be eds1 ADR1-L2D484V were transplanted individually into pots, monitored for size, and genotyped for the T-DNA insertion of the lsd1-2 mutation.
|
10.1371/journal.pbio.2004455 | p38α blocks brown adipose tissue thermogenesis through p38δ inhibition | Adipose tissue has emerged as an important regulator of whole-body metabolism, and its capacity to dissipate energy in the form of heat has acquired a special relevance in recent years as potential treatment for obesity. In this context, the p38MAPK pathway has arisen as a key player in the thermogenic program because it is required for the activation of brown adipose tissue (BAT) thermogenesis and participates also in the transformation of white adipose tissue (WAT) into BAT-like depot called beige/brite tissue. Here, using mice that are deficient in p38α specifically in adipose tissue (p38αFab-KO), we unexpectedly found that lack of p38α protected against high-fat diet (HFD)-induced obesity. We also showed that p38αFab-KO mice presented higher energy expenditure due to increased BAT thermogenesis. Mechanistically, we found that lack of p38α resulted in the activation of the related protein kinase family member p38δ. Our results showed that p38δ is activated in BAT by cold exposure, and lack of this kinase specifically in adipose tissue (p38δ Fab-KO) resulted in overweight together with reduced energy expenditure and lower body and skin surface temperature in the BAT region. These observations indicate that p38α probably blocks BAT thermogenesis through p38δ inhibition. Consistent with the results obtained in animals, p38α was reduced in visceral and subcutaneous adipose tissue of subjects with obesity and was inversely correlated with body mass index (BMI). Altogether, we have elucidated a mechanism implicated in physiological BAT activation that has potential clinical implications for the treatment of obesity and related diseases such as diabetes.
| Accumulation of fat in adipose tissue is essential to store energy and insulate the body; however, excessive body fat leads to obesity. Of the 2 existing types of adipose tissue, white adipose tissue (WAT) stores energy, whereas brown adipose tissue (BAT) can produce heat. Activation of BAT and transformation of WAT into brown-like ‘brite/beige’ adipocytes have recently emerged as novel strategies against obesity. The uncoupling protein 1 (UCP1) is a hallmark of BAT and is responsible for triggering these 2 processes under the regulation of the p38 MAP kinase (p38MAPK) pathway, but the underlying mechanisms remain unknown. Here, we have analysed this process in detail and demonstrate that a protein kinase called p38α directly correlates with UCP1 levels in human adipose tissue, while it inversely correlates with body mass index (BMI). We find that mice lacking p38α in adipose tissue are protected against diet-induced obesity due to increased body temperature. In addition, another p38 family member, p38δ, is activated in these adipocytes lacking p38α and reduces their thermogenic capacity. Our results suggest that these 2 members of the p38 family have opposite roles in controlling thermogenesis.
| p38α has emerged as one of the main player that could activate the thermogenic capacity of adipose tissue. Because the thermogenesis of adipose tissue is reduced in obesity [6, 7, 21], we wondered whether expression of this kinase changes in human WAT during obesity. Using 2 cohorts for visceral fat and subcutaneous fat (sWAT) of adult patients with 80 and 170 samples, respectively, we found that the expression of p38α (Mapk14) in visceral fat and sWAT from individuals with obesity was reduced compared with those without obesity (Fig 1A and 1D). In fact, mRNA levels of Mapk14 in visceral fat inversely correlated with body mass index (BMI) (Fig 1B). It has been suggested that p38α in WAT activates browning by triggering the expression of UCP1 [18], the main protein responsible for adipose tissue thermogenic capacity [22]. In visceral fat and sWAT from individuals with obesity and those without obesity, we found that expression of Mapk14 correlated positively with the levels of Ucp1 (Fig 1C and 1E). This correlation reinforced the idea that p38α in visceral fat and sWAT controls the levels of UCP1 and could regulate browning in humans.
Then, we evaluated the function of p38α in adipose tissue using conditional mice (p38αFab-KO), which lacked p38α in WAT and BAT (S1 Fig). Under normal-chow diet (ND), p38αFab-KO mice had the same weight gain as the control Fab-Cre mice (S2A Fig). However, they presented reduced fat mass, in concordance with lower eWAT, perirenal WAT (pWAT), and BAT weight (S2B and S2C Fig). This reduction in fat accumulation was associated with higher energy expenditure and slight increase of body temperature (S2G and S2H Fig). In fact, these mice presented lower blood glucose levels in fasted and fed conditions (S2D Fig) and increased glucose tolerance (S2E Fig), with no differences in insulin sensitivity or insulin-stimulated glucose transporter type 4 (GLUT4) translocation in adipose tissue (S2E and S2F Fig). These data suggest that lack of p38α might protect against type 2 diabetes. Moreover, we evaluated whether lack of p38α affects adipogenesis, browning, and metabolism in eWAT and BAT. BAT from p38αFab-KO mice presented an increase of Cidea, a marker of browning, together with higher expression of glycolytic and β oxidation genes (S3 Fig).
To further evaluate the role of p38α in adipose tissue, mice were fed an HFD, and we observed that p38αFab-KO mice were completely protected from diet-induced obesity because their weight was identical to the weight of the control animals in ND (Fig 2A). This reduced weight gain was in line with lower fat mass (Fig 2B) and reduced weight of the different fat depots, including eWAT, sWAT, iWAT, pWAT, and BAT (S4A Fig). Moreover, liver weight was also reduced in agreement with protection against HFD-induced liver steatosis in p38αFab-KO mice (Fig 2C and S4A Fig). The protection against HFD-induced obesity was associated with reduced fasted and fed hyperglycaemia in p38αFab-KO mice, with no differences in triglyceridemia (Fig 2D and S4E Fig). In addition, p38αFab-KO mice were protected against HFD-induced glucose intolerance even when glucose dose was adjusted to lean mass (Fig 2E, S4B Fig.) and insulin resistance as shown by the reduced glucose levels during the insulin tolerance test (ITT) (Fig 2E). HFD-induced obesity was associated with liver insulin resistance and reduced insulin-stimulated Akt phosphorylation in livers from HFD-fed Fab-Cre mice (S4C Fig). Evaluation of insulin sensitivity in several tissues indicated that HFD-fed p38αFab-KO mice presented higher insulin-induced phosphorylation of Akt at Thr308 and Ser473 than HFD-fed Fab-Cre mice in liver and muscle but not in eWAT nor BAT (S4D Fig). Furthermore, we observed a slight increase of insulin-stimulated GLUT4 translocation in eWAT (Fig 2F). Together, these results demonstrate that p38αFab-KO mice are protected against diet-induced obesity and diabetes.
Histological analysis showed that interscapular BAT depot from HFD-fed p38αFab-KO mice had small multilocular adipocytes (Fig 2G), whereas in eWAT, we observed a slight decrease of adipocyte size (Fig 2G), which correlates to reduced cell size in BAT and WAT adipocytes from HFD-fed p38αFab-KO with respect to HFD-fed Fab-Cre (S5A and S5C Fig). Then, we evaluated HFD-induced WAT adipocyte expansion by bromodeoxyuridine (BrdU) staining [23], observing reduced expansion in p38αFab-KO (Fig 3A). However, no differences in Ki67 staining were observed after HFD in WAT or BAT adipocytes (S5A and S5C Fig).
To further investigate the mechanism by which lack of p38α in adipose tissue could protect against HFD-induced obesity, we evaluated whole-body metabolism using metabolic cages. HFD-fed p38αFab-KO mice showed a significant increase in whole-body energy expenditure analysed by ANCOVA, with no changes in food intake or respiratory exchange ratio (Fig 3B). These data are consistent with the observation that HFD-fed p38αFab-KO mice have higher skin temperature in the region of BAT compared with Fab-Cre mice (Fig 3C). Western blot analysis of BAT indicated that HFD-fed p38αFab-KO mice presented a slight increase of UCP1 expression associated with higher AMPK and Creb phosphorylation (Fig 3D and 3E). In addition, higher expression of UCP1 levels was observed in iWAT from HFD-fed p38αFab-KO mice (S5B and S7A Figs), suggesting an increased browning of this adipose depot. In contrast with the up-regulated UCP1 levels in iWAT, analysis of eWAT by western blot and immunohistochemistry showed that HFD-fed p38αFab-KO mice have reduced UCP1 levels in this tissue (S6 and S7B Figs). These results are in agreement with the results found in human visceral fat (Fig 1C) suggesting that, in visceral fat, p38α directly correlates with UCP1.
In vitro–differentiated brown adipocytes from p38αFab-KO mice confirmed a key role of this kinase inhibiting browning in a cell-autonomous manner because several browning markers (UCP1, PGC1b, Cidea, Cox7a1, Cox7a2, and Cox8b) were up-regulated in p38αFab-KO brown adipocytes (S8A Fig). In concordance with the results observed in the BAT tissue, glycolytic genes were also up-regulated, while many lipogenic genes that correlated with the lower triglyceride content in p38αFab-KO brown adipocytes were down-regulated (S8B, S8C, S8D and S8E Fig). In addition, p38αFab-KO brown adipocytes have increased expression of perilipin with no changes in adiponectin, suggesting same differentiation capacity but smaller and more abundant lipid droplets (S8B Fig). On the other hand, p38αFab-KO white adipocytes presented the same in vitro differentiation rate judging by red-oil staining and the expression levels of adipocyte markers such as adiponectin and perilipin (S8F and S8G Fig). However, p38αFab-KO white adipocytes have increased expression of leptin (S8F Fig). To further confirm the autonomous role of p38α in BAT, we crossed p38α loxP mice with UCP1-Cre mice [24], which express Cre recombinase specifically in the interscapular brown fat at room temperature, generating p38αUCP1-KO mice. In agreement with our previous results, these mice were protected against HFD-induced obesity and presented lower fat mass and increased temperature. Furthermore, they had lower blood glucose levels and partial glucose tolerance, indicating that they were protected against HFD-induced diabetes (Fig 4A–4F).
Our data at 23 °C demonstrated that lack of p38α resulted in increased whole-body energy expenditure due to the activation of BAT and iWAT thermogenesis. At this temperature, BAT is already fully differentiated; because it is complicated to detect an even higher level of UCP1, genetic modifications that up-regulate UCP1 levels cannot be easily detected [25]. For this reason, we therefore evaluated p38αFab-KO phenotype in thermoneutrality (30 °C) because it has been suggested to be more similar to the human situation [25]. At 30 °C, p38αFab-KO mice were also protected against HFD-induced obesity (Fig 5A) and presented lower body fat mass and increased BAT thermogenesis (Fig 5B and 5C), indicating that, even at temperatures at which BAT is impeded, these mice maintain BAT activation. In fact, UCP1 expression was much higher in BAT from p38αFab-KO than in the control Fab-Cre mice at 30 °C (Fig 5D). In addition, p38αFab-KO were also protected from HFD-induced diabetes at thermoneutrality (Fig 5E and 5F). Together, these data confirm that lack of p38α protects against HFD-induced obesity and diabetes due to an activation of BAT thermogenesis.
To gain insight into the molecular mechanism that might account for increased UCP1 levels and thermogenic capacity, we studied the signalling in the different adipose tissue depots. The p38MAPK pathway has been shown to trigger BAT activation in several models [18, 26–28]. Additionally, it has been found that p38α can inhibit the other p38 isoforms by a negative feedback loop that blocks the activation of the upstream kinases of this pathway [29]. Therefore, we evaluated the expression and phosphorylation state of the other p38s, with a phospho-p38 antibody that recognises all p38 isoforms [30]. Using adipocytes lacking p38γ/δ, we confirmed that p38α/β run around 38 kDa, while p38γ/δ run higher—around 41 kDa—allowing us to distinguish the phosphorylation of these kinases (S9A Fig). Under ND condition, p38δ and p38γ were hyperactivated in eWAT and iWAT from p38αFab-KO (S9B Fig). In agreement, p38δ/γ were activated more when cells were treated with sorbitol and p38α inhibitor SB203580 (S9C Fig). HFD resulted in reduced RNA expression of all the p38 isoforms in BAT, while in eWAT, only p38δ and p38γ decreased (S9D Fig). p38δ and p38γ were hyperactivated in iWAT and BAT from HFD-fed p38αFab-KO, whereas elevated p38δ (Mapk13) RNA levels were also found in BAT and eWAT from HFD-fed p38αFab-KO animals (Fig 6A, 6B and S7, S9E and S9F Figs). Activation of p38δ in BAT was diminished when mice were maintained at 30 °C (Fig 6A), suggesting that this p38 isoform might activate BAT thermogenesis. To further evaluate this hypothesis, mice lacking p38δ in adipose tissue (p38δFab-KO) were generated. In agreement with the importance of this kinase in BAT activation, p38δFab-KO mice fed with ND presented higher body weight, associated with increased fat mass and weight of all fat depots (Fig 6C and 6D and S10A Fig). In concordance, p38δFab-KO presented reduced energy expenditure, whole-body temperature, and decreased BAT thermogenesis (Fig 6E and 6F) as well as lower expression levels of Ucp1 and Ppargc1β in BAT (S10B Fig) with no differences in protein kinase A (PKA) phosphorylation (S10C Fig).
p38δ is activated in BAT upon cold exposure and in adipocytes after stimulation with the thyroid hormone T3 or norepinephrine (NE) (Fig 6G and 6H), suggesting that this p38 isoform might activate BAT thermogenesis. In fact, at 4 °C, p38δFab-KO mice have lower body and skin temperature in the BAT region (Fig 6I). Moreover, HFD-fed p38δFab-KO mice were more obese with higher fat mass and weight of all fat depots (S11A–S11C Fig). This increased adiposity correlated with lower BAT thermogenesis and lower UCP1, Ppargc1a, and Cidea levels in BAT (S11D–S11F Fig).
Our data indicated that p38δ was triggering thermogenesis because in vitro–differentiated brown adipocytes lacking p38δ have reduced expression of important genes implicated in BAT thermogenesis (Ppargc1b, Ppargc1a, Cidea, and Cox8b) and a slight decrease of Ucp1 and Cox7a1 supporting the cell-autonomous effect of p38δ in BAT thermogenesis (Fig 7A), with no differences in amount of mitochondrial DNA (Fig 7B and 7C).
Therefore, we evaluated respiration profiles in brown adipocytes lacking p38α and p38δ. Brown adipocytes lacking p38α presented higher leak respiration after isoproterenol (ISO) or NE treatment (Fig 7D). However, this augmented respiration capacity induced by NE or ISO was diminished when p38δ was chemically inhibited by BIRB796, a known inhibitor p38δ [31], as well as in p38δ-deficient brown adipocytes (Fig 7E and 7F), supporting the important role of this kinase in brown adipocyte activation.
In conclusion, we demonstrated that p38α in BAT inhibits p38δ activation, which in turn regulates BAT thermogenesis, energy expenditure, and body weight. We demonstrated that p38α and p38δ have opposite roles in BAT: whereas p38α inhibits BAT thermogenesis, p38δ induces it upon several physiological stimuli (Fig 8).
Adipose tissue has become an important target for the treatment of obesity, not only because its dysfunction could be responsible for diabetes development but also because increasing BAT thermogenesis and/or browning of WAT could lead to new therapeutic approaches against obesity [32, 33]. In this scenario, p38MAPK signalling has been proposed to be a key activator of these processes. Consequently, there is an increasing interest to understand the function of this pathway in the regulation of adipose tissue metabolism, remodelling, and browning.
A growing number of studies have defined p38MAPK as one of the main pathways that stimulates browning and BAT thermogenesis [18,21–23]. However, using genetically modified mice lacking specific p38 family members in adipose tissue, we have shown that lack of p38α in adipose tissue protects against HFD-induced obesity by increasing energy expenditure through the activation of BAT thermogenesis. Mechanistically, lack of p38α results in hyperactivation of p38δ in BAT together with increased UCP1 expression and higher Creb and AMPK phosphorylation. Negative feedback controls by p38α through the regulation of upstream activators of the pathway—such as TAB1 phosphorylation or MKK6 expression—have been previously reported [29, 34]. Here, we show that negative control of the pathway by p38α has biological and pathological implications. However, it would be interesting to examine the epistatic relationship between p38α and p38δ genetically in future studies.
We also demonstrated that p38δ is activated in BAT by 3 stimuli widely known to activate this tissue: cold exposure, NE, and thyroid hormone treatment [35, 36], whereas its phosphorylation is reduced under thermoneutrality conditions. In addition, p38δ expression in BAT was reduced in obese mice, while this down-regulation was ablated in p38αFab-KO mice, suggesting that activation of p38δ in p38αFab-KO mice is responsible for the protection against diet-induced obesity observed in these mice. Indeed, inhibition of p38δ in p38αFab-KO brown adipocytes abolished the increased respiratory capacity induced by β3-adrenergic stimuli. In agreement with the role of p38δ-promoting thermogenesis, mice lacking this kinase in adipose tissue developed overweight, even in ND, and showed decreased whole-body energy expenditure associated with lower temperature and reduced BAT activation. Moreover, we confirmed the cell-autonomous role of p38δ inducing browning using differentiated adipocytes.
Our results were completely unexpected because the p38MAPK pathway has been shown to trigger BAT activation in several models [18, 26–28], and—until now—it was thought that the only implicated family member was p38α. Moreover, we have recently found that hyperactivation of p38α in MKK6-deficient animals induces browning of eWAT [37]. These finding might indicate opposite effects of p38α in eWAT versus iWAT or BAT. While p38α would activate browning in eWAT—increasing energy expenditure—it would prevent it in iWAT, and it would block thermogenesis through the negative regulation of p38δ in BAT. In agreement with this hypothesis, we observed reduced levels of UCP1 in epididymal fat lacking p38α. In fact, our data from human samples indicated that the p38α mRNA levels in visceral fat directly correlates with UCP1 expression and inversely correlates with the BMI, suggesting that p38α triggers visceral fat browning. We also found that p38α in sWAT inversely correlates with UCP1. This is in accordance with results observed in mouse models, in which we found a decrease of all p38s after HFD in all fat depots. However, the levels of UCP1 expression in these human fat depots is quite low judging by the low Ct obtained (higher than 29), and evaluation of UCP1 protein expression in human fat depots would be necessary. Moreover, further studies to determinate the expression of p38 family members and upstream kinases in other human fat depots would help us to understand the role of these kinases in human adipocytes.
It has been proposed that p38α induces adipogenesis [38–40]. However, using genetically modified animals, we showed here that lack of p38α in preadipocytes did not affect their differentiation to adipocytes, nor did it affect changes in the differentiation markers evaluated in the major fat depots. This capacity of cells lacking p38α to still differentiate to adipocytes could be due to the hyperactivation of the other members of the family: p38γ and p38δ. In fact, it has been shown that p38 isoforms can compensate for each other [30]. Here, we demonstrated the cell-autonomous and opposite effects of 2 p38 isoforms in adipocytes, p38α and p38δ. The cell-specific actions of p38α in each fat depot could be explained by the specific expression pattern of p38 family members—p38α being the main isoform expressed in eWAT, whereas p38δ or p38γ are abundant in BAT or iWAT. Furthermore, our results suggest a different regulation of p38s expression in adipose tissue during obesity, with only decrease of p38δ and p38γ in eWAT and no effects in p38α or p38β. More studies would be necessary to elucidate the function of p38γ in adipose tissue.
We also evaluated the controversial role of p38α in GLUT4 translocation [41–43]. Under ND, insulin-induced GLUT4 translocation was the same in both control and p38αFab-KO mice. However, p38αFab-KO mice maintained the insulin-induced translocation after the HFD, perhaps due to the fact that these animals did not gain weight and were protected against diet-induced insulin resistance. In fact, our data suggest that these mice are more glucose tolerant using a dose of glucose based on their total body weight.
Due to the potential clinical implications of these results, it would be necessary to further evaluate the function of each p38 family member in browning to better understand how this pathway controls adipose tissue metabolism.
In summary, we have demonstrated that p38α and p38δ in adipose tissue have opposite roles: p38α negatively regulates BAT thermogenesis, energy expenditure, and body weight, while p38δ induces thermogenesis in BAT in response to several physiological stimuli. These results have potential clinical implications because inhibition of p38α or activation of p38δ might be of therapeutic interest against obesity.
This population study was approved by the Ethics Committee of the University Hospital of Salamanca and the Carlos III (CEI PI 09_2017-v3) with the all subjects providing written informed consent to undergo visceral fat biopsy under direct vision during surgery. Data were collected on demographic information (age, sex, and ethnicity), anthropomorphic measurements (BMI), smoking and alcohol history, coexisting medical conditions, and medication use.
All animal procedures conformed to EU Directive 86/609/EEC and Recommendation 2007/526/EC regarding the protection of animals used for experimental and other scientific purposes, enacted under Spanish law 1201/2005. The protocols are CNIC 08/13 and PROEX 49/13.
For the analysis of visceral fat, the study population included 71 patients (58 adult patients with BMI ≥35), while for the analysis of sWAT, the study population included 170 patients (140 adult patients with BMI ≥35), recruited from patients who underwent elective bariatric surgery at the University Hospital of Salamanca. Patients were excluded if they had a history of alcohol use disorders or excessive alcohol consumption (>30 g/day in men and >20 g/day in women) or had chronic hepatitis C or B. Control subjects (n = 13 for visceral fat study; n = 30 for sWAT study) were recruited among patients who underwent laparoscopic cholecystectomy for gallstone disease. Before surgery, fasting venous blood samples were collected for measuring complete cell blood count, total bilirubin, aspartate aminotransferase (AST), alanine aminotransferase (ALT), total cholesterol, high-density lipoprotein, low-density lipoprotein, triglycerides, creatinine, glucose, and albumin (S1 and S2 Tables).
Mice with a germ-line mutation in Mapk14 (p38α) and Mapk13 (p38δ) have been reported before [44, 45]. These animals were crossed with Tg (Fabp4-cre)1Rev/J [46] line or B6.FVB-Tg(Ucp1-cre)1Evdr/J [24] on the C57BL/6J background (Jackson Laboratory) to generate the mice lacking p38α or p38δ in adipose tissue (both WAT and BAT or just in BAT, respectively). All mice were maintained on a C57BL/6J background (back-crossed 10 generations). Genotype was confirmed by PCR analysis of genomic DNA. Mice were fed with an ND or an HFD, Research Diets Inc.) for 8 weeks ad libitum. For fat expansion measurement, mice were treated with BrdU (0.4 mg/ml; Sigma) in the drinking water (water was refreshed every 3 days) during the first week of a 6-week HFD. For temperature experiments, mice were housed at 30 °C for 8 weeks while feeding an HFD in case of thermoneutrality analysis. Mice were exposed to 4 °C for 1 hour, 1 day, or 1 week in case of cold adaptation studies.
Immortalised and primary brown preadipocytes from WT, Fab-Cre, p38αFab-KO, and p38δ-KO mice were differentiated to brown adipocytes in 10% FCS medium supplemented with 20 nM insulin, 1 nM T3, 125 μM indomethacin, 2 μg/ml dexamethasone, and 50 mM IBMX for 48 hours and maintained with 20 nM of insulin and 1 nM of T3 for 8 days. For some experiments, cultures were incubated with 100 nM T3 for 48 hours before extraction.
Immortalised white preadipocytes from Fab-Cre and p38αFab-KO mice were differentiated to adipocytes for 9 days in 8% FCS medium supplemented with 5 μg/ml insulin, 25 μg/ml IBMX, 1 μg/ml dexamethasone, and 1 μM troglitazone. For some experiments, cultures were incubated with 1 μM NE for 1 hour before extraction.
Primary brown preadipocytes were plated and differentiated in gelatin-coated (0.1%) 96 seahorse plates. MitoStress oxygen consumption rate (OCR) was assessed in XF medium containing 25 mM glucose, 2 mM L-glutamine, and 1 mM sodium pyruvate using a XF-96 Extracellular Flux Analyzers (Seahorse Bioscience, Agilent Technologies). Cells were stimulated with following drugs: NE or ISO, oligomycin, FCCP, and antimycin A plus rotenone (1 μM finally; all from Sigma Aldrich). The protocol for the all drugs followed a 3-minute mix, 2-minute wait, and 3-minute measure cycle that was repeated 3 times. After the analysis, data were normalised to protein level assessed by Bradford quantification. Basal Respiration Capacity (OCR basal − OCR nonmitochondrial) and oxygen consumption in response to NE (OCR NE − OCR basal) or ISO (OCR ISO–OCR basal) were calculated. For some experiments, cultures were pretreated with 10 μM BIRB796 for 1 hour.
Samples were lysed with RIPA buffer containing protease and phosphatase inhibitors (Tris-Hcl 50 mM [pH 7.5]; Triton X-100 1%; EDTA 1 mM [pH 8]; EGTA 1 mM; NaF 50 mM; β-glycerophosphate-Na 1 mM; sodium pirophosphate 5 mM; orthovanadate-Na 1 mM; sucrose 0.27 M; PMSF 0.1 mM; β-mercaptoethanol 1 mM; aprotinin 10 μg/ml; leupeptin 5 μg/ml). Lysates were separated by SDS-PAGE and incubated with antibodies diluted 1/1,000 against P-Akt308 (Cell Signaling, 9275s), P-Akt473 (Cell Signaling, 9271s), Akt (Cell Signaling, 9272s), UCP1 (Abcam, AB10983), P-ATF2 (Cell Signaling, 9225s), ATF2 (Cell Signaling, 9226s), P-CREB (Cell Signaling, 9198), CREB (Cell Signaling, 4820s), P-p38 (Cell Signaling, 9211s)—which recognises the phosphorylation in the activation sites of all the p38 isoforms—p38α (Santa Cruz, sc-535), P-AMPKα (Cell Signaling, 2531s), AMPKα (Cell Signaling, 2603s), P-ACC (Cell Signaling, 3661s), ACC (Cell Signaling, 3676s), PGC1α (Santa Cruz, sc13067), GAPDH (Santa Cruz, sc25778), tubulin (Sigma, T6199), and vinculin (Sigma, V9131), followed by an incubation with a secondary antibody conjugated with HRP. Reactive bands were detected by chemiluminescence and quantified by Image J software. Specificity of UCP1 antibody was evaluated using brown and eWAT from UCP1 KO animals [47].
For the immunoprecipitation assay, cell extracts were incubated with 4 μg of anti-p38 delta coupled with protein-G-Sepharose. After an overnight incubation at 4°C, the captured proteins were centrifuged at 10,000 g, the supernatants discarded, and the beads washed 4 times in lysis buffer. Beads were boiled for 5 minutes at 95 °C in 10 μl sample buffer. The antibodies employed were anti-phospho p38 and anti-p38δ (Santa Cruz, sc7585). Immune complexes were detected by enhanced chemiluminescence (NEN).
Mouse bone marrow (BM) and spleens were collected, and single-cell suspension was obtained. Erythrocytes were lysed with a red cell lysis buffer incubation for 3 minutes on ice. Spleen samples were enriched using CD3 (BioLegend 79751 clone 145-2C11) and B220 (BioLegend 79752 clone RA3-6B2) biotinylated antibodies and magnetic Dynabeads Myone streptavidin T1 (invitrogen). Myeloid cells from spleen were labelled by surface staining with FITC-conjugated CD11b (BioLegend 79749 clone M1/70), PE-conjugated Gr1 (Ly6G/Ly6C) (BDBioscience 79750 clone RB6-8C5), and APC-conjugated F4/80 (eBiosciences 25-4801-82 clone BM8) antibodies, and myeloid cells from BM were labelled by FITC-conjugated Gr1 (Ly6G/Ly6C) (Invitrogen 11-5931-82 clone RB6-8C5), PE-conjugated CD115 (eBioscience 12-1152-82 clone AFS98), and APC-conjugated F4/80 (eBiosciences 25-4801-82 clone BM8) antibodies. Nuclei were stained with DAPI. Cells were sorted with a fluorescence-assisted cell sorting (FACS) Aria (BD) as follows: spleen macrophages (Gr1− Cd11bmedium F4/80+), spleen neutrophils (Gr1high Cd11b+), and BM monocytes (CD115+ F4/80−). Isolated myeloid cells were lysed and analysed by western blot.
Overnight-starved mice were injected intraperitoneally with 1 g/kg of body weight of glucose, and blood glucose levels were quantified with an Ascensia Breeze 2 glucose meter at 0, 15, 30, 60, 90, and 120 minutes post injection. Alternatively, GTT was performed injecting intraperitoneally 1 g/kg of lean mass of glucose.
ITT was performed by injecting intraperitoneally 0.75 IU/kg of insulin at mice starved for 1 hour and detecting blood glucose levels with a glucometer at 0, 15, 30, 60, 90, and 120 minutes post injection.
Energy expenditure, respiratory exchange, and food intake were quantified using the indirect calorimetry system (TSE LabMaster, TSE Systems, Germany) for 3 days.
Body temperature was detected by a rectal thermometer (AZ 8851 K/J/T Handheld Digital Thermometer-Single, AZ Instruments Corp., Taiwan).
BAT-adjacent interscapular temperature was quantified by thermographic images using a FLIR T430sc Infrared Camera (FLIR Systems, Inc., Wilsonville, OR) and analysed through FlirIR software.
Body, fat, and lean mass were quantified by nuclear magnetic resonance (Whole Body Composition Analyzer; EchoMRI, Houston, TX) and analysed by ImageJ software.
Blood triglyceride content was quantified using a Dimension RxL Max analyser (Siemens). For triglyceride analysis in cells, brown adipocyte cultures were lysed in isopropanol, centrifuged at 10,000 g for 15 minutes at 4 °C, and triglycerides were detected in the supernatant with a commercial kit (Sigma).
Brown adipocyte cells were scraped in PBS and pellet lysed in TNES buffer supplemented with Proteinase K (20 mg/ml) overnight at 55 °C. Reaction was stopped with sodium chloride 6 M and samples centrifuged 5 minutes at 13,000 g. DNA was precipitated in supernatants with 100% ethanol and washed with 70% ethanol. After drying, DNA was resuspended in DNase free water, quantified, and analysed by RT-PCR. Mitochondrial DNA was detected using primers for COII and nuclear DNA, using primers for Sdh1 (S3 Table).
RNA 500ng—extracted with RNeasy Plus Mini kit (Quiagen) following manufacturer instructions—was transcribed to cDNA, and qRT-PCR was performed using Fast Sybr Green probe (Applied Biosystems) and the appropriated primers in the 7900 Fast Real Time thermocycler (Applied Biosystems). Relative mRNA expression was normalised to Gapdh mRNA measured in each sample. Primers used are listed in S3 Table.
Fresh livers, brown, and epididymal white fat were fixed with formalin 10%, included in paraffin, and cut in 5 μm slides followed by a haematoxylin–eosin staining.
Fat droplets were detected by oil red staining (0.7% in propylenglycol) in 8 mm slides included in OCT compound (Tissue-Tek) and in differentiated brown and white adipocytes.
Brown adipocytes were stained with Mito Tracker Deep Red (Invitrogen) and Bodipy (Invitrogen). Images were captured using Leica SPE confocal microcope (Leica Microsystems, Wetzlar, Germany).
For UCP1 immunostaining, brown and epididymal white fat were fixed with formalin 10%, included in paraffin, cut in 5 μm slides, and sequentially stained with a UCP1 antibody (1/500, Abcam, AB10983), a biotinylated goat anti-rabbit secondary antibody (1/500, Jackson Immuno Research Laboratories), a streptavidin-conjugated ABC complex (Vector Laboratories), and the substrate 3,3’-diaminobenzidene conjugated with horseradish peroxidase (Vector Laboratories), followed by brief counterstaining with Nuclear Fast Red haematoxylin (Sigma).
For immunofluorescence analysis, the 5 μm tissue sections were deparaffinised and rehydrated, followed by antigen retrieval in 10 mM sodium citrate (pH 6.0) under pressure in a CertoClav EL (CertoClav Sterilizer GmbH). For BrdU staining, sections were treated with DNase 30 minutes at 37 °C. Blocking and staining was performed in 5% BSA in PBS. Sections were incubated in primary antibodies including rat-anti-Ki67 (eBioscience, 14-5698-82; clone:SolA15) (1:100), rabbit-anti-GLUT4 (Abcam, ab654) (1:1000), mouse-anti-Caveolin-1 (Sigma, SAB4200216) (1:500), rat anti-BrdU (Abcam, Ab6326; clone: BU1/75 [ICR1]) (1:200), and rabbit anti-Perilipin (Cell Signaling, 9349; clone: D1D8) (1:400) overnight at 4 °C. Secondary antibodies including goat anti-rabbit-A488, goat anti-rat-A647, and chicken anti-mouse-A647—all used at 1:500—were purchased from Molecular Probes and incubated with tissue for 1 hour at room temperature. Nuclei were stained with DAPI, and slides were mounted with Vectashield mounting medium (Vector Laboratories) and examined using SP5 multi-line inverted confocal microscope. Several confocal images of each tissue section were acquired and analysed for the translocation of GLUT4 or the presence of Ki67 or BrdU in adipocyte nuclei. BAT and WAT cellularity were quantified using Fiji software. Adipocyte nuclei were identified by their location inside adipocyte membranes as described [23].
Results are expressed as mean ± SEM. Statistical analysis was evaluated by student t test and 2-way ANOVA coupled with Bonferroni’s post-tests with values of p < 0.05 considered significant. When variances were different, Welch’s test was used. For human studies, variables were compared by means of Mann-Whitney U test or χ2 test.
|
10.1371/journal.pgen.1003244 | A Quartet of PIF bHLH Factors Provides a Transcriptionally Centered Signaling Hub That Regulates Seedling Morphogenesis through Differential Expression-Patterning of Shared Target Genes in Arabidopsis | Dark-grown seedlings exhibit skotomorphogenic development. Genetic and molecular evidence indicates that a quartet of Arabidopsis Phytochrome (phy)-Interacting bHLH Factors (PIF1, 3, 4, and 5) are critically necessary to maintaining this developmental state and that light activation of phy induces a switch to photomorphogenic development by inducing rapid degradation of the PIFs. Here, using integrated ChIP–seq and RNA–seq analyses, we have identified genes that are direct targets of PIF3 transcriptional regulation, exerted by sequence-specific binding to G-box (CACGTG) or PBE-box (CACATG) motifs in the target promoters genome-wide. In addition, expression analysis of selected genes in this set, in all triple pif-mutant combinations, provides evidence that the PIF quartet members collaborate to generate an expression pattern that is the product of a mosaic of differential transcriptional responsiveness of individual genes to the different PIFs and of differential regulatory activity of individual PIFs toward the different genes. Together with prior evidence that all four PIFs can bind to G-boxes, the data suggest that this collective activity may be exerted via shared occupancy of binding sites in target promoters.
| An important issue in understanding mechanisms of eukaryotic transcriptional regulation is how members of large transcription-factor families, with conserved DNA–binding domains (such as the 162-member Arabidopsis bHLH family), discriminate between target genes. However, the specific question of whether, and to what extent, closely related sub-family members, with potential overlapping functional redundancy (like the quartet of Phytochrome (phy)-Interacting bHLH transcription Factors (PIF1, 3, 4, and 5) studied here), share regulation of target genes through shared binding to promoter-localized consensus motifs does not appear to have been widely investigated. Here, using ChIP–seq analysis, we have identified genes that bind PIF3 to conserved, sequence-specific sites in their promoters; and, using RNA–seq, we have identified those genes displaying altered expression in various pif mutants. Integration of these data identifies those genes that are likely direct targets of transcriptional regulation by PIF3. Our data suggest that the PIF quartet members share directly in transcriptional activation of numerous target genes, potentially via redundant promoter occupancy, in a manner that varies quantitatively from gene to gene. This finding suggests that these PIFs function collectively as a signaling hub, selectively partitioning common upstream signals from light-activated phys at the transcriptional-network interface.
| A key component of the successful colonization of land by terrestrial flowering plants was the evolution of a developmental strategy termed skotomorphogenesis (etiolated growth). This strategy enabled post-germinative seedlings emerging from buried seed to grow heterotrophically, on seed reserves, rapidly upwards through the subterranean darkness to the soil surface. Coupled with this was the evolution of a photosensory mechanism to trigger a switch to autotrophic, photomorphogenic (deetiolated) development upon emergence into sunlight.
Genetic evidence indicates that a small subfamily of basic helix-loop-helix (bHLH) transcription factors, termed PIFs (for Phytochrome (phy)-Interacting Factors) are centrally critical to the promotion of such skotomorphogenic development in dark-grown seedlings [1]. A quadruple pif mutant (pifq), lacking PIF-family members PIF1, PIF3, PIF4 and PIF5 (termed the PIF quartet), displays morphogenic development in total darkness that strongly phenocopies that of normal light-grown seedlings [2], [3]. This observation establishes that these factors act constitutively to promote skotomorphogenic development and that their absence induces the switch to photomorphogenic development. All four quartet members have been shown individually to bind preferentially to a core G-box DNA-sequence motif (CACGTG) (a variant of the canonical E-box motif (CANNTG)) [4]–[9], and to function as transcriptional activators in transfection or heterologous systems [4]–[7], [10]. Because monogenic mutants at each of these loci have no, or minimal, visible effects on skotomorphogenesis, and the various double and triple pif-mutant combinations progressively exhibit increasingly photomorphogenic phenotypes in darkness, it appears that the PIF quartet members act with partially additive or overlapping redundancy to drive the skotomorphogenic pathway [2], [3], [11]–[13].
The phy family of sensory photoreceptors (especially phyA and phyB) has a central role in inducing the switch from skotomorphogenic to photomorphogenic development (deetiolation) in response to initial exposure of dark-grown seedlings to light [1], [14], [15]. The existing evidence indicates that this is achieved in large part by rapid phy-triggered degradation of the PIF proteins. The mechanism underlying this process involves the rapid, light-induced translocation of the activated (Pfr) conformer of the phy molecule from the cytoplasm into the nucleus, where it physically interacts with PIF-quartet members. This interaction induces phosphorylation of the PIF proteins which in turn triggers ubiquitylation and proteolytic degradation of the transcription factors (half-lives of 5–20 min) via the proteasome system. The altered transcriptional landscape resulting from the consequent robust reduction in steady-state abundance of these factors is the major driving force in the switch from heterotrophic to autotrophic development inherent in the deetiolation process.
A limited number of transcriptome analyses, using Affymetrix ATH1 microarrays, aimed at identifying genes regulated by the phy-PIF signaling pathway during deetiolation have been reported [3], [12], [16]–[19]. The data show that 80% of the genes that display altered expression in the pifq mutant in the dark are normally altered by prolonged light in fully deetiolated wild-type (WT) seedlings [17], but that only a relatively small fraction of these are misexpressed in dark-grown pif1 [19], pif3 [16]–[18] and pif4pif5 mutants [12]. These results affirm the central collective regulatory function of these four PIFs in regulating the overall transcriptional network that drives the developmental switch from skotomorphogenesis to photomorphogenesis, and provide initial indications of functional redundancy at the gene expression level. These genes could be either direct or indirect targets of PIF transcriptional regulatory activity [20]. Identification of those genes that respond rapidly (within 1 h) to initial light exposure has defined a subset of PIF-regulated genes that are likely to be enriched for loci that are directly transcriptionally regulated by the PIF-quartet proteins [17]. PIF-regulated genes that conversely respond rapidly to vegetative shade in fully-green, light-grown plants have also been identified by microarray-based expression profiling [11], [21]. It is notable that these early-response genes are enriched for transcription-factor-encoding loci, suggesting a potential hierarchal network that drives a transcriptional cascade. However, rapid responsiveness alone obviously does not establish that transcriptional regulation is direct.
The advent of ChIP-chip and ChIP-seq technology has provided the opportunity to identify genes that contain binding sites for transcription factors of interest, on a genome-wide scale [20], [22], [23]. When combined with full transcriptome analysis, the data provide identification of genes that are direct targets of transcriptional regulation by the factor(s) under study. A number of such studies have recently been reported for a diversity of factors in Arabidopsis, using either ChIP-chip or ChIP-seq analysis of factor binding sites, coupled predominantly with Affymetrix ATH1 microarrays (representing about 80% of the protein-coding genes in the genome) for expression analysis [21], [23]–[29]. These data have begun to provide insight into the complexity of the transcriptional networks that coordinate a variety of the fundamental processes underlying plant growth and development.
Despite these advances, the use of the ATH1 microarray for expression analysis in many of these studies means that important expression changes in genes not present on this array might have been missed. In addition, the question of whether, and to what extent, closely related transcription-factor family members, such as the PIF quartet, with apparently shared DNA-target-sequence specificity, contribute toward the transcriptional regulation of common target genes does not appear to have been addressed in many existing studies of eukaryotic systems [27], [30]–[34], although a recent report by Hornitschek et al shows differential binding of recombinant PIF4 and PIF5 to various E-box variants in vitro using protein-binding microarrays, as well as shared binding in vivo to four selected promoters using ChIP-PCR analysis [21]. Here, using ChIP-seq analysis, we have identified PIF3-binding sites, genome wide, and, in parallel, using RNA-seq analysis of selected pif-mutant lines, we have defined the genes regulated by PIF3, genome-wide, in dark-grown seedlings. By merging these datasets, we have identified those genes whose expression is, at least partially, directly regulated by promoter-bound PIF3. In addition, by profiling the expression of a selected subset of these direct PIF3-targets in multiple additional pif-mutant combinations, we have addressed the question of whether PIF1, 3, 4 and 5 display qualitative and/or quantitative functional divergence in regulating shared target genes.
Two-day-old dark-grown wild-type (WT) and MYC-epitope-tagged-PIF3 (P3M)-expressing, pif3-3 null-mutant seedlings were used for ChIP-seq analysis. DNA prepared from MYC-antibody-generated immunoprecipitates from four independent biological replicates of each genotype was subjected to high-throughput sequencing. Statistically-significant binding peaks were defined by comparing the parallel P3M and WT ChIP samples within each replicate using the MACS algorithm [35]. Replicate-specific peaks (Table S1) were defined as reproducible if they were identified at the same genomic location in two or more biological replicates (overlapping Venn sectors in Figure 1A; also Table S2). For each reproducible peak, we assigned a common summit as the mean of the individual replicate-specific summits.
This analysis identified 1064 reproducible peaks which form our “high-confidence” set of PIF3-binding sites (Table S2). These sites are evenly distributed on the five chromosomes and 89% are located in intergenic regions (Figure 1B). In ChIP-qPCR validation assays, all but 1 of the 38 tested regions exhibited strong binding enrichment in the P3M samples compared to the WT controls (Figure S1), indicating a low false positive rate for our ChIP-seq procedure.
Using the MEME program [36], we performed de novo motif discovery on the +/−100 bp regions surrounding the 1064 “high-confidence” PIF3 binding-peak summits described above. Two E-box (CANNTG) variants were identified as statistically overrepresented motifs within these PIF3-binding regions (Figure 1C). The CACGTG (‘G-box’) variant is well-established as a preferred PIF-binding motif [4]–[9]. By contrast, the CACATG variant is previously undescribed as a PIF3-binding motif, although PIF1 [37] and, recently, PIF4 [21] have been reported to bind. We conclude that this variant is a strong candidate for being a general alternative binding motif for PIF3 across the genome, and define it, therefore, as the PIF-binding E-box (PBE-box). The relative distribution of these two motifs across the 1064 PIF3 binding-sites is summarized in Figure 1D. A majority (73%) of the sites contain one or both motifs (G-box 50% and PBE-box 36%, with 13% overlap) within the 200-bp window.
A broader analysis shows that 64% of the G-box and 30% of the PBE-box motifs present in the 2 kb windows surrounding the PIF3-binding summits cluster within the designated 200-bp binding sites (Figure 1E). Similarly, both motifs are strongly enriched in these 200-bp windows compared to random 200-bp genome segments, and this enrichment increases toward the PIF3-binding summit (Figure 1F). These data establish the highly significant coincidence between PIF3 and these two specific cis-elements.
To examine the potential direct interaction of PIF3 with the newly-identified PBE-box compared to that of the G-box, we performed DNA-Protein-Interaction (DPI)-ELISA [38]. We tested the binding of PIF3 to several G-box- (PIL1, PHYB, and RGA1) or PBE-box- (IAA2, IBH1, and AT4G30410) containing probes generated from various genomic PIF3-binding sites identified in the ChIP-seq analysis. Figure 1G shows that recombinant PIF3 binds sequence-specifically to all G-box- and PBE-box-containing probes, although the apparent affinity for the G-box seems overall to be higher than for the E-box. An EMSA analysis showed similar results (Figure S2). These in vitro binding-assay data indicate that the coincidence of PIF3-binding sites with the G-box or PBE-box motif in the ChIP-seq assay likely results from their direct interaction in vivo, and that the PBE-box is indeed another sequence-specific PIF-binding, E-box variant genome wide. Because all of the PIF3-binding sites tested by ChIP-qPCR in Figure S1 contain coincident G- or PBE-box motifs, these data validate the in vivo-binding of PIF3 to these motifs.
The binding of PIF3 to the ATHB-2 probe, which contains one G-box and one PBE-box, provides an interesting insight. Neither the competitor mutated in both motifs (Figure 1G; also Figure S2), nor the competitor mutated only in the G-box motif (Figure S3) displayed competitive activity, whereas the probe mutated only in the PBE-box motif showed competitive efficiency similar to that of the WT sequence (Figure S2). These findings suggest that PIF3 may have differential binding affinity toward these two motifs in specific genomic contexts.
Although all ChIP-defined transcription factor binding sites may prove to be functionally significant, we have chosen here to focus on identifying those genes displaying motif-coincident PIF3-binding sites located in conventional promoter regions (defined here as “PIF3-bound genes”). Initially, from the 1064 binding sites defined above, we identified 709 sites that are both intergenic and G- and/or PBE-box-coincident (Table S2). For these 709 sites, we defined PIF3-bound genes as having a binding site in the 5′ flanking DNA, within 5 kb of the transcription start site (TSS), in the absence of intervening genes. This analysis identified 596 PIF3-binding sites, with 828 associated genes, where some sites are associated with two genes on opposite strands. Of these genes, 88% have PIF3-binding sites within 3 kb of TSS, whereas the remaining 12% have sites between 3 and 5 kb upstream (Table S3). These 828 genes thus constitute a set of PIF3-bound genes whose transcription is potentially directly regulated by PIF3.
To provide genome-wide visualization of the ChIP-seq analysis, we developed a platform using the Integrated Genome Browser [39]. Figure 2A shows the chromosomal regions around PIL1 and ATHB-2, as examples. The chromosomal region surrounding the PIL1 gene shows a single PIF3-binding peak that is coincident with three G-box motifs located in the PIL1 promoter region. ATHB-2 is somewhat unusual in that it displays five specific PIF3-binding peaks in its extensive 5′-upstream region, each coincident with 1 to 3 G-box motifs (Figure 2A). ChIP-qPCR analysis scanning across the PIL1 genomic region provides robust validation of the ChIP-seq data for this gene (Figure 2B).
To identify the genes regulated by PIF3, genome-wide, in the promotion of skotomorphogenic development, we performed 3′-end-capture directional RNA-seq analysis, comparing the expression profiles of 2-d dark-grown WT, pif3, pif1pif4pif5 (pif145) and pif1pif3pif4pif5 (pifq) Arabidopsis seedlings. Genes displaying Statistically-Significant Two-Fold (SSTF) expression changes in the three mutant genotypes compared to the WT and each other were identified as being regulated by the relevant mutated PIF(s) (Figure 3; listed in Tables S4, S5, S6, S7, S8). The degree of overlap between SSTF genes identified in these comparisons is depicted in the Venn diagrams in Figure 3B, 3C and 3D. We defined a combined total of 345 genes in the pif3/WT and pifq/pif145 gene-sets as the composite PIF3-regulated gene-set (Table S9). Similarly, a combined total of 1454 genes in the pif145/WT and pifq/pif3 gene-sets were defined as the composite PIF1/4/5-trio regulated gene-set (Table S10). Comparison of these composite gene-sets is displayed in Figure 3D.
The data indicate that 254 (74%) PIF3-regulated genes are also redundantly transcriptionally regulated by one or more of the other three PIF-quartet proteins. Conversely, 1740 (86%) PIF-quartet-regulated genes show no significant PIF3 dependence (Figure 3B), whereas 918 (45%) do display PIF1/4/5 regulation (Figure 3D), indicating that one or more of the other PIF-quartet members function non-redundantly with PIF3 in regulating the expression of many target genes. The general robustness of our genome-wide RNA-seq expression profiling is demonstrated by the extensive RT-qPCR validation data presented in Figure S3.
To identify the genes that both physically bind PIF3 in their promoters, in a G-box- or PBE-box-coincident manner, and display PIF-regulated expression (defined here as “direct-target genes”), we merged our ChIP-seq and RNA-seq data. This permitted gene-by-gene visualization of the PIF3-binding peaks and PIF-dependent transcription, genome-wide, as shown for PIL1 and ATHB2 in Figure 2A. The expression data for these two genes show a clear difference in transcript levels between the WT and pifq mutant, demonstrating the robust dependence of full expression on the presence of the PIF-quartet. Comparison of the expression peaks for the pif145 and pifq mutants also suggests that PIF3 acts in the absence of the other three quartet members to promote a moderate increase in transcript levels. Overall, the combined PIF-regulated expression-patterns and promoter-located PIF3-binding sites displayed by these two genes render them likely direct-targets of transcriptional regulation by PIF3 and one or more other quartet members in promoting skotomorphogenic development.
The Venn diagrams in Figure 4 show the genome-wide overlap of the genes identified independently as displaying PIF-quartet- and/or PIF3-dependent expression in a SSTF manner, with those exhibiting promoter-located, motif-coincident PIF3-binding sites (Figure 4A, Classes X, Y and Z; listed in Table S11). By these criteria, a total of 22 genes (Classes X and Z) were identified as robustly-likely, direct-target genes of autonomous-PIF3 transcriptional regulation. Of these, 21 genes (19 PIF3-induced; 2 PIF3-repressed) also display collective PIF-quartet regulation (Class Z). The 19 PIF3-induced Z-Class genes are listed in Table 1. The bar graphs in Figure 4B and 4C portray the mean expression level (relative to WT) of all the genes in each class, for each pif genotype. The quantitatively robust responsiveness of the PIF3-bound genes to the presence of PIF3 in the pif145 mutant background is evident from these data (Class Z-associated bar graphs). This robust PIF3-responsiveness was validated using RT-qPCR for selected members of the 19 PIF3-induced, Z-Class gene-set, having a range of quantitative dependence on this bHLH factor (Figure S3A).
A striking feature of our data is the relatively large number of PIF-quartet-regulated genes (107 total genes; 88 PIF-induced and 19 PIF-repressed) that display promoter-located, PIF3-binding sites, but lack evidence of SSTF-level PIF3 regulation in our RNA-seq analysis (Figure 4, Class Y; also Table S11). Nevertheless, the bar graphs of the mean expression of these genes suggest a tendency toward a consistent difference in expression between the pif145 and pifq mutants, across the gene-set. In addition, combined analysis of the full set of PIF-induced, PIF3-bound genes (Classes Y and Z together) shows that there is a reciprocal continuum in the magnitude of the relative contributions of PIF3 and the PIF1/4/5-trio to the collective activity of the PIF quartet in transcriptionally activating these genes (Figure S4). To more closely examine the PIF3 contribution to the total PIF-quartet activity in the Y-Class genes, we therefore arrayed these genes by the pif145/pifq fold-change value and assayed the relative expression levels in 20 selected PIF-induced loci by RT-qPCR. Figure S3B shows that the 17 genes in this group with fold-changes >1.5 by the RNA-seq analysis, all exhibit statistically-significant (Student's t-test, P<0.05), PIF3-promoted expression increases in the pif145 mutant compared to the pifq mutant by RT-qPCR. This suggests that a subset of Y-Class genes may represent additional bona fide autonomously-PIF3-regulated genes that are below the resolution of the SSTF criteria we imposed on our RNA-seq analysis. We have therefore designated these 17 as Class YZ1.5 genes, having moderate (>1.5-fold), but statistically significant, regulation by PIF3 (Table 1). The evidence indicates, therefore, that a combined total of at least 38 YZ1.5- and Z-Class genes are direct targets of moderate to robust transcriptional regulation by promoter-bound PIF3. Because an additional 34 of the 88 Y-Class genes also display >1.5-fold PIF3-induced expression (Figure S3B and Table S11), it is possible that the number of direct targets of partial PIF3 transcriptional regulation is yet larger.
The W-class genes are those that display promoter-localized, G- or PBE-box-coincident PIF3-binding peaks, but no differential expression between the pifq mutant and wild type (Figure 4A). This observation is consistent with data from a variety of organisms that have shown that transcription factors vary greatly in their number of genomic binding sites, and that binding events can vastly exceed the number of known or possible direct gene targets [40]. The reasons for this phenomenon here are unclear but could include functional redundancy with other factors, including other PIF proteins. Consistent with this possibility, a subset of 41 of the total 699 W-class genes exhibit rapid light responsiveness [17] upon initial exposure (Table S12).
The reciprocal continuum in relative PIF3 and PIF1/4/5-trio contributions to the collective PIF-quartet transcriptional activation of PIF-induced, Y- and Z-class genes referred to above (Figure S4), indicates that PIF1, 4 and/or 5 contribute substantially to the regulation of these PIF3-bound genes. To identify the individual genes in this set displaying a significant PIF1/4/5 contribution, we compared the Y- and Z-class genes (Table S11) with those defined above as PIF1/4/5-regulated (Figure 3D; also Table S10). Overall, 92 (72%) of the 128 combined Y- and Z-class genes exhibit regulation by PIF1, 4 and/or 5 (Table S11), as shown by significant differences in the pif145/WT and/or pifq/pif3 comparisons. More notably, all 38 PIF3-induced direct-target genes (Class YZ1.5 and Z) are also PIF1/4/5-induced (Table 1). Because all four PIFs have been shown to bind to the G-box motif in sequence-specific fashion [4]–[9], it appears probable that these PIF-quartet-regulated genes, displaying promoter-located PIF3-binding sites (Figure 4; also Table S11), may be directly regulated by one or more of the other quartet members, in addition to, or instead of, PIF3.
Categorization of the YZ-class genes by the known or predicted functions of their encoded products reveals substantial enrichment in a diversity of transcription-factor-encoding genes (Table 1; also Figure S5 and Table S11), consistent with the concept that these multiple direct targets of the PIF quartet function at the apex of a primary transcriptional-cascade to regulate the downstream transcriptional network. It is also notable, however, that a considerable number of the YZ-class genes that have other cellular functions are also apparent direct targets of transcriptional regulation by the PIFs, including two non-protein-encoding genes of unknown function (Table 1).
Previously, by microarray profiling, we identified a subset of genes (designated Class 7) that, in dark-grown seedlings, exhibit a PIF-quartet-dependent expression pattern, that is rapidly reversed (within 1 h) upon initial exposure to phy-activating R light [17]. Of the 24 rapidly light-repressed Class 7 genes displaying promoter-localized, G- or PBE-box-coincident PIF3-binding peaks, 21 (88%) are either Class YZ1.5 or Z genes here (Table 1). These genes are thus identified as a subset whose expression is directly promoted, at least partially, by PIF3 transcriptional activation in the dark, and is rapidly reduced in the light, at least in part, by photoactivated-phyB-induced PIF3 degradation. It is notable that 9 of these genes (43%) encode transcription factors (Table 1), indicative of being master regulators at the apex of the downstream transcriptional cascade controlled by the phy signaling pathway.
In striking contrast to the light-repressed Class 7 genes, only 7 of the 115 rapidly light-induced Class 7 genes (6%) [17] display PIF3-binding peaks that are coincident with a G-box or PBE-box, and of these only 2 genes (<2%) (PSY and KAI2) exhibit derepression here in the dark-grown pifq mutant. No individual PIF3 contribution to this repression was detectable here. Collectively, these data indicate that PIF3 acts predominantly, if not exclusively, to activate the expression of direct-target genes in dark-grown seedlings. Conversely, the 94% (108/115) of light-induced Class 7 genes that do not display G- or PBE-box coincident PIF3-binding peaks, might suggest that one or more of the proposed direct targets of the PIF quartet (Table 1) can act as key repressor(s) that regulate a diverse set of light-induced genes.
Recently, we defined a small core set of 14 Class 7 PIF-quartet-regulated genes (called M-Class genes) that display rapid, reciprocal, transcriptional responsiveness to light and vegetative shade in dark-grown and light-grown seedlings, respectively [11]. Our present analysis shows that 11 (79%) of these M-Class genes are identical to those identified here as dual Class 7 and YZ1.5-/Z-Class genes (Table 1), indicating that they are likely direct targets of PIF3 regulation, not only during skotomorphogenesis and deetiolation, but also subsequently, on a continuing basis, through juvenile vegetative development. The correlated PIF3-binding and PIF-regulated transcriptional behavior of several of these M-Class genes, determined by merging the ChIP-seq and RNA-seq data, is depicted in Figure S6.
Because our data indicate that the contribution of PIF3 to the total level of expression collectively regulated by the PIF-quartet is quantitatively variable between genes (Figure 2A; Figure 4B, 4C; Figures S3 and S4), we wished to determine whether the other members of the PIF-quartet display a similarly variable pattern of regulation. For this purpose, we assayed the expression by RT-qPCR of a selected set of apparently PIF3 direct-target genes (Classes Z and YZ1.5), in the four different pif triple mutants compared to the pifq mutant and WT (Figure 5A). The relative autonomous contribution of each individual PIF (in the absence of the other three quartet members) to the total, collective PIF-quartet-supported expression was calculated as a percentage of the total difference in expression between the WT and pifq mutant, for each separate gene. The data reveal a striking diversity of relative contributions, both between the individual PIFs, and between genes for any individual PIF, in two-dimensional-matrix fashion (Figure 5B). Particularly notable is the dominant role played by PIF1 in promoting the expression of the majority of these genes. On the other hand, PIF3 contributes strongly to ARF18, SNRK2.5 and BBX28 expression, PIF4 strongly activates ST2A and ATHB-2, PIF5 contributes actively to AT5G02580, ATHB-2 and XTR7 expression, while all four PIFs contribute substantially to IAA19 transcription. Because all these tested genes are prospective direct-target genes of multiple PIF-quartet members, our findings suggest that there is an intricate combinatorial network, in which the individual PIF-quartet factors collaborate to transcriptionally regulate an array of direct-target genes, through potentially common DNA binding sites, with quantitatively differential regulatory activity.
Comparison of our data with recently published ChIP-seq-identified PIF4- and PIF5-binding sites [21], [24] supports this conclusion (Figure S7). Although the three studies were performed under contrasting experimental conditions, our analysis shows that 82% of genes with promoter-located, G- or PBE-box-associated PIF3-binding peaks identified here, also display PIF4- and/or PIF5-binding peaks (Figure S7A). Perhaps more striking, 89% of the 128 PIF3-binding, PIF-quartet-regulated genes identified here (Y- and Z-class genes in Figure 4A), are also bound by PIF4 and or PIF5, with 52% being bound by all three PIFs (Figure S7B; Table S11).
One possible mechanistic basis for the differential control of shared target genes by the individual PIF-quartet members described above (Figure 5) is that each PIF transcription factor has a different spatial expression pattern across the plant. To examine this possibility, we expressed pPIF:GUS fusions for each of the PIF genes transgenically in Arabidopsis seedlings, and assayed the distribution of GUS expression histochemically. The data show that all four PIF promoters support expression broadly throughout the seedling shoot tissue, with largely similar distribution patterns between the quartet members, within the resolution of this procedure (Figure 6A–6D).
In principle, differences in absolute expression levels among the PIF-quartet members could also be a fundamental determinant of differences in PIF-promoted expression of target genes (Figure 5). However, this does not appear to be the case here. Examination of the RNA-seq profiles, and independent RT-qPCR analyses, of PIF1, PIF3, PIF4 and PIF5 expression, shows that, while there are marked differences in expression between these genes in wild-type seedlings (Figure 6F), these are not strongly correlated with the respective patterns of target-gene expression (Figure 5). In particular, PIF1 and PIF3 display expression levels that are robustly converse to their respective general levels of transcriptional activation. Similarly, and more importantly, although the expression levels of PIF4 and PIF5 are significantly elevated in the relevant triple mutant compared to wild-type (Figure 6G), these differences also do not correlate with the overall differential expression patterns of the target genes. While these elevated levels could indicate that the computation in Figure 5B overestimates the normal, relative contributions of these two PIFs to the collective PIF-quartet activity, displayed when all four PIFs are present, they do not account for the apparent dominance of PIF1 or the diversity of response-patterns between the genes.
Taken together, these results suggest that the sometimes strikingly different quantitative contributions of the individual PIFs to the expression of a given target gene appears unlikely to be primarily due to either differences in transcriptionally-driven PIF abundance or differences in spatially-determined abundance of the PIF-quartet-members. It appears more likely that these differences are due to intrinsic differential activities of the individual PIFs in the context of the individual target-gene promoters. In addition, because the GUS expression pattern driven by the CaMV 35S promoter (Figure 6E) overlaps substantially with that driven by the PIF promoters (Figure 6A–6D), it seems reasonable to expect that the majority of PIF3-binding sites detected by ChIP-seq analysis here, using 35S-driven PIF3-Myc expression, will reflect sites that are normally available to PIF3 generated by endogenous PIF3 promoter activity.
The robust binding of PIF3 to the G-box-containing region of the PIL1 promoter detected by ChIP-seq analysis (Figure 2) and in vitro assay (Figure 1G and Figure S2), and the partial autonomous promotion of PIL1 expression by PIF3 observed by RNA-seq analysis (Figure 2A, Figure 5, and Figure S3A), provides strong evidence that PIL1 is a direct target of PIF3 transcriptional regulation via physical interaction of the bHLH factor with these cis-elements. Conversely, because the non-PIF3 quartet members contribute robustly to the collective PIF-quartet-dependent expression of PIL1 (Figure 2A and Figure 5), and given that these non-PIF3 members also bind selectively to G-box motifs [4]–[9], it might be predicted that PIF1, 4 and/or 5 transcriptional activation of PIL1 will, like PIF3, be exerted through interaction with the G-boxes in the PIL1 promoter [5], [6]. To examine this prediction, we tested the functional necessity of these G-box motifs to PIL1 expression using reporter constructs in transgenic seedlings. The data show that activation of the PIL1 promoter requires both the presence of one or more of the PIF quartet and one or more of the G-boxes (Figure 5C), indicating that the G-box elements are the major, if not sole, targets of PIF-quartet transcriptional activation activity. By contrast, it is notable that, although a recent report shows that PIF7 also binds to the G-box region of the PIL1 promoter in a manner that is functionally important for shade-induced expression of this gene in light-grown seedlings [41], the extremely low residual levels of PIL1 expression in dark-grown pifq seedlings compared to wild-type (Figure 2A and Figure 5A) indicate that PIF7 has minimal, if any, contribution under these conditions. Together, the evidence suggests that, to the extent that the PIF-quartet members share transcriptional activation of PIL1 (Figure 5A and 5B), they do so by sharing the G-box motifs as interaction sites. This conclusion is consistent with the demonstration that PIF3 binds to all three G-box motifs in the PIL1-promoter cluster, both in vivo (Figure 2B) and in vitro (Figure 1G and Figure S2). By extrapolation, the other Y- and Z-Class, PIF-quartet-regulated genes, established here as being direct-targets of PIF3 transcriptional regulation through G- or PBE-box binding motifs, are strong candidates for likewise being targets of functionally active, direct binding-site sharing among the four PIF factors.
Previous genetic studies have established that the overarching biological function of the PIF quartet is to promote skotomorphogenic growth and development in post-germinative seedlings in darkness, and to promote shade-avoidance behavior in deetiolated seedlings in response to exposure to neighboring vegetation [2], [3], [11], [42]. The evidence shows that the quadruple pifq mutant is strongly impaired in skotomorphogenic growth and development in dark-grown seedlings and has reduced shade-avoidance responsiveness to signals from neighboring vegetation in green seedlings [2], [3], [11]. In addition, there are indications that the contributions of individual PIF members to the collective activities of the quartet vary quantitatively, both between the PIFs for a given morphogenic-response feature, and between morphogenic-response features for a given PIF. For example, experiments comparing single, double, triple and quadruple pif-mutant combinations indicate that the individual PIFs appear to contribute additively or synergistically, in more or less equivalent fashion, to the promotion of hypocotyl-cell elongation growth in dark-grown seedlings [2], [3], [11]. By contrast, PIF1 appears to dominate the concomitant suppression of cotyledon separation that occurs in these same seedlings during dark-growth [2], [11]. In green seedlings, on the other hand, PIF4 and/or PIF5 appear to have a major role in promoting the stem and petiole elongation intrinsic to shade-avoidance in response to vegetative shade [21], [42], whereas PIF3 [43], together with PIFs 4 and 5 [44], contribute strongly to growth during the night period under diurnal light/dark cycles. Consistent with this general pattern, another related bHLH factor, PIF7, displays only moderate involvement in seedling deetiolation [45], but has a prominent role in shade avoidance [41]. Although a limited number of previous studies have examined the transcriptome regulated by PIF-quartet members in seedlings in darkness [3], [12], [16]–[19] and vegetative shade [11], [21] using the Affymetrix ATH1 array, these studies did not provide full genome coverage and did not permit dissection of potential quantitative differences in transcriptome profiles controlled by the individual PIFs.
The RNA-seq analysis performed here defines, with full genome coverage, the transcriptome collectively regulated by the PIF quartet in promoting skotomorphogenesis, and provides initial definition of the extent, and quantitative partitioning, of shared transcriptional control of the genes within of this network between PIF3 and the PIF1/4/5-trio. Superimposed on these data, our ChIP-seq analysis has identified a subset of these genes that are likely direct targets of PIF3 transcriptional regulation, exerted by physical binding of this factor to promoter-localized G- or PBE-box recognition-motifs (Class X, Y and Z genes, combined; Figure 4). The predominant pattern of PIF-regulated expression of these PIF3-bound genes (108 (84%) of 129 total) is one of high levels in the presence of the wild-type PIF factors, and reduced levels in the genetically-imposed absence of these factors in dark-grown seedlings, indicative of transcriptional activation by PIF3 and/or one or more of the other three PIF-quartet members. This pattern is consistent with the existing reports that all four factors function intrinsically as transcriptional activators, at least in transfection or heterologous expression systems [4]–[9], and with the demonstration here and elsewhere [21] that these PIFs function to activate PIL1-promoter-driven expression in transgenic seedlings (Figure 5C). We have therefore focused here primarily on this predominant class of PIF-transcriptionally-activated genes.
Our data indicate that there is a continuum, from robust to marginal, in the extent of the contribution of PIF3 to the combined transcriptional regulatory activity of the PIF quartet toward the PIF-induced, Class Y and Z genes (Figure S4). Conversely, by definition, there is a complimentary continuum in the share of this combined activity provided by the collective actions of PIFs 1, 4 and 5. These data imply at least some degree of shared, but quantitatively differential, transcriptional-regulatory activity among the PIF-quartet members toward individual genes that are apparent direct targets of PIF3-induced expression. Our RT-qPCR analysis of the expression patterns of selected genes from this subset, in all pif triple-mutant combinations, confirms that all four quartet members display such intra-subfamily differential activity toward individual genes in this set. Moreover, this analysis shows, conversely, that the individual PIF proteins induce differential levels of transcription in each different gene (Figure 5A). The three-dimensional response surface generated by this comparison (Figure 5B) suggests that this pattern may be iterated across all PIF-regulated genes genome wide, and points to the potential for considerable signaling and regulatory complexity at the PIF-target-gene interface.
Because it has been shown that all four PIF-quartet members bind robustly to the G-box motif [4]–[9], it appears likely that many of the direct targets of PIF3 transcriptional regulation are also direct targets of these other PIFs [5], [6], and that the shared activation of genes by the individual quartet members observed here will involve some degree of shared occupancy of these binding sites by the different PIFs. This may also apply to the newly discovered PBE-box motif, as there is recent evidence that PIF4 also recognizes this motif [21]. However, there is also evidence of potential divergence in motif recognition, as PIF5 was shown in the same report not to bind to the PBE-box motif [21]. The prominent contribution of PIF1 to the transcriptional activation of many of the genes examined here (Figure 5), despite its apparent considerably lower expression level than PIF3 (Figure 6F), is particularly intriguing in this respect, as this may imply that PIF1 may dominate promotion of target gene expression in dark-grown seedlings.
Comparison of the genes identified here as direct targets of PIF3 transcriptional activation (Class Z and YZ1.5 genes), with those previously identified as being rapidly (within 1 h) repressed by initial exposure of dark-grown seedlings to red light [17], has defined an overlapping subset of 21 genes (22 including PIL1) (Table 1). The evidence is strong, therefore, that these 22 genes form a core set that are directly transcriptionally activated by PIF3 in darkness and repressed in light, at least in part, by direct, photoactivated-phy-induced PIF3 degradation. Moreover, because all of these genes are also transcriptionally activated, either collectively (Table 1 and Table S11), or individually (Figure 5A and 5B) by PIF1, 4 and/or 5 in darkness, it appears likely that these PIFs share similarly directly in the light-reversible trans-activation of this core gene-set via photoactivated-phyB-induced degradation of the PIF-trio members.
The predicted or established functional diversity of the PIF direct-target genes identified here (Figure S5) suggests that PIF3 and/or one or more other PIF-quartet members act pleiotropically to directly regulate the transcription of a diversity of genes involved in a spectrum of cellular processes that sustain the skotomorphogenic developmental pathway. Consistent with previous analyses [3], [11], [17], [21], the PIF-induced genes are strikingly enriched for transcription-factor-encoding loci (40% of the annotated genes in this set). These data support the proposition, therefore, that the PIFs regulate an extensive transcriptional network via direct activation of a battery of primary target-genes in a hierarchal transcriptional cascade [20]. Because the encoded target-proteins represent multiple major classes of transcription factors (including bHLH, homeobox, bZIP, ARF, AUX/IAA, AP2-EREBP, BBX and TCP), it appears likely that they act concomitantly to activate multiple, diverse downstream pathways in parallel. Interestingly, however, many apparent PIF direct-target genes are involved in other cellular processes (including cytokinin metabolism, auxin-responsiveness, protein phosphorylation and cell-wall metabolism), suggesting a more immediate mode of PIF regulation of these processes.
A central issue in understanding mechanisms of eukaryotic transcriptional regulation is how members of large transcription-factor families, with conserved DNA-binding domains (such as the 162-member Arabidopsis bHLH family [46]), discriminate between target genes [22], [30], [47]. However, the specific question of whether, and to what extent, closely-related sub-family members, with potential overlapping functional redundancy (like the PIF quartet), share regulation of target genes through shared binding to promoter-localized consensus motifs, does not appear to have been widely investigated [31]–[34]. Our data, together with those of others [21], [24], provide evidence suggesting that the PIF quartet members share directly in transcriptional activation of numerous target genes, potentially via redundant promoter occupancy, in a manner that varies quantitatively from gene to gene (Figure 7). This finding suggests that these PIFs function collectively as a signaling hub, selectively partitioning common upstream signals from light-activated phys at the transcriptional-network interface. Definition of the mechanistic basis and functional consequences of this apparent complexity will require further investigation.
The Colombia-0 ecotype of Arabidopsis thaliana was used for all experiments. The 35S:6×His-PIF3-5×MYC (P3M) transgenic line [48], pif3 [18], pif1pif4pif5 (pif145) [11], pif1pif3pif4 (pif134), pif1pif3pif5 (pif135), pif3pif4pif5 (pif345), and pif1pif3pif4pif5 (pifq) [2] were as described.
Stratified seeds were irradiated with WL at 21°C for 3 h to induce germination, followed by a FR pulse for 15 min to suppress pseudo dark effects [2], and grown in darkness at 21°C for 2 d before harvest.
The ChIP assay was performed using about 2 g of Arabidopsis 2-d-old dark-grown whole seedlings as described [49] under green safelight. Polyclonal anti-MYC antibodies (Abcam, ab9132) were used with BSA-blocked Protein G Agarose beads (Millipore) to immunoprecipitate the P3M-DNA complex. Wild-type Arabidopsis seedlings grown under the same conditions were used as the negative control following the same assay procedure.
The ChIP-seq library was constructed according to Illumina's instructions (www.illumina.com) with some modifications. Four ChIP samples from technical replicates of each biological replicate were pooled together and concentrated to increase the starting amount of DNA. The end repair of DNA fragments was performed using End-It DNA End-Repair Kit (Epicentre). The A-tailing was added to the end-repaired DNA fragments using Klenow Fragment (NEB), and then Illumina's PE adapters were ligated by T4 DNA Ligase (Promega) at 16°C overnight. The adapter-ligated DNA fragments in the 200–300 bp size-range were selected by the gel purification, and then were amplified using Phusion High-Fidelity DNA Polymerase (NEB) with the Illumina PE PCR primer set. The library was purified using an Agencourt AMPure XP system (Beckman Coulter Genomics), and then validated by Bioanalyzer 2000 (Agilent).
The parallel libraries from P3M and WT ChIP samples were assayed by single-end sequencing on an Illumina GAIIx platform. The 36-nt reads were aligned to the TAIR9 assembly of the Arabidopsis genome using Bowtie [50] with up to 2 mismatches allowed. Only reads mapped uniquely to the nuclear genome with the lowest number of mismatches were retained for binding-peak identification. To increase the uniformity of read-counts across biological replicates, two technical-replicate sequencing runs were performed on the 4 libraries from the 1st and 2nd ChIP experiments (two of the four biological replicates). The aligned reads from the two technical sequencing replicates of each library were combined and processed as single biological replicate data.
The statistical identification of PIF3-binding peaks was performed separately for each biological replicate using MACS [35] with the default 10−5 P-value cutoff. MACS analysis was customized to ensure a more uniform analysis across biological replicates, and to decrease the size of the window for detecting background enrichment (due to the small size of the Arabidopsis genome) by employing modified parameters (gsize = 1.1e8, bw = 100, nomodel, shiftsize = 50, slocal = 1000, and llocal = 2000).
Four independent biological replicates of ChIP-seq data were collected, and replicate-specific binding peaks, identified in at least one other replicate, were defined as reproducible, if the distance between the summits of each replicate were less than 100 bp. For each reproducible peak, a mean summit position was assigned as the average position of the individual replicate-specific summits, and the PIF3-binding sites were defined as the 201 bp windows centered at each reproducible mean-summit position.
De novo PIF3-binding motif discovery was performed on the 201-bp defined binding sites using MEME [36], and the enrichment significance of identified G-box and PBE-box motifs beyond the genome background was quantified by 100 random simulations, where in each simulation 1064 randomly selected genomic regions of the same size were searched for the occurrence of each motif. The tight association of PIF3 binding with a specific motif was defined as the distance between the peak summit and the closest motif less than 100 bp.
Definition of the closest neighboring genes to each binding peak was approached by scanning the regions within +/−5 kb centered at each peak summit, using CisGenome [51], and the potential target genes downstream of each summit with no intervening genes were selected manually.
Total RNA was extracted from 2-d-old dark-grown seedlings using QIAshredder column and RNeasy Plus Mini Kit (Qiagen) according to the manufacturer's instructions. The sequencing library construction was adapted from 3′-end RNA-seq protocol [52]. The mRNA was fractionated from 20 µg of total RNA using Dynabeads Oligo (dT)25 (Invitrogen), and fragmented using Fragmentation Reagents (Ambion) at 70°C for 2.5 min in 20 µl of reaction. The polyA-tailed 3′-end fragments were captured by another run of mRNA purification as described above, and then treated by Antarctic Phosphatase (NEB) and T4 Polynucleotide Kinase (NEB) at 37°C for 1 h and 2 h, respectively. The sample was purified using RNeasy MinElute Cleanup Kit (Qiagen) according to Illumina's protocol. The eluted mRNA fragments were ligated with 2.5 µM of Illumina's SRA 5′ adaptor by T4 RNA Ligase 1 (NEB) at 20°C for 4 h. The 3′ cDNA adapter derived from Illumina's v1.5 sRNA 3′ adapter was conjugated with the anchored oligo (dT)20 primer, and introduced through reverse transcription using the SuperScript III First-Strand Synthesis System (Invitrogen). The first-strand cDNA was purified using the Agencourt AMPure XP system, and then amplified by PCR reaction using Phusion High-Fidelity DNA Polymerase with Illumina's sRNA PCR primer set. The size of purified library DNA was validated by Bioanalyzer 2000.
Libraries from the 1st biological replicate were assayed by 36-cycle single-end sequencing on the Illumina GAIIx platform, while libraries from the 2nd and 3rd biological replicates were assayed by 50-cycle single-end sequencing on the HiSeq2000 platform. For consistency, only the 5′-end 36-nt trimmed reads from the 2nd and 3rd replicates, as well as the full-length 36-nt reads from the 1st replicate, were aligned to the TAIR9 representative transcriptome using Bowtie with zero mismatches allowed. Only reads mapping uniquely to the 3′-end 500-bp region of the coding strand were counted for gene expression. Differentially expressed genes were identified using the edgeR package [53], and SSTF genes were defined as those that differ by ≥2-fold with an adjusted P value ≤0.05 as described [17].
The recombinant protein GST-PIF3-Flag and the GST control were purified from E. coli as described previously [9]. DNA probes were generated by annealing a 5′ biotinylated oligonucleotide (IDT) to a complementary unmodified oligonucleotide (IDT). The complementary oligonucleotides were diluted in annealing buffer (10 mM Tris-HCl (pH 7.5), 50 mM NaCl, 1 mM EDTA) to a final concentration of 40 µM, heated to 95°C for 5 min, and cooled down slowly (0.1°C/second) to 12°C. The same procedure was followed to generate unmodified dsDNA fragments for competition assays. Probes are listed in Table S13.
The DPI-ELISA assays were performed as described [38]. Biotinylated probes were bound to Reacti-Bind Streptavidin High Binding Capacity Coated 96-Well Plates (Thermo Scientific) by applying 2 pmol/well of the probes in TBS-T buffer (20 mM Tris-HCl (pH 7.5), 180 mM NaCl, 0.1% (v/v) Tween 20) for 1 h at 37°C. The wells were blocked with 5% (w/v) non-fat dry milk in TBS-T buffer for 30 min, and then incubated with 100 ng of GST-PIF3-Flag or GST for 1 h. For competition assays, 2, 10 or 50 pmol/well of the unlabeled probes were added at the same time with the proteins. After incubation, the wells were washed 3 times with TBS-T/PBS-T buffer, and then were incubated with 1∶2000 diluted THE GST Antibody [HRP] (GenScript, A00866) in PBS-T buffer for 1 h. The wells were then washed twice with PBS-T and PBS buffers, respectively, after incubation. The protein binding was detected by adding the OPD solution (Thermo Scientific), and the reaction was stopped by 2.5 M sulfuric acid. The color extinction was measured at 490 nm, using 650 nm as a reference wavelength in the ELISA reader.
The EMSA assays were performed as described [9]. 100 ng of recombinant proteins and the biotinylated DNA probes were used in each assay. Gel electrophoresis using native 5% PAGE gel in ice cold 0.5× TBE buffer (280 V, 15 min) was followed by wet-transfer electro blotting to Biodyne B Nylon membrane (Pierce) in 0.5× TBE buffer (80 V, 1 h). The Lightshift Chemiluminescent DNA EMSA kit (Pierce) was used for detection of the biotinylated probes according to the manufacturer's instructions.
RT-qPCR was performed as described [17]. Each PCR reaction was repeated at least twice, and the mean value of technical replicates was recorded for each biological replicate. Data from biological triplicates were collected, and the mean value with standard error is represented in the bar graphs. Primers and gene accession numbers are listed in Table S13.
The 1.8 kb PIL1 promoter region (pPIL1) upstream of the ATG was amplified by PCR using the pPIL1-F1/R1 primer set, and then the XhoI/EcoRI fragment was cloned into the pBluescript II SK (pBSK) vector (Stratagene) to produce pBSK-pPIL1. The G-box mutations were introduced by two-step PCR amplification (using pPIL1-F7/R7, pPIL1-F8/R8, pPIL1-F9/R9, and pPIL1-F10/R10 primer sets), and the XhoI/MfeI fragment from the pPIL1-F1/R5 primer set was cloned to replace the unmutated fragment of pBSK-pPIL1 to produce pBSK-mpPIL1. The HindIII/BamHI fragment containing the omega-LUC+-rbcS terminator from the pENTR/D-TOPO\arGIp::LUC+ construct was cloned into pBSK-pPIL1 and pBSK-mpPIL1, respectively, to produce pBSK-pPIL1:LUC and pBSK-mpPIL1:LUC. The CDS of Renilla Luciferase (RLUC) was amplified by PCR using the Rluc-F1/R1 primer set, and then the NcoI/PmlI fragment was cloned into the pCAMBIA1302 binary vector to produce pC1302-35S:RLUC. The PstI/SacI fragments from pBSK-pPIL1:LUC and pBSK-mpPIL1:LUC were then sub-cloned into pC1302-35S:RLUC to produce pC1302-pPIL1:LUC-35S:RLUC (pPIL1:LUC) and pC1302-mpPIL1:LUC-35S:RLUC (mpPIL1:LUC), respectively.
The 2.5–3.0 kb promoter regions upstream of the ATG of PIF3, PIF4 and PIF5 were amplified from Arabidopsis (Col-0 ecotype) genomic DNA by PCR using the TOPO-PIF3p-LP1/RP1, TOPO-PIF4p-LP1/RP1 and TOPO-PIF5p-LP1/RP1 primer sets, respectively. The PCR products were cloned into the pENTR/D-TOPO vector (Invitrogen) to produce the pPIF3, pPIF4 and pPIF5 entry clones. For the PIF1 promoter, the first 2 kb fragment upstream of the ATG was amplified by PCR using the TOPO-PIF1p-LP3/RP1 primer set, and then was cloned into the pENTR/D-TOPO vector to produce the intermediate entry clone. The second fragment of 2–4 kb upstream of ATG was amplified using the NotI-PIF1p-LP/XcmI-PIF1p-RP primer set, and then the NotI/XcmI fragment of the PCR product was subcloned into the intermediate entry clone to produce the pPIF1 entry clone. All four entry clones were subcloned into the gateway compatible pGWB3 binary vector [54] using Gateway LR Clonase II Enzyme Mix (Invitrogen) to produce pPIF:GUS constructs.
The constructs were transformed into Arabidopsis plants as described [55], and the individual transgenic lines were selected on MS medium containing 25 mg/L of Hygromycin B (Invitrogen).
The 2-d dark-grown seedlings of independent transgenic lines were ground in liquid nitrogen, and total protein was extracted in LUC extraction buffer (1× PBS, 4 mM EDTA, 2 mM DTT, 5% glycerol, 1 mg/ml BSA, 2 mM PMSF and 1× complete protease inhibitor cocktail (Roche) at 3× w/v) as described [7]. 20 µl of the supernatant were used to measure the LUC and RLUC activity using a Dual-Luciferase Reporter Assay System (Promega) according to the manufacturer's instruction. The relative expression of LUC was represented by its enzyme activity compared to the RLUC internal control.
Histochemical GUS staining assays were performed on 2-d-old dark-grown seedlings as described [56] using a modified substrate buffer (1× PBS (pH 7.0), 1 mM K3Fe(III)(CN)6, 0.5 mM K4Fe(II)(CN)6, 1 mM EDTA, 1% Triton X-100, 1 mg/ml X-gluc). Data of biological triplicates were collected from two independent transgenic lines, and representative images are shown for each transgene.
ChIP-seq and RNA-seq data reported in this study have been deposited in the Gene Expression Omnibus database under the accession number GSE39217.
|
10.1371/journal.ppat.1000039 | The MHV68 M2 Protein Drives IL-10 Dependent B Cell Proliferation and Differentiation | Murine gammaherpesvirus 68 (MHV68) establishes long-term latency in memory B cells similar to the human gammaherpesvirus Epstein Barr Virus (EBV). EBV encodes an interleukin-10 (IL-10) homolog and modulates cellular IL-10 expression; however, the role of IL-10 in the establishment and/or maintenance of chronic EBV infection remains unclear. Notably, MHV68 does not encode an IL-10 homolog, but virus infection has been shown to result in elevated serum IL-10 levels in wild-type mice, and IL-10 deficiency results in decreased establishment of virus latency. Here we show that a unique MHV68 latency-associated gene product, the M2 protein, is required for the elevated serum IL-10 levels observed at 2 weeks post-infection. Furthermore, M2 protein expression in primary murine B cells drives high level IL-10 expression along with increased secretion of IL-2, IL-6, and MIP-1α. M2 expression was also shown to significantly augment LPS driven survival and proliferation of primary murine B cells. The latter was dependent on IL-10 expression as demonstrated by the failure of IL10−/− B cells to proliferate in response to M2 protein expression and rescue of M2-associated proliferation by addition of recombinant murine IL-10. M2 protein expression in primary B cells also led to upregulated surface expression of the high affinity IL-2 receptor (CD25) and the activation marker GL7, along with down-regulated surface expression of B220, MHC II, and sIgD. The cells retained CD19 and sIgG expression, suggesting differentiation to a pre-plasma memory B cell phenotype. These observations are consistent with previous analyses of M2-null MHV68 mutants that have suggested a role for the M2 protein in expansion and differentiation of MHV68 latently infected B cells—perhaps facilitating the establishment of virus latency in memory B cells. Thus, while the M2 protein is unique to MHV68, analysis of M2 function has revealed an important role for IL-10 in MHV68 pathogenesis—identifying a strategy that appears to be conserved between at least EBV and MHV68.
| Gammaherpesviruses are able to maintain life-long, quiescent infections (latency) in lymphocytes characterized by intermittent production of infectious progeny virus (reactivation). The murine gammaherpesvirus 68 (MHV68) has extensive genetic and phenotypic similarities to the human gammaherpesviruses Epstein-Barr virus (EBV) and Kaposi's sarcoma associated herpesvirus (KSHV). Similar to EBV pathogenesis, MHV68 establishes long-term latency in memory B cells. A unique MHV68 protein designated M2 is known to play an important role in both establishment of latency and reactivation from latency. Efficient transition of MHV68 latency to the memory B cell population is hindered in the absence of M2, leading to the hypothesis that M2 may be involved in MHV68-driven B cell differentiation. In this study we show that M2 expression enhanced primary murine B cell survival, proliferation, and differentiation in culture. M2 expressing B cells secreted of high levels of IL-10 that is necessary for the observed expansion of the M2-expressing B cell population in vitro. Mice infected with a M2-null MHV68 mutant had a significant decrease in serum IL-10, and this correlated with an increased frequency of MHV68-specific CD8+ T cells in these animals. Thus, M2 manipulation of IL-10 signaling appears to both drive expansion of the major latency reservoir (B cells), as well as dampen the immune response to the virus facilitating both viral latency and reactivation.
| Herpesviruses establish life-long, latent infections characterized by episodic virus reactivation and subsequent virus shedding. Chronic infections with the lymphotropic gammaherpesviruses are associated with a variety of lymphomas and carcinomas which in humans includes Burkitt's lymphoma, nasopharyngeal carcinoma, Hodgkin's disease and Kaposi's sarcoma. The narrow host range of the gammaherpesviruses that infect humans, Epstein-Barr Virus (EBV) and Kaposi's sarcoma-associated herpesvirus (KSHV), has severely hindered detailed pathogenesis studies. Murine gammaherpesvirus 68 (MHV68; also known as γHV68 and murine herpesvirus 4) shares extensive genetic homology and biological similarity with both EBV and KSHV and is a natural pathogen of wild murid rodents. As such, MHV68 infection of inbred strains of mice has gained favor as a small animal model in which to evaluate viral and host determinants of gammaherpesvirus pathogenesis in vivo.
Upon intranasal infection, MHV68 infection results in acute viremia in the lung that is later resolved into a latent infection of B cells, dendritic cells, and macrophages [1]. B cells are necessary for trafficking of virally infected cells to the spleen, leading to the establishment of splenic latency [2],[3]. The CD8+ T cell response is critical for control of lytic infection in the lung as well as establishment of latent viral load in the spleen [4]. MHV68 infection results in a CD4+ T cell-dependent expansion of splenic B cells and both virus-specific and non-specific hypergammaglobulinemia [5],[6]. Similar to EBV pathogenesis, memory B cells are the primary long-term reservoir of latent MHV68 in mice [7],[8],[9].
All herpesviruses manipulate the host's immune system to establish and maintain a long-term, latent infection, and many of these immunomodulatory mechanisms are conserved among the members of the gammaherpesvirus family. Both KSHV and MHV68 encode proteins, K3 and mK3, respectively, that downregulate MHC I [10]. MHV68 also encodes a viral bcl-2 homolog, a viral cyclin, and a chemokine-binding protein, M3 [11],[12],[13]. The EBV proteins LMP1 and LMP2a mimic CD40 and tonic BCR signals, respectively, to manipulate B cell development and are believed to enable the virus to gain access to the memory B cell compartment independent of antigenic stimulation of the host B cell [14],[15]. KSHV encodes both a viral IL-6 and a viral MIP-1α ortholog, while EBV encodes a viral IL-10 homolog, BCRF1 (or vIL-10) [16],[17].
Interleukin-10 (IL-10) was first noted as a cytokine synthesis inhibitory factor (CSIF) that is secreted by TH2 cells and suppresses the activity of TH1 cells [18]. IL-10 enhances murine B cell viability and can activate human B cell proliferation and class switching in culture [17],[19]. In addition, IL-10 suppresses TH1 responses through modulation of macrophage function by downregulation of MHC II and costimulatory molecules as well as inhibition of cytokine production and macrophage effector functions [20],[21]. Dendritic cells exposed to IL-10 do not down-regulate costimulatory molecules but do secrete lower levels of IL-12, impairing their ability to induce a TH1 response [22].
The M2 protein, unique to MHV68, has been shown to play a critical role in both the establishment of latency as well as reactivation from latency [23],[24],[25]. A M2-null strain of MHV68 (MHV68/M2.Stop) replicates with wild-type efficiency in mice following intranasal inoculation but exhibits a dose-dependent defect in the establishment of latency at day 16 post-infection [23],[24]. Under conditions in which the MHV68/M2.Stop mutant can efficiently establish a latent infection (high dose intranasal inoculation or low dose intraperitoneal inoculation), the M2-null virus exhibits a profound reactivation defect, revealing dual roles for the M2 protein in the viral life-cycle [23],[24]. Additionally, efficient transition of latently-infected B cells from the germinal center reaction to the memory B cell reservoir appears to be stalled in the absence of M2, suggesting M2 may manipulate B cell signaling or differentiation to facilitate establishment of long-term latency in the memory B cell pool [24],[26]. Numerous candidate SH3 binding motifs throughout M2 suggest the protein may function as a molecular scaffold that may modulate specific cellular signal transduction pathways. Consistent with this hypothesis, M2 has been shown to interact with a number of cellular proteins in vitro. M2 co-immunoprecipitates with Vav1 in S11 B cells, a MHV68 latently infected cell line, and M2 and Vav1 overexpression in A20 B cells leads to Vav1 phosphorylation, trimerization with Fyn, and downstream activation of Rac1 [27]. In fibroblast cultures, M2 interacts with DDB1/COP9/cullin repair complex and ATM to suppress DNA-damage induced apoptosis [28]. In addition, M2 can suppress STAT1/2 expression, leading to inhibition of the interferon response [29]. However, to date the impact of M2 expression in primary murine B cells has not been reported.
Here we show that one function of the M2 protein is to induce expression of IL-10 in primary B cells, demonstrating a common immunomodulatory strategy utilized by those gammaherpesviruses encoding a viral IL-10 homolog and MHV68.
The MHV68 M2 protein has been shown to be critical for both establishment and reactivation from B cell latency. M2 has no known homologous proteins, viral or cellular, and contains numerous SH3 binding motifs through which it can potentially manipulate B cell biology. Proliferating B cells harbor the majority of latent MHV68 genomes, and splenic B cell activation is associated with MHV68 infection at the onset of latency [30],[31]. We asked whether expression of M2 in primary murine B cells in vitro altered proliferation or activation. B cells were purified by negative selection from mouse splenocytes, stimulated overnight with LPS, and transduced with either an M2 protein expressing recombinant murine stem cell virus (MSCV) retrovirus, MSCV-M2-IRES-Thy1.1, or a control retrovirus, MSCV-M2.Stop-IRES-Thy1.1, which harbors a translation termination codon near the 5′ end of the M2 open reading frame at amino acid 13 (Figure 1A). LPS stimulation is necessary for efficient retroviral transduction in this system because MSCV infection requires the cells to be in cycle [32]. The presence of an IRES-Thy1.1 cassette readily allowed retroviral transduction efficiency to be monitored by flow cytometry for surface expression of Thy1.1. Notably, expression of the M2 protein from the retroviral construct could be detected as demonstrated by immunoprecipitation and immunoblotting of whole cell lysates harvested from primary B cells transduced with MSCV-M2-IRES-Thy1.1 at four days post-transduction (Figure 1B). It should be noted that detection of M2 expression in the transduced primary murine B cells required immunoprecipitation with a chicken anti-M2 antisera raised against two M2 peptides, followed by immunoblotting with a rabbit polyclonal antiserum raised against a bacterially expression recombinant M2 protein. In contrast, M2 expression in the MHV68 latently infected B lymphoma cell line S11 can be detected by immunoblotting S11 lysates with the rabbit polyclonal anti-M2 antiserum (data not shown). Thus, it does not appear that M2 is “over-expressed” in transduced primary murine B cells.
Two days post-transduction, there were similar frequencies of transduced, Thy1.1+ B cells in the control and M2-transduced B cell cultures. However, by 5–6 days post-transduction, nearly 100% of the B cell culture transduced with MSCV-M2-IRES-Thy1.1 was Thy1.1+ as compared to ca. 20% of the culture transduced with the control vector, MSCV-M2.Stop-IRES-Thy1.1 (Figure 1C). Notably, the dominance of M2-expressing, Thy1.1+ B cells in the M2-transduced cultures was observed repeatedly. The increase in the percentage of Thy1.1+ cells in the M2-transduced culture was gradual, and it did not correspond to an increase in overall cell number in the cultures or a decrease in cell death (Figure 1D). The latter result suggests that in a mixed culture (M2 expressing and non-expressing cells), the non-transduced primary B cells are actively selected against. This could either be due to the secretion of a “toxic” factor by the M2 expressing cells or competition for a limiting factor necessary for cell survival.
Upon observing the M2-transduced cells dominating the culture, we asked whether M2 was influencing B cell survival, proliferation, or both. To directly assess B cell survival in M2-transduced and control retrovirus cultures (M2.Stop), cells were stained with anti-Thy1.1, Annexin V, and 7-AAD and analyzed by flow cytometry. In contrast to the results obtained by trypan blue exclusion which measured the live/dead ratio in the entire culture (Figure 1D), flow cytometry of the transduced and untransduced populations within the culture revealed a survival advantage of the M2-transduced B cells. At day 2 post-transduction, 20% more of the M2-transduced B cells were alive (AnnexinV− 7-AAD−) than the untransduced cells in the same culture (Figure 2B). At day 3 post-transduction, there was a four-fold higher frequency of live cells in the M2-transduced population as compared to the untransduced cells in culture (Figure 2, panels A & B). M2-transduced cells continued to survive better than the untransduced cells in the population, despite an equal frequency of cells entering apoptosis (data not shown).
At day 2 post-transduction, the M2-transduced cells have equal frequencies of live cells as the control M2.Stop retrovirus transduced cells (Figure 2B). Analysis at day 3 post-transduction revealed a 20% increase in the frequency of live cells in the M2-transduced population as compared to the cells transduced with the control retrovirus (Figure 2B). This trend continued until the end of the time-course, with a higher frequency of live cells found in the M2-transduced population versus the M2.Stop retrovirus control (Figure 2B). The increased frequency of live cells in the M2-transduced population versus both the untransduced cells within the culture as well as the control retrovirus transduced population reveals a pro-survival effect of M2 protein expression in B cells.
We next addressed whether M2 protein expression altered proliferation in the B cell cultures thereby contributing to the expansion of transduced cells. To directly assess B cell proliferation in M2-transduced and control retrovirus cultures, cells were pulsed with bromodeoxyuradine (BrdU) for 24 hours at different time points post-transduction. Cells were surface stained for Thy1.1 and proliferation was measured by intracellular staining for incorporation of BrdU. The time course analyses revealed that B cells transduced with either the M2 or control retrovirus exhibited equivalent frequencies (84–90%) of proliferating cells 2 days post-transduction (Figure 2D). However, 80–90% of M2-transduced B cells continued to proliferate 3 and 4 days post-transduction as compared to 40–50% of the cells transduced with the control retrovirus (Figures 2, panel C & D). By 5–6 days post-transduction there was a significant drop in the proliferation of M2-transduced B cells (Figure 2D). These results indicate that M2 protein expression is able to transiently augment murine B cell proliferation. Thus, these analyses indicated that the M2 protein contributes to both enhanced B cell survival as well as promoting continued B cell proliferation post-LPS stimulation – which together leads to dominance of M2-transduced B cells in the primary murine B cell cultures over the time-course analyzed.
The transition from the germinal center B cell population to the long-lived memory B cell compartment is critical for establishment of MHV68 latency [8],[9]. Latent genomes in mice infected with M2-deficent MHV68 accumulate in the germinal center compartment late in infection, leading to the hypothesis that M2 is capable of manipulating B cell differentiation [26]. To determine whether M2-transduction leads to differentiation of B cells, surface expression of B cell differentiation markers was analyzed by flow cytometry. At four days post-transduction, B cells expressing M2 were CD19+, CD25high, GL7high, B220low, I-Ab low, surface IgD− (sIgD), sIgG+, and CD138low when compared to untransduced cells within the culture (Figure 3A). Strikingly, M2-transduced cells expressed higher levels of CD25 as compared to cells transduced with the control retrovirus, although the MFI of CD25 was similar between the two populations (Figure 3A). Both transduced populations (M2 and M2.Stop) became surface IgG positive, likely due to LPS stimulation coupled with retrovirus infection selecting for the LPS-driven proliferating B cell population. However, the M2 expressing B cells expressed higher levels of surface sIgG than the control M2.Stop retrovirus transduced cells. Similarly, the M2 and M2.Stop transduced populations both upregulated CD138, although the presence of M2 did not lead to the high levels of CD138 indicative of plasma cell differentiation. Notably, the other changes observed in B cell differentiation were unique to the M2-transduced B cell population versus the cells transduced with the M2.Stop control retrovirus. In addition, the M2-transduced B cells secreted significantly higher levels of IgG on days 4–6 post-transduction than the cells transduced with the control retrovirus (Figure 3B). Secreted IgM levels remained similar throughout the time-course for M2 and control retrovirus B cell cultures (Figure 3C). Importantly, M2-transduced cells express surface IgG and remain CD138low, indicating that they have not fully differentiated into plasma cells. Together, these data provide strong evidence that M2 expression leads to B cell activation and differentiation similar to a functional activated, pre-plasma memory B cell phenotype, namely CD19+, sIgG+, sIgD−, B220low, CD138low [33],[34]. However, we cannot formally rule out that M2 expression leads to differential survival and expansion of a population of pre-plasma memory B cells present in the transduced culture – although this seems unlikely based on the very low frequency of this population in the purified naïve splenic B cells used for these studies.
To further investigate the proliferative effects of M2 protein expression in primary murine B cells, the supernatants of the transduced B cells were screened for a variety of cytokines using a mouse cytokine antibody array (see Materials and Methods). Supernatants of B cell cultures transduced with M2 and control retrovirus were compared at four days post-transduction (Figure 4A). Cytokine arrays performed in duplicate time-course experiments revealed substantial increases in IL-10, IL-2, IL-6, and MIP-1α in the culture supernatants of B cells expressing M2 compared to the control retrovirus transduced B cell cultures (Figure 4A). Cytokine levels throughout the time-course analyses were subsequently quantitated by ELISA. IL-2 levels in the M2-transduced cultures peaked at 50 pg/mL of supernatant at day 4 and waned by day 6 post-transduction, while only 1–2 pg/mL of IL-2 were detected in the control retroviral supernatants (Figure 4B). From 3 days post-transduction until the end of the time-course, the supernatants from M2-expressing B cells contained levels of IL-6 twice as high as those of B cells transduced with the control retrovirus (780 pg/mL vs. 400 pg/mL) at day 6 post-transduction (Figure 4C). There was also a 10-fold increase in the level of MIP-1α with M2-transduced cultures containing an average of 1845 pg/mL of MIP-1α versus 139 pg/mL in the control retrovirus supernatant at the end of the time course (Figure 4D). Notably, we observed a 20-fold increase in IL-10 levels in the B cell cultures transduced with M2 with 17.5 ng/mL of IL-10 in the M2-transduced cultures as compared to 0.9 ng/mL in the control retroviral supernatants at day 6 post-transduction (Figure 4E). Notably, the number of cells in the M2 protein expressing and control B cell cultures were not significantly different, and thus the observed differences in cytokine levels cannot be explain by an increase in cell number. These data demonstrate that M2 expression in primary murine B cells leads to enhanced secretion of several cytokines, most notably IL-10. Finally, to further assess the ability of M2 expression to up-regulate IL-10 secretion from B cells, we transfected the murine A20 B cell line with either a control expression vector (pIRES-EGFP) or an M2 expression vector (pM2-IEGFP) and assessed IL-10 secretion by ELISA (Figure 4F). Untreated A20 cells secrete significant levels of IL-10, which were only modestly enhanced by LPS treatment (Figure 4F). In addition, transfection of the control expression vector had no impact of the levels of IL-10 secreted by A20 cells (Figure 4F). However, transfection with the M2 expression vector lead to a substantial increase in the levels of IL-10 secretion (Figure 4F). The latter result provides further evidence that M2 is able to increase IL-10 secretion by B cells – independent of LPS stimulation.
IL-10 has been demonstrated to be involved in the establishment of a latent MHV68 infection, and we asked whether IL-10 played a role in M2-driven B cell proliferation [35],[36]. To address the role of IL-10 in M2-driven proliferation, B cells were isolated from wild-type and IL-10−/− mice, transduced with M2 or the control retrovirus, and surface Thy1.1+ expression was monitored over a six day time course. Although the percentage of M2-transduced C57Bl/6 B cells increased from 40% to 85% of the culture, as previously observed (see Figure 1C), there was only a modest expansion of the Thy 1.1+ population from 37% to 49% in the IL-10−/− cultures transduced with M2 expressing MSCV retrovirus (Figure 5A). ELISAs of the supernatants from the transduced cultures confirmed that the IL-10−/− B cells do not secrete detectable levels of IL-10 (Figure 5B). We noted an approximately 10% increase in the percentage of IL-10−/− M2 protein expressing B cells over the time course experiments, and we hypothesize that this small increase might be due to the ability of the M2 protein to manipulate proliferation and/or survival pathways independent of IL-10. Notably, IL-10−/− mice have been shown to have 20-fold higher levels of serum IL-6 than IL-10-sufficient mice [37], and indeed we observed a two-fold increase in IL-6 in the IL-10−/− B cell supernatants of the untransduced population at day 2 (Figure 5C). Expression of the M2 protein led to a four-fold increase in the levels of IL-6 in the culture supernatants of IL-10−/− B cells (Figure 5C). In addition, MIP-1α levels were four-fold higher in the IL-10−/− B cell cultures at day 2 post-transduction, and this increase was observed throughout the time-course (data not shown). However, the increased levels of IL-6 and MIP-1α observed in the M2-transduced IL-10−/− cultures could not compliment the loss of IL-10 in the cultures, leading us to hypothesize that IL-10 secretion is required for the expansion of the M2 protein-expressing B cells.
To more directly assess the role of IL-10 in M2 protein-mediated B cell proliferation and survival, we tested the ability of the cytokine enriched supernatants from transduced wild-type and IL-10−/− B cells to compliment loss of M2 and IL-10 expression in culture. After analysis of transduction efficiency at two days post-transduction, one third of the supernatant from the WT MSCV-M2.Stop and IL-10−/− MSCV-M2 transduced cultures was replaced with an equal volume of supernatant from C57BL6 MSCV-M2 cultures from the respective days post-transduction. B cells were analyzed for Thy1.1 expression for the remainder of the time-course, and IL-10 levels were measured by ELISA (data not shown). Interestingly, addition of culture supernatants from C57BL6 M2 protein expressing B cells to C57BL6 B cells transduced with MSCV-M2.Stop failed to induce significant proliferation of the transduced B cells (Figure 5D). In contrast, addition of IL-10 containing culture supernatants to IL-10−/− B cells expressing the M2 protein led to a steady proliferation nearly equivalent to that of C57BL6 B cell cultures expressing the M2 protein (Figure 5D).
Finally, to formally demonstrate that IL-10 is required for the observed phenotype, we transduced IL10−/− B cells with either the M2 or M2.Stop control recombinant MSCV viruses and assayed the frequency of Thy 1.1. cells in the culture over time in the presence and absence of recombinant IL-10 (Figure 5E). As expected, the addition of recombinant IL-10 had no discernable effect on M2 expressing IL-10-sufficient B cells recovered from C57Bl/6 mice. However, addition of IL-10 to the M2 transduced IL-10−/− B cells (but not the M2.Stop transduced IL-10−/− B cells) rescued the dominance phenotype (Figure 5E). These results demonstrate that both intracellular M2 expression and IL-10 secretion are necessary for the observed proliferative expansion of the transduced B cell population, and that neither one alone is sufficient to induce this expansion. These data suggest that M2 manipulates intracellular signaling pathways which enhance the response to IL-10 signaling as well as induce IL-10 secretion.
Previous studies have shown that in the absence of a functional M2 gene, establishment of MHV68 latency following intranasal inoculation is severely reduced [24]. Similarly, inoculation of IL-10−/− mice with wild-type MHV68 leads to a decrease in the establishment of latency [35],[36]. To determine whether M2 expression leads to IL-10 secretion in vivo, C57Bl/6 mice were infected (either 1,000 pfu via intranasal inoculation or 100 pfu via intraperitoneal inoculation) with either a recombinant MHV68 harboring the same translation termination codon near the 5′ end of the M2 open reading frame as used in control retrovirus construction (MHV68/M2.Stop) or with a genetically repaired marker rescue isolate of the same locus (MHV68/M2.MR). Both intranasal and intraperitoneal inoculation of the M2-null mutant were assessed, since we have previously reported that route of inoculation impacts the latency phenotype observed [24]. Serum IL-10 was measured by in vivo cytokine capture and ELISA (Figure 6, panels A & B). Notably, mice infected with MHV68/M2.Stop had serum IL-10 levels that were only slightly elevated over the levels present in naïve mice and were 2- to 3-fold lower than the levels observed in mice infected with the marker rescue virus (MHV68/M2.MR). Notably, this phenotype was independent of the route of inoculation (Figure 6, panels A & B). As we have previously reported [24], we observed defects in both establishment of latency (which was accentuated following intranasal inoculation), as well as reactivation from latency with the M2-null mutant MHV68 (Figure 6, panels C & D). Intraperitoneal infection with MHV68/M2.Stop increased the establishment of latency eight-fold over intranasal inoculation, yet serum IL-10 levels were very similar to those observed following intranasal inoculation (Figure 6). Importantly, the serum levels of IL-10 we observed in MHV68/M2.MR infected mice were similar to those previously observed [35]. These results provide strong evidence that M2 induction of IL-10 secretion, either from latently infected B cells or some other latency reservoir (e.g., infected macrophages or dendritic cells), contributes significantly to the serum levels of IL-10 observed during MHV68 infection following either intranasal or intraperitoneal virus inoculation.
We next examined whether loss of M2 expression and the concomitant reduction in IL-10 expression might alter the CD8 T cell response to MHV68 since IL-10 is known to suppress T cell responses [18]. Thus, we examined the MHV68-specific CD8+ T cell response following infection of mice with either MHV68/M2.Stop or MHV68/M2.MR. Mice were infected intraperitoneally with 100 pfu of MHV68/M2.Stop or MHV68/M2.MR and splenocytes were harvested at day 16 post-infection, a time at which lytic virus has been cleared and latency established. As previously reported, there was a ten-fold decrease in establishment of latency with a 20-fold decrease in reactivation (Figure 6D). Splenocytes from individual mice were stained for activated, tetramer positive CD8+ T cells using tetramers specific to two MHV68 antigens encoded by ORF6 and ORF61 (Figures 7, panels A–C). Both tetramers used in this analysis were specific for viral antigens expressed during the virus lytic replication cycle. In two independent experiments, tetramer staining for two different lytic antigens revealed a statistically significant increase in the frequency of tetramer-specific, activated CD8+ T cells in mice infected with MHV68/M2.Stop compared to MHV68/M2.MR (Figure 7A–C). In contrast, there was no global change in overall CD8 activation as determined by the percentage of CD8+ CD11ahigh T cells in the spleens of infected mice (Figure 7D). CD4+ T cell activation, as well as the percentage of CD44high CD62Llow CD4+ and CD8+ T cells, was the same in the two groups of infected mice (data not shown). These data indicate that the loss of M2 during MHV68 infection specifically enhanced the MHV68-specific CD8+ T cell response, despite a significant decrease in viral latency and reactivation (see Figure 6D).
Overall, the immune response in the absence of M2 protein expression during infection is unique in that the MHV68-specific T cell response is increased correlating with a decrease in serum IL-10 levels. These data point to a potential role of M2 protein-mediated IL-10 secretion in the quiescence of the virus-specific T cell response in vivo which may facilitate both the efficient establishment of latency as well as reactivation from latency.
The latency-associated M2 protein is critical for establishing splenic latency following low dose intranasal inoculation and for virus reactivation from latency following low dose intraperitoneal inoculation [23],[24]. In the absence of M2, infected B cells are unable to efficiently transition from the germinal center to the follicles [26]. Early in latency, there is an accumulation of latently infected naïve B cells in the absence of the M2 protein, indicating a role for the M2 protein in manipulating B cell development during infection [24]. Epstein-Barr virus is hypothesized to drive naïve B cells to enter the germinal center reaction in order to establish latency in the memory B cell pool [7],[38],[39]. In long-term EBV carriers, lytic EBV gene transcripts are preferentially found in the plasma cell population, leading to a model whereby reactivation from latency is associated with differentiation from memory to plasma cell [40]. B cell proliferation is necessary for the establishment of MHV68 latency, and, similar to EBV, memory B cells are the primary long-term latency reservoir [8],[9],[30]. Reactivation is hypothesized to be needed for efficient seeding of the spleen during the establishment phase of MHV68 infection, and, as such, the M2-associated defects in establishment of latency and reactivation from latency may, in fact, be functionally linked. In this study we explored the impact of M2 protein expression in primary murine B cells - a system capable of differentiation.
M2 expression in primary B cells led to proliferation of transduced B cells, driving a rapid expansion of transduced cells within the culture, regardless of initial transduction efficiency. Although both the transduced (Thy1.1+) and untransduced (Thy1.1−) B cell populations could be shown to be proliferating (by BrdU incorporation), the enhanced proliferation and survival of the M2-transduced B cells rapidly led to this population dominating the mixed culture. In primary murine B cells, M2-driven proliferation was dependent on the B cell's ability to secrete IL-10 and respond to IL-10 signaling. Notably, transfer of culture supernatants from M2 expressing C57Bl/6 B cells, or addition of recombinant murine IL-10, did not result in dominance of the M2.Stop retrovirus transduced (i.e., Thy1.1+) population in the absence of M2 protein expression, leading us to hypothesize that some other function(s) of the M2 protein augments IL-10 signaling. Culturing stimulated human memory B cells with IL-10 or IL-2 and IL-6 leads to plasma cell differentiation [41],[42]. Also, an increase in MIP-1α transcription is associated with differentiation to a plasma cell phenotype [43]. Human germinal center B cells can be induced to differentiate into plasma cells rather than memory B cells in the presence of IL-10 [44]. In contrast to human B cells, IL-10 enhances murine B cell viability but does not drive proliferation [19]. Our data suggest that the MHV68 M2 protein uniquely increases the murine B cell proliferative response to IL-10, mimicking the role of IL-10 signaling in human B cells.
M2-transduced B cells were B220low, I-Ab low, sIgD−, yet retained surface expression of CD19 and IgG, and remained CD138low, indicating that they did not fully differentiated into plasma cells. Instead, the surface phenotype of the B cells expressing M2 most closely resembled that of a pre-plasma memory B cell, an intermediate stage in development between the memory and plasma cell phenotypes [33],[45]. Together, this data supports a model wherein infection of naïve B cells in the lung with MHV68 leads to M2 expression, B cell proliferation and activation, and differentiation to a pre-plasma memory B cell phenotype. Depending on other cytokines and signals in the area, M2-expressing B cells may further differentiate into memory B cells, establishing long-term latency, or plasma B cells, potentiating virus reactivation. Thus, in this model of MHV68 pathogenesis, M2 protein manipulation of B cell differentiation to an intermediate pre-plasma memory B cell phenotype could facilitate both virus reactivation as well as establishment of viral latency.
IL-10 has potent immunoregulatory activity, suppressing proinflammatory cytokine secretion and activation of antigen-presenting cells - functions which result in suppressed NK cell and T cell activity [46]. IL-10 plays an important role in MHV68 pathogenesis, but prior to our analyses of M2 protein function no specific viral antigen had previously been shown to stimulate cellular IL-10 production. It has been shown that ex vivo stimulation of MHV68 latently infected splenocytes with MHV68-infected antigen presenting cells resulted in IL-10 secretion peaking at the onset of splenic latency, and B cells were shown to be responsible for a significant portion of the IL-10 secreted [47]. Dendritic cells isolated from MHV68 infected mice express IL-10 transcripts, and dendritic cells infected ex vivo secrete IL-10 only when concurrently stimulated with LPS [36]. Interestingly, these investigators showed that M2 is transcribed by infected dendritic cells, although they did not demonstrate that IL-10 secretion was mediated by M2 [36]. In the absence of IL-10, establishment of MHV68 latency is decreased concurrent with an increase in serum IL-12 p70 and splenomegaly, demonstrating a role for IL-10 in both establishment of latency as well as immunosuppression [1],[35],[47].
We observed a significant decrease in serum IL-10 in mice infected with an M2-null MHV68 mutant. Notably, the decreased serum IL-10 levels correlated with an increase in the percentage of MHV68-specific CD8+ T cells. Furthermore, it is important to note that this increased CD8+ T cell response was in the setting of an infection where virus reactivation was severely attenuated. Therefore, increased persistent virus replication cannot explain the increase in the tetramer-specific response. Thus, we hypothesize that during M2-mediated reactivation there is concurrent IL-10 secretion, locally dampening the ability of the MHV68-specific CD8+ T cells to clear the infected cells, leading to enhanced establishment and reactivation from latency.
Manipulation of the IL-10 signaling pathway appears to be a conserved mechanism used by a number of herpesviruses. EBV encodes a viral IL-10 homolog, BCRF1 (or vIL-10), that has been shown to increase human B cell proliferation following surface immunoglobulin crosslinkinking and induce B cells to secrete increased levels of IgM, IgG, and IgA in a similar manner to cellular IL-10 [17]. In the absence of vIL-10, EBV can still efficiently establish latent, long-term lymphoblastoid lines (LCLs) [48]. Exogenous vIL-10 added during the transformation of B cells by EBV enhanced both the rate and frequency of growth transformation, and antisense oligonucleotides to vIL-10 could negate this enhancement [49],[50]. Human IL-10 could complement the loss of vIL-10 during EBV infection of B cells, demonstrating that it is IL-10 mediated signaling that augments B cell transformation following EBV infection [50]. LMP1, a functional CD40 ortholog encoded by EBV, is both IL-10 responsive and induces secretion of cellular IL-10 in stimulated Burkitt's lymphoma cells [51],[52]. Patients with EBV-associated post-transplant lymphoproliferative disease also have elevated serum IL-10, but not IL-6 [53]. However, in vivo, whether the primary role of vIL-10 is to suppress the immune response, trigger B cell proliferation and differentiation or both is unclear. Serum cellular IL-10 levels are elevated both during primary EBV infection as well as during EBV reactivation from latency, implying that IL-10 plays a role in both the establishment and reactivation from latency [54].
Human cytomegalovirus (HCMV), a beta herpesvirus, encodes a viral IL-10 (cmvIL-10) that has only 27% homology to cellular IL-10, but is nevertheless capable of binding the IL-10 receptor and mediating downstream STAT1/STAT3 signaling [55]. Human cmvIL-10 is capable of downregulating MHC I and II, suppressing PBMC proliferation, and decreasing IFNγ, IL-1α, GM-CSF, IL-6, and TNF-α secretion in response to stimulation [56]. Transcripts encoding cmvIL-10 have been detected in the bone marrow and mobilized peripheral blood during natural HCMV latency, indicating that cmvIL-10 may play a role in either establishment, maintenance, or reactivation from latency [57]. Murine CMV (MCMV) does not encode an IL-10 homolog, although, parallel to the studies of Flano et al. on dendritic cells infected with MHV68 [36] , in vitro MCMV infection of primary macrophages results in secretion of cellular IL-10 and downregulation of MHC II [58]. Recently, IL-10 production by CD4 T cells has been shown to be of key importance in regulating MCMV persistence in the salivary glands. Blockade of the IL-10R resulted in a significant decrease in the titer of MCMV in the salivary glands with a concurrent increase in the frequency of IFNγ-producing CD4 T cells, indicating a role for IL-10 in the maintenance of a MCMV infection, possibly through reactivation from latency [59]. It seems reasonable to speculate that suppression of the host response may help these viruses both establish latency as well as reactivate from latency, reseeding the latency reservoir without clearance by memory T cell responses.
IL-10 has been implicated in the pathogenesis of both autoimmune as well as viral diseases, and the fact that many viruses carry IL-10 orthologs speaks to the potency of IL-10 in manipulating the host immune system [60],[61]. The parapoxvirus, ORF virus, encodes a viral IL-10 homolog with 80% homology to ovine IL-10 and is capable of inhibiting T cell proliferation [62]. Finally, two independent reports have demonstrated that blockade of the IL-10 receptor during chronic lymphocytic choriomeningitis virus infection led to clearance of the infection with enhanced IL-10 production by dendritic cells [63],[64].
We also observed a significant increase in IL-6 and MIP-1α in the B cell cultures transduced with M2. Interestingly, KSHV encodes homologs of both IL-6 and MIP-1α, suggesting that these cytokines have key roles in gammaherpesvirus immunomodulation that we have yet to appreciate [16]. MIP-1α can be detected in the BAL and lung homogenate of MHV68-infected mice at the peak of lytic replication, but the contribution of this cytokine to latency and reactivation has not been studied directly [65],[66]. During MHV68 infection, MIP-1α secretion in the lungs may attract B cells to the area of acute replication, facilitating viral infection and trafficking to the spleen. There is no significant difference in pathology or viral latency in IL-6−/− mice, despite the fact that upon ex vivo stimulation infected splenocytes secrete IL-6 [67]. Further study is necessary to explore the links between M2 and these cytokines.
Does M2-driven IL-10 secretion play a critical role in MHV68 latency? The studies presented here provide an indication that the M2 protein has multiple functions – some of which are necessary for primary murine B cells to respond to IL-10 signaling in the retroviral transduction assays we have described. Our attempts to neutralize IL-10 in the B cell culture system have been unsatisfactory (data not shown) – perhaps owing to the difficulty of neutralizing the autocrine activity of IL-10 expressed from primary murine B cells. Thus, we anticipate that attempting to neutralize IL-10 in vivo during MHV68 infection will be difficult. As a distinct approach, we have recently published the analysis of a panel of point mutations in candidate functional motifs in the M2 protein [68]. The latter studies have also provided evidence for the presence of multiple functionally important domains in M2 [68]. With respect to M2-driven B cell proliferation and IL-10 secretion, we analyzed in primary B cells three M2 mutants which, in the context of virus infection, were severely attenuated in establishment and reactivation from MHV68 latency. Notably, two of these mutations (Y129F/P7 and P8) ablated the IL-10 dependent proliferative dominance phenotype in primary B cell cultures while the other mutation (P9) was similar to wild type M2 [68]. The latter result underscores that M2 is a multifunctional protein. In addition, these studies link M2 functional domains that play a critical role in MHV68 latency in vivo to M2-driven IL-10 secretion. Further studies will be required to assess the contribution of M2-driven IL-10 expression to chronic MHV68 infection.
In summary, the analysis of M2 protein function provides a unique insight into an immunomodulatory mechanism that is employed by many viruses, particularly the herpesvirus family. Our work demonstrates that the M2 protein, a unique viral protein, manipulates B cell signaling to induce cellular IL-10 secretion and make cells more responsive to IL-10 signaling, leading to proliferation and enhanced survival of M2-expressing primary B cells in culture. M2 expression in primary murine B cells results in differentiation to a pre-plasma memory B cell phenotype, an intermediate in mature B cell development. In addition, M2 protein expression correlates with high serum IL-10 levels and an increased frequency of virus-specific CD8+ T cells during MHV68 infection. We conclude that driving B cell proliferation, survival and differentiation, while simultaneously dampening the host immune response to the virus, is an elegant immunomodulatory mechanism used by MHV68 to both facilitate establishment of latency and subsequent episodic virus reactivation from latency.
Female C57Bl/6 and IL-10−/− mice 6 to 8 weeks of age were purchased from the Jackson Laboratory. Mice were sterile housed and treated according to the guidelines at Emory University School of Medicine (Atlanta, GA). Following sedation, mice were infected intranasally with 1000 pfu of either MHV68/M2.Stop or MHV68/M2.MR in 20 µL of cMEM. Mice were infected with 100 pfu of MHV68/M2.Stop or MHV68/M2.MR in 500 µL of cMEM intraperitoneally. Mice were allowed to recover from anesthesia before being returned to their cages.
Spleens were homogenized and erythrocytes removed by hypotonic lysis. B cells were enriched using negative selection by magnetic cell separation with the mouse B Cell Isolation Kit (Miltenyi Biotech). Purity was confirmed by staining for CD19, and B cells used in experiments were 93–97% pure. Cells were cultured in RPMI-1640 supplemented with 10% FCS, 100 U/mL penicillin, 100 mg/mL streptomycin, 2 mM L-glutamine, 10 mM HEPES, 1 mM sodium pyruvate, 10 mM non-essential amino acids, and 25 µg/mL of LPS (Sigma) overnight before retroviral transduction.
BglII sites were cloned flanking the M2 ORF with primers 5′ CAG CTC AGA TCT ATG GCC CCA ACA CCC 3′ and 5′ CAG CTC AGA TCT TTA CTC CTC GCC CCA 3′ and cloned into pCR-Blunt (Invitrogen). Positive clones were sequenced, digested with BglII, and cloned into the pMSCV-IRES-Thy1.1 vector (a gift from Philippa Marrack) to construct pMSCV-M2-IRES-Thy1.1. pMSCV-M2Stop-IRES-Thy1.1 was constructed in a similar manner. Retroviruses were produced using the BOSC23 producer cell line (ATCC). 2×106 BOSC23 cells were plated on 60 mM Collagen II coated plates. The following day, 10 ug of pMSCV vector was transfected into the BOSC23 cells using the LT-293T reagent from Mirus Biotech. Retroviral supernatants were harvested 48 to 72 hours post-transfection, centrifuged at 2000 rpm for 10 minutes to clear cell debris, and supplemented with 5 µg/mL of polybrene. B cells were transduced by removing 700 µL of media and replacing it with 1 mL of retroviral supernatant/polybrene. Cells were spun at 2500 rpm at 30°C for one hour. 750 uL of retroviral supernatant was removed and replaced with fresh, complete RPMI. Cells were rested for 48 hours before analysis.
In some analyses recombinant IL-10 was added back to transduced primary B cell cultures, as follows. Primary murine B cells were harvested from C57Bl6 and IL-10−/− mice as previously described. Transduction efficiencies were measured 48 hours post-transduction by flow cytometry. Following day 2 analysis, indicated cultures received 20 ng/mL of murine recombinant IL-10 (Peprotech). B cell populations were analyzed on days 3–5 post-transduction by flow cytometry.
Cells were lysed in a suitable volume of ELB buffer on ice for 20 minutes. Lysates were pre-cleared with pre-immune chicken IgY, and M2 precipitated with chicken anti-M2 IgY followed by capture by agarose anti-IgY beads (Aves Labs, Inc.). Precipitates were run on a 15% acrylamide gel, transferred to nitrocellulose membranes, and blotted with rabbit anti-M2 antisera followed by donkey anti-rabbit HRP. Protein was detected using chemiluminescence on Kodak X-Omat Blue XB-1 film.
Rat anti-mouse CD16/32 (Fc block) was used prior to staining in most experiments. Cells were stained with the following antibodies: Thy1.1-FITC, -PE, or –APC (eBiosciences), CD44-FITC (Caltag), CD62L-PE (Caltag), GL7-FITC, IgG1, 2a, 2b, 3-FITC, CD25-PE, CD138-PE, I-Ab-PE, CD4-PerCP, CD11a-PE-Cy7, CD19-APC, B220-APC, CD8-PacficBlue (BD Pharmigen except where noted). Tetramers were synthesized at the NIH Tetramer Core Facility at Emory University and conjugated to streptavidin-APC (Molecular Probes) according to core protocol. Intracellular bromodeoxyuradine incorporation was measured using BrdU-APC according to the manufacturer's protocol (BD Pharmigen). AnnexinV-PacificBlue and 7-AAD reagents were purchased in the Vybrant® Apoptosis Assay Kit #14 (V35124) from Molecular Probes and used per manufacturer's protocol. Cells were analyzed on FACScalibur or LSR II flow cytometer. Data was analyzed using FlowJo software (TreeStar, Inc., San Carols, CA).
TranSignalTM Mouse Cytokine Antibody Arrays 1.0 (Panomics, Inc.) were used to screen for secreted cytokines as per manufacturer's instructions. Membranes were blocked in Blocking Buffer for two hours, washed, and then incubated for two hours at room temperature with day four supernatants from B cells transduced with MSCV-M2 or MSCV-M2.Stop. Membranes were washed and incubated with Biotin Conjugated Anti-Cytokine Mix as per protocol. Membranes were washed and incubated with Streptavidin-HRP. After a final wash, bound cytokine was detected using chemiluminescence on Kodak X-Omat Blue XB-1 film. Cytokines were quantitated by ELISA. IL-6 and IL-10 were detected with reagents from BD Biosciences, and IL-2 and MIP-1α were detected with reagents from R&D Biosystems. IgM and IgG were detected with reagents from Bethyl Biosciences.
Triplicate cultures of 1×106 A20 B cells were nucleofected (Amaxa Biosystems) with 4 ug of pIRES2-EGFP (BD Biosciences Clontech), pM2-IRES-EGFP [68], or pBluescriptIISK (Stratagene) using Solution T with setting T-01 on an Amaxa Nucleofector I (Amaxa Biosystems). Cells transduced with pBluescriptIISK were stimulated with 100 ng/mL of LPS following nucleofection as indicated. 48 hours post-nucleofection, supernatants were harvested and secreted IL-10 measured by ELISA (BD Biosciences).
A MHV68 genomic fragment containing the region from bp 2403 to bp 6262 (WUMS sequence) [69] was cloned into the Litmus-38 plasmid (Lit38-M2) as previously described [23]. With Lit38-M2 as a template, a stop codon was introduced into the M2 ORF using the following oligonucleotides: Oligo1 (5′ CCA CCA GGC CGA AGC TTA CGG ATT GGG AAT C) and Oligo2 (5′ CCA ATC CGT AAG CTT CGG CCT GGT GGA TG) generating a translational stop codon at bp 4566 and introducing a Hind III restriction site. The resultant product was ligated into the pCR Blunt plasmid (Invitrogen). In addition, an M2 marker rescue pCR Blunt plasmid was generated by PCR using Lit38-M2 as a template and designated as M2.MR. M2.Stop pCR Blunt plasmid and M2.MR were sequenced to verify the introduction of the site directed point mutations and the absence of unwanted mutations. Recombinant viruses were generated by allelic exchange in E. coli, as described by Smith and Enquist [70],[71]. Briefly, the Not I and Bam HI restriction sites within pCR Blunt were used to liberate the MHV68 genomic region contained within the plasmid. This fragment was cloned into the suicide vector pGS284 which harbors an ampicillin gene and a levansucrase cassette for positive and negative selection, respectively. The resulting plasmid was transformed into S17λpir E. coli cells and mated to GS500 E. coli (RecA+) containing wt MHV68 BAC. Cointegrants were selected on Luria-Bertani (LB) agar plates containing chloramphenicol (Cam) and ampicillin (Amp) and were resolved following overnight growth in LB medium with Cam. Next, bacteria were plated on LB agar plates containing Cam and 7% sucrose to select for loss of pGS284 vector sequence. Individual colonies harboring site specific point mutations within M2 were identified by colony PCR followed by restriction digest. Positive clones were grown in LB medium with Cam, and BAC DNA was purified with a Midi Prep Kit (Qiagen, Hilden, Germany) as described by the modified manufacturer's protocol. The presence of site specific point mutations and the absence of unwanted mutations within the region of homologous recombination were confirmed by sequencing and southern blot. Virus stocks were generated by Superfect (Qiagen, Hilden, Germany) transfection of recombinant MHV68 BAC DNA into Vero-Cre cells as previously described [70]. In wells showing cytopathic effect (CPE), virus was harvested, cleared of cell debris, and used to infect Vero-Cre cells in order to generate high-titer stocks. Following the presence of CPE in Vero-Cre cells, samples were harvested, homogenized, clarified, and aliquoted for storage at −80°C. Virus stock titers were determined by plaque assay as previously described [23],[72].
Limiting dilution assays for frequency of latent were performed as previously described [23],[24]. To determine the frequency of cells harboring latent viral genomes, single-copy-sensitive nested PCR was performed. Frozen samples were thawed, washed in isotonic buffer, counted, and plated in three-fold serial dilutions in a background of 104 NIH 3T12 cells in 96 well plates. Cells were lysed by protease K digestion for six hours at 56°C. Two rounds of nested PCR were performed per sample with twelve samples per dilution, and the products were resolved on 2% agarose gels. In order to measure the frequency of reactivating splenocytes, bulk splenocytes were resuspended in cMEM and plated in serial two-fold dilutions on mouse embryonic fibroblast (MEF) monolayers in 96-well tissue culture plates. Parallel samples of mechanically disrupted cells were plated to detect preformed infectious virus. Wells were scored for cytopathic effect 14 to 21 days post-explant.
The Mouse IL-10 In Vivo Capture Assay Set (BD Biosciences) was used to detect IL-10 in vivo during infections. On day 14–15 p.i., parallel groups of five mice were injected with 10 µg of biotinylated rat anti-mouse IL-10 antibody in 200 µl of sterile PBS. Mice were bled on day 15–16 p.i. and serum collected. Samples were prepared and assayed by ELISA as per protocol. The limit of detection of this assay is 31.3 pg IL-10/mL of serum.
Data analysis was conducted using GraphPad Prism software. Error bars in all graphs depict standard error of the mean. For limiting-dilution analysis, data was subjected to nonlinear regression analysis with a sigmoidal dose-response algorithm for best-fit. Poisson distribution predicts that the frequency at which 63.2% of wells are positive for an event (PCR or reactivation) is the frequency at which there is at least one event present in the population. Statistical significance of the flow cytometry and ELISA data was determined by two-tailed, unpaired Student's T test with a confidence level of 95%. |
10.1371/journal.pgen.1005987 | An Indel Polymorphism in the MtnA 3' Untranslated Region Is Associated with Gene Expression Variation and Local Adaptation in Drosophila melanogaster | Insertions and deletions (indels) are a major source of genetic variation within species and may result in functional changes to coding or regulatory sequences. In this study we report that an indel polymorphism in the 3’ untranslated region (UTR) of the metallothionein gene MtnA is associated with gene expression variation in natural populations of Drosophila melanogaster. A derived allele of MtnA with a 49-bp deletion in the 3' UTR segregates at high frequency in populations outside of sub-Saharan Africa. The frequency of the deletion increases with latitude across multiple continents and approaches 100% in northern Europe. Flies with the deletion have more than 4-fold higher MtnA expression than flies with the ancestral sequence. Using reporter gene constructs in transgenic flies, we show that the 3' UTR deletion significantly contributes to the observed expression difference. Population genetic analyses uncovered signatures of a selective sweep in the MtnA region within populations from northern Europe. We also find that the 3’ UTR deletion is associated with increased oxidative stress tolerance. These results suggest that the 3' UTR deletion has been a target of selection for its ability to confer increased levels of MtnA expression in northern European populations, likely due to a local adaptive advantage of increased oxidative stress tolerance.
| Although molecular variation is abundant in natural populations, understanding how this variation affects organismal phenotypes that are subject to natural selection remains a major challenge in the field of evolutionary genetics. Here we show that a deletion mutation in a noncoding region of the Drosophila melanogaster Metallothionein A gene leads to a significant increase in gene expression and increases survival under oxidative stress. The deletion is in high frequency in three distinct geographic regions: in northern European populations, in northern populations along the east coast of North America, and in southern populations along the east coast of Australia. In northern European populations the deletion shows population genetic signatures of recent positive selection. Thus, we provide evidence for a regulatory polymorphism that underlies local adaptation in natural populations.
| Natural populations adapt constantly to their changing environments, with alterations in protein sequences and gene expression providing the main sources of variation upon which natural selection can act. At present, understanding how changes in gene expression contribute to adaptation is one of the major challenges in evolutionary genetics. The fruit fly Drosophila melanogaster has populations distributed throughout the world, with environments ranging from tropical to temperate. On the basis of biogeographical, anatomical and population genetic studies, the center of origin of D. melanogaster has been inferred to be in sub-Saharan Africa [1–3]. Several genomic studies concluded that D. melanogaster underwent a population expansion around 60,000 years ago within Africa that set the ground for an out-of-Africa expansion 13,000–19,000 years ago and the subsequent colonization of Europe and Asia 2,000–5,000 years ago [4–6]. Because the colonization of new habitats requires that species adapt to new environmental conditions, there has been considerable interest in identifying the genetic and phenotypic changes that occurred during the out-of-Africa expansion of D. melanogaster [7–9].
In order to identify genes that differed in expression between a D. melanogaster population from Europe (the Netherlands) and one from sub-Saharan Africa (Zimbabwe), whole-transcriptome comparisons were carried out using adult males and females [10,11], as well as the dissected brains and Malpighian tubules of each sex [12,13]. These studies identified several hundred genes that were differentially expressed between the two populations and which represent candidates for adaptive regulatory evolution. One of the candidate genes that showed a large difference in expression between populations in the brains of both sexes was the metallothionein (MT) gene Metallothionein A (MtnA). MtnA lies on chromosome arm 3R (Fig 1) and belongs to a gene family of five members that also includes MtnB, MtnC, MtnD and MtnE [14,15]. Metallothioneins are present in all eukaryotes and have also been identified in some prokaryotes [16]. In general, MTs are cysteine-rich proteins, a feature that makes them thermostable, and have a strong affinity to metal ions, especially zinc and copper ions [17]. Some of the biological functions that have been described for MTs include: sequestration and dispersion of metal ions; zinc and copper homeostasis; regulation of the biosynthesis of zinc metalloproteins, enzymes and zinc dependent transcription factors; and protection against reactive oxygen species, ionizing radiation and metals [18]. In natural isolates of D. melanogaster, increased MtnA expression has been linked to copy number and insertion and deletion (indel) variation and is associated with increased tolerance to heavy metals [19,20].
In this paper we show that the expression difference of MtnA between a European and a sub-Saharan African population is not associated with copy number variation, but is associated with a derived 49-bp deletion in the MtnA 3’ untranslated region (UTR). Outside of sub-Saharan Africa, the deletion shows a latitudinal cline in frequency across multiple continents, reaching very high frequencies in northern Europe. Using transgenic reporter genes, we show that the indel polymorphism in the 3’ UTR contributes to the expression difference observed between populations. Furthermore, we use hydrogen peroxide tolerance assays to show that the deletion is associated with increased oxidative stress tolerance. Population genetic analyses indicate that MtnA has been the target of positive selection in non-African populations. Taken together, these results suggest that a cis-regulatory polymorphism in the MtnA 3’ UTR has undergone recent positive selection to increase MtnA expression and oxidative stress tolerance in derived northern populations of D. melanogaster.
A previous RNA-seq study of gene expression in the brain found MtnA to have four times higher expression in a European population (the Netherlands) than in a sub-Saharan African population (Zimbabwe) [12]. Of the members of the Mtn gene family, only MtnA showed high levels of expression and a significant difference in expression between populations (Fig 2A). To confirm this expression difference, we performed qRT-PCR on RNA extracted from dissected brains of flies from each population following the same pooling strategy used previously [12]. With this approach, we found MtnA to have 5-fold higher expression in the European population than in the African population (Fig 2B).
The RNA-seq and qRT-PCR analyses were performed on a "per gene" basis and did not discriminate between the two annotated transcripts of MtnA, which differ only in the length of their 3' UTR (Fig 1). The MtnA-RA transcript completely overlaps with that of MtnA-RB and contains no unique sequence. The MtnA-RB transcript, however, contains an extra 371 bp at the 3' end that can be used to assess isoform-specific expression. Using RNA-seq data [12], we found that the MtnA-RB isoform represents only a small proportion of total MtnA expression (1.50% in the European population and 0.13% in the African population). Thus, almost all of the observed expression difference in MtnA can be attributed to the MtnA-RA isoform. Although present at very low levels, the MtnA-RB transcript showed much higher expression (50-fold) in Europe than in Africa (S1 Table).
Previous studies found copy number variation (CNV) for MtnA in natural isolates of D. melanogaster and showed that an increase in copy number was associated with higher MtnA expression [19,20]. To determine if CNV could explain the observed expression difference between the European and the African populations, we assayed MtnA copy number in flies of both populations by quantitative PCR. We found no evidence for CNV within or between the populations (Fig 3). In both populations, MtnA copy number was equal to that of the control single-copy gene RpL32 and was about half that of the nearly-identical paralogs AttA and AttB [21], which can be co-amplified by the same PCR primers and serve as a positive control. These results indicate that CNV cannot account for the observed variation in MtnA gene expression.
To identify cis-regulatory variants that might be responsible for the difference in MtnA expression between European and African flies, we sequenced a 6-kb region encompassing the MtnA transcriptional unit (Fig 1) in 12 lines from the Netherlands (NL) and 11 lines from Zimbabwe (ZK). In addition, we quantified MtnA expression in a subset of eight lines from each population in both the brain and the gut by qRT-PCR. Across the 6-kb region, only a polymorphic 49-bp indel and a linked single nucleotide polymorphism (SNP) in the MtnA 3’ UTR showed a large difference in frequency between the populations, being this deletion present in 10 of the 12 European lines, but absent in Africa (Fig 4A). This indel was previously observed to segregate in natural populations from North America [20]. A comparison with three outgroup species (D. sechellia, D. simulans, and D. yakuba) indicated that the deletion was the derived variant. The qRT-PCR data revealed that the two European lines that lacked the deletion had MtnA expression that was similar to that of the African lines, but much lower than the other European lines. This result held for both brain and gut expression. Taken together, these results suggest that the 3' UTR polymorphism contributes to MtnA expression variation in natural populations. Furthermore, the expression variation is not limited to the brain, but shows a correlated response in at least one other tissue (Fig 4B).
To test if the 49-bp deletion in the MtnA 3' UTR has an effect on gene expression, we designed expression constructs in which the MtnA promoter was placed upstream of either a green fluorescent protein (GFP) or lacZ reporter gene. Two versions of each reporter gene were made, one with the ancestral MtnA 3' UTR sequence and one with the derived MtnA 3' UTR sequence, which has the 49-bp deletion (Fig 5A). The reporter genes were then introduced into the D. melanogaster genome by PhiC31 site-specific integration [22,23].
Our analysis of MtnA expression in the brain and gut indicated that the difference in expression observed between African and European populations is not brain-specific (Fig 4B). This is further supported by the expression of the reporter gene constructs. For the GFP reporter gene, the presence of the 3’ UTR deletion led to increased expression in both the brain and body (Fig 5B), with the difference in expression being 2.3-fold and 1.75-fold, respectively. A similar result was found for the lacZ reporter gene, where the 3’ UTR deletion led to 1.7-fold and 1.4-fold higher expression in the head and gut, respectively (Fig 5C).
MtnA shows high expression in most D. melanogaster organs, including the fat body, digestive system, Malpighian tubule, and brain [24]. Although it has been documented that MtnA and its paralogs are involved in heavy metal homeostasis and tolerance, it is poorly understood which other functions MtnA might have and in which cells it is expressed. To get a more detailed picture of MtnA expression in the brain, we examined the expression of the GFP reporter gene by confocal imaging of dissected brain tissue (Fig 6).
GFP expression driven by the MtnA promoter is evident in cells that form a mesh-like structure surrounding the brain and in between the neuropiles (Fig 6). MtnA does not appear to be expressed at a discernible level in neurons, as the cells expressing GFP do not have dendrites or axonal processes. The shape and localization of the cells expressing GFP in the brain suggest that they are glia, which provide neurons with developmental, structural and trophic support as well as with protection against toxic elements [25–27]. In a genome-wide expression profiling study it was found that MtnA is expressed in the astrocyte glial cells of larvae and adults of D. melanogaster [28]. Although we cannot be certain that MtnA expression is limited to the glia in the brain, our results provide direct evidence that MtnA is expressed in cell types other than the copper cells of the midgut and Malpighian tubules, as previously reported [29].
To better characterize the geographical distribution of the indel polymorphism in the MtnA 3' UTR, we used a PCR-based assay to screen ten additional D. melanogaster populations across a latitudinal range spanning from tropical sub-Saharan Africa to northern Europe (Table 1). We found that the deletion was at very low frequency in sub-Saharan Africa, but nearly fixed in populations from northern Europe. This suggests that, at least outside of the ancestral species range, there is a latitudinal cline in the deletion frequency. Indeed, when the sub-Saharan populations are excluded, there is a highly significant correlation between latitude and deletion frequency (linear regression; R = 0.95, P = 0.0004). This correlation still holds when the sub-Saharan populations are included (using the absolute value of latitude), but is weaker (R = 0.80, P = 0.001).
To investigate if the clinal distribution of the MtnA 3’ UTR deletion is present on other continents, we analyzed pooled sequencing (pool-seq) data from North America and Australia [30,31]. In North America, there is a significant correlation between latitude and deletion frequency (R = 0.94, P = 0.005) (Table 2). A similar pattern was seen in Australia, although data from only two populations were available. The deletion is at a frequency of 42% in Queensland (latitude 16 S) and 61% is Tasmania (latitude 42 S). The difference in deletion frequency between the two populations is significant (Fisher’s exact test, P = 0.02).
To test for a history of positive selection at the MtnA locus, we performed a population genetic analysis of the 6-kb MtnA region in the original European (the Netherlands) and African (Zimbabwe) population samples. In addition, we sequenced this region in 12 lines of a Swedish population, in which the 49-bp 3' UTR deletion was at a frequency of 100% (Table 1). Across the entire region, the Zimbabwean population showed the highest nucleotide diversity, having 1.43- and 2.50-fold higher values of π than the Dutch and Swedish populations, respectively (Table 3). Tajima’s D was negative in all three populations, and was significantly negative in both Zimbabwe and the Netherlands (Table 3). This could reflect a history of past positive or negative selection at this locus, but could also be caused by demographic factors, such as population expansion.
A sliding window analysis was performed to determine the distribution of nucleotide diversity (θ) (Fig 7A) and population differentiation (Fst) (Fig 7B) across the MtnA region. The region flanking the 3’ UTR indel polymorphism showed very low sequence variation in Zimbabwe and Sweden, but higher variation in the Netherlands. This pattern is due to the fact that the ancestral state of the indel polymorphism is fixed in the Zimbabwean population and the derived state is fixed in the Swedish population. In the Dutch population, the MtnA 3’ UTR is polymorphic for the deletion (two of the 12 lines have the ancestral state). This leads to higher nucleotide diversity than in the Swedish population, because the ancestral, non-deletion alleles contain more SNPs than the derived, deletion alleles. On average, Sweden and Zimbabwe showed the greatest population differentiation, with Fst reaching a peak in the 3’ UTR of MtnA, whereas values of Fst were lowest for the comparison of the Dutch and Swedish populations, indicating that there is very little differentiation between them (Fig 7b).
If positive selection has favored the derived MtnA allele (with the 49-bp 3' UTR deletion) in northern populations, then in this region of the genome one would expect there to be less variation among chromosomes containing the deletion than among those with the ancestral form of the allele. Indeed, this is what we observe in the Netherlands, where both alleles are segregating. Across the 6-kb region, there are 41 segregating sites within the Dutch population (Table 3). Among the 10 chromosomes with the deletion, there are 18 segregating sites, while between the two chromosomes lacking the deletion there are 23 segregating sites. This indicates that chromosomes with the deletion, which are in high frequency, shared a much more recent common ancestor. To test if this pattern differs from that expected under neutral evolution, we performed the Hudson's haplotype test (HHT) [36] using three different demographic models of the D. melanogaster out-of-Africa bottleneck for neutral simulations. Under the model of Werzner et al. [6], HHT was significant (P = 0.031). Under the models of Thornton and Andolfatto [35] and Duchen et al. [5], HHT was marginally significant (P = 0.076 and P = 0.094, respectively). These results suggest that neutral evolution and demography are unlikely to explain the observed patterns of DNA sequence variation.
To further test if the MtnA locus has experienced recent positive selection in northern Europe, we used the composite likelihood ratio (CLR) test to calculate the likelihood of a selective sweep at a given position in the genome, taking into account the recombination rate, the effective population size, and the selection coefficient of the selected mutation [37,38]. Within the Dutch population, the CLR statistic shows a peak in the region just adjacent to the MtnA 3' UTR deletion (Fig 7C). This peak was significant when the demographic models of Duchen et al. [5], Werzner et al. [6], and Thornton and Adolfatto [35] were used for neutral simulations, which provides compelling evidence for a recent selective sweep at the MtnA locus in the Netherlands population. A similar result was obtained for the Swedish population (Fig 7D), where the CLR statistic was above the 5% significance threshold determined from all three of the bottleneck models, suggesting that the selective sweep was not limited to a single population, but instead affected multiple European populations.
To test the possibility that the deletion in the MtnA 3’ UTR might have risen to high frequency as a result of hitchhiking with another linked polymorphism, we examined linkage disequilibrium (LD) across a 100 kb region flanking the MtnA locus in the Netherlands population (S1 Fig). The degree of linkage disequilibrium, r2 [39], was calculated between all pairs of SNPs present in the 100 kb region, excluding singletons. The SNP corresponding to the indel polymorphism (Fig 4a), position 53 of the linkage disequilibrium matrix, is not in significant LD with any of the 94 SNPs present along the 100 kb region analyzed (S1 Fig). These results indicate that the high frequency of the MtnA 3’ UTR deletion cannot be explained by linkage with another positively selected locus.
MtnA expression has been linked to increased heavy metal tolerance [19,20,40] and metallothioneins in general have been associated with protection against oxidative stress [18,41]. To test if MtnA plays a role in oxidative stress and/or heavy metal tolerance, we used RNA interference (RNAi) to knockdown MtnA expression; these flies, along with their respective controls, were exposed to either hydrogen peroxide or copper sulfate. A knockdown in MtnA expression was significantly associated with increased mortality in the presence of hydrogen peroxide (P < 0.001; Fig 8A) and copper sulphate (P = 0.026; Fig 9A and 9B), although for the latter, this decrease was only significant in females.
To further test if the deletion in the MtnA 3’ UTR could be associated with an increase in oxidative stress and/or heavy metal tolerance, a subset of D. melanogaster lines from the Dutch and Malaysian populations, either with or without the deletion, were exposed to hydrogen peroxide and copper sulfate. The 3’ UTR deletion was associated with a significant increase in survival in the presence of hydrogen peroxide in both the Dutch (P = 0.001; Fig 8B) and Malaysian (P = 0.001; Fig 8B) populations. The 3’ UTR deletion had no significant effect on survival in the presence of copper sulfate in Dutch and Malaysian females (P = 0.976 and P = 0.732 respectively; Fig 9D) or males (P = 0.578 and P = 0.904 respectively; Fig 9C). Thus, the deletion in the MtnA 3’ UTR was associated with increased oxidative stress tolerance, but not increased heavy metal tolerance.
Differential expression of MtnA between a European and an African population of D. melanogaster was first detected in a brain-specific RNA-seq analysis [12]. In the present study, we confirm this inter-population expression difference by qRT-PCR and show that it is associated with an indel polymorphism in the MtnA 3’ UTR. We also perform reporter gene experiments to demonstrate that a large proportion of the expression difference can be attributed to this indel polymorphism. The ancestral state of the 3’ UTR contains a 49-bp sequence that is deleted in a derived allele that is present in worldwide populations. The deletion is nearly absent from sub-Saharan Africa, but present in frequencies >80% in northern Europe (Table 1). The deletion is present at intermediate frequency in Egypt (60%), Cyprus (65%) and Malaysia (45%). These findings suggest that positive selection has favored the 3' UTR deletion, at least within northern European populations. This interpretation is supported by population genetic analyses that indicate a recent selective sweep at the MtnA locus in populations from the Netherlands and Sweden (Fig 7). Furthermore, a clinal relationship between deletion frequency and latitude is also seen in North America and Australia, suggesting that there is a common selection gradient affecting all populations outside of sub-Saharan Africa.
Although chromosome arm 3R is known to harbor inversion polymorphisms that vary in frequency with latitude in cosmopolitan populations [42], we can rule out linkage to a segregating inversion as a cause for the clinal pattern seen for the MtnA 3’ UTR deletion. A previous analysis of the same Dutch population used in our study found that only one of the isofemale lines harbored an inversion on 3R, In(3R)P [43]. This was line NL13, which is one of the 10 lines with the MtnA 3’ UTR deletion (Fig 4A). Thus, there is no evidence for linkage between the inversion and the deletion. Moreover, the MtnA gene lies 7 Mb outside of the nearest breakpoint of In(3R)P.
Using hydrogen peroxide tolerance assays, we found evidence that knocking down MtnA expression decreases oxidative stress tolerance (Fig 8B). The association of the deletion in the MtnA 3’ UTR with increased survival in the presence of hydrogen peroxide (Fig 8A) suggests that the deletion has been selectively favored in some environments because it confers increased tolerance to oxidative stress. While cytotoxic reactive oxygen species (ROSs) are generated by natural metabolic processes, they can also be introduced via abiotic factors in the environment, such as radiation, UV light or exposure to toxins. The significant correlation between the frequency of the 3' UTR deletion and latitude, coupled with its association with increased oxidative stress tolerance suggests that environmentally induced oxidative stress may vary clinally, with greater stress in northern European environments. Regulation of the oxidative stress response usually occurs via upregulation of antioxidant protective enzymes in response to the binding of a cis-acting antioxidant-responsive element (ARE), which contains a characteristic sequence to which stress-activated transcription factors can bind [41]. A recent example of adaptation to oxidative stress in Drosophila is the insertion of the Bari-Jheh transposable element into the intergenic region of Juvenile Hormone Epoxy Hydrolase (Jheh) genes, which adds additional AREs that upregulate two downstream Jheh genes and was associated with increased oxidative stress tolerance [44]. Interestingly, the Bari-Jheh insertion also shows evidence for a partial selective sweep in non-African D. melanogaster [45], suggesting that oxidative stress may have imposed an important selective constraint on the colonization of Europe. However, the MtnA 3’ UTR deletion cannot mediate its associated increase in oxidative stress tolerance in a similar way, since it does not add any new AREs.
Due to their high inducibility in response to heavy metals, metallothioneins have traditionally been thought to play a role as detoxifiers specifically of heavy metals. However, this view has come into question recently, and metallothioneins are now thought to be a part of the general stress response and may function as scavengers of free radicals [41]. The association of the MtnA 3’UTR deletion with increased oxidative stress tolerance (Fig 8A) is in line with this more recent view of the role of metallothioneins, while the observed increased mortality after copper exposure in females in which MtnA expression has been knocked down (Fig 9D) is in keeping with the more traditional view. However, we found no association between the presence of the deletion and copper tolerance. This may be because the RNAi knockdown results in an MtnA expression level that is much lower than that of naturally occurring alleles, and copper tolerance is only affected when MtnA expression falls below a minimal threshold. The precise mechanisms of how metallothioneins interact with other metal processing systems after their initial binding and help remove excess of heavy metals, remain unclear [41].
At present, the mechanism by which the 3' UTR deletion affects MtnA gene expression is unknown. Although the deletion appears to have an effect on the usage of the MtnA-RB transcript isoform (S1 Table), this isoform is too rare (<2% of all MtnA transcripts) to account for the observed 4-fold difference in MtnA expression. Another possibility is that the deleted 3' UTR region contains one or more binding sites for a microRNA (miRNA). miRNAs are short, non-coding RNAs that modulate the expression of genes by inhibiting transcription or inducing mRNA degradation [46]. They are known to bind to a seed region that consists of 6–8 nucleotides in the 3’ UTR of their target mRNA. Post-transcriptional gene expression regulation by miRNAs can result in the fine-tuned regulation of a specific transcript or can cause the complete silencing of a gene in a particular tissue or developmental stage [46–48]. To identify miRNAs that might bind specifically to the 49-bp sequence present in the ancestral form of the MtnA 3’ UTR, we used the UTR predictor [49]. The UTR predictor takes into account the three-dimensional structure of the miRNA and the 3’ UTR, as well as the energetic stability of the miRNA-3’ UTR base-pair binding. The score given by the UTR predictor is an energetic score, with the most negative scores indicating the most probable interactions. Our analysis of the MtnA 3' UTR identified five candidate miRNAs with scores below -6 that had predicted binding sites overlapping with the 49-bp indel region (Table 4). These candidates should serve as a good starting point for future functional tests of putative miRNA-3' UTR interactions.
Genetic variation provides the substrate upon which natural selection acts, resulting in an increase in the frequency of alleles that are beneficial in a given environment. Because changes in gene expression, especially those caused by variation in cis-regulatory elements, are predicted to have fewer pleiotropic effects than changes occurring within coding regions, it has been proposed that they are the most frequent targets of positive selection [50–52]. In contrast to structural changes in protein sequences, changes in gene expression can be specific to a particular a tissue or developmental stage. Our results indicate that the observed variation in MtnA expression is not specific to the brain, as a similar expression pattern is also seen in the gut (Fig 4). This suggests that the 3' UTR deletion has a general effect on MtnA expression, which is present at high levels in almost all organs of D. melanogaster [24]. However, tissue-specific effects of the difference in MtnA expression cannot be ruled out. As shown in Fig 6, GFP expression driven by the MtnA promoter in the brain is limited to what seems only one cell type, which according to their morphological and anatomical characteristics, could correspond to glia. It has been reported that glia cells protect neurons and other brain cells from ROS damage caused by oxidative stress [53,54] and the fact that MtnA has been found to be expressed in the astrocyte glial cells in larva and adult flies [28], suggests that MtnA expression in glia could serve as neuronal protection against environmental factors, such as exposure to xenobiotics, that trigger an oxidative stress response [29,55–57]. Our functional experiments showing an association between genetic variation in MtnA and oxidative stress tolerance are consistent with MtnA expression in glia providing protection against oxidative stress, which may be especially important in the brain, as neurons are highly susceptible to ROS damage.
This study used isofemale lines from 12 populations of D. melanogaster, including: Zimbabwe (Lake Kariba), Zambia (Lake Kariba), Rwanda (Gikongoro), Cameroon (Oku), Egypt (Cairo), Cyprus (Nicosia), Malaysia (Kuala Lumpur), France (Lyon), Germany (Munich), the Netherlands (Leiden), Denmark (Aarhus) and Sweden (Umeå). The lines from Zimbabwe and the Netherlands were the same as those used in previous expression studies [10–12]. Flies from Germany were collected from different locations in the greater Munich area. Flies from Cyprus were collected from a single location near Nicosia. Flies from Denmark were kindly provided by Volker Loeschcke (Aarhus University). Flies from Sweden and Malaysia were kindly provided by Ricardo Wilches and Wolfgang Stephan (University of Munich). The remaining fly lines were collected as part of the Drosophila Population Genomics Project [8] and were kindly provided by John Pool and Charles Langley (University of California, Davis).
Flies expressing hairpin RNA targeted against MtnA mRNA under the control of the GAL4/UAS system (RNAi-MtnA; transformant ID: 105011) and the host line used in their creation (control; transformant ID: 60100) were obtained from the Vienna Drosophila Stock Center [58] Act5C/Cyo flies expressing GAL4 under the control of an Act5C driver were kindly provided by Ilona Grunwald Kadow. For tolerance assays, Act5C/Cyo females were crossed to RNAi-MtnA and control males and the progeny (RNAi-MtnA/Act5C-GAL4 and control/Act5C-GAL4) were used in tolerance assays. Using qRT-PCR as described below, MtnA expression was confirmed to be knocked down by 90.03% in males and 87.58% in females in RNAi-MtnA/Act5C-GAL4 flies in comparison to control/Act5C-GAL4. Flies were maintained on standard cornmeal-molasses medium at a constant temperature of 22° with a 14 hour light/10 hour dark cycle.
Validation of the MtnA expression results obtained from brain RNA-seq data [12] was performed by qRT-PCR using TaqMan probes (Applied Biosystems, Foster City, California, USA). For population-level comparisons, six brains were dissected from males and females of each of the 11 lines from Zimbabwe (ZK84, ZK95, ZK131, ZK145, ZK157, ZK186, ZK191, ZK229, ZK377, ZK384, ZK398) and five brains were dissected from males and females of each of the 12 lines from the Netherlands (NL01, NL02, NL11, NL12, NL13, NL14, NL15, NL16, NL17, NL18, NL19, NL20). The dissected brains of each population and sex were pooled following the RNA-seq strategy previously described [12]. The above procedure was repeated in two biological replicates for each sex and population. To compare the MtnA expression of individual lines within populations, subsets of eight lines were chosen from Zimbabwe (ZK84, ZK95, ZK131, ZK145, ZK157, ZK186, ZK377, ZK384) and the Netherlands (NL01, NL02, NL11, NL12, NL15, NL16, NL17, NL18). Thirty whole brains and digestive tracts (from foregut to hindgut) were dissected per line. Two biological replicates of each line (each consisting of 30 brains or guts) were processed. Tissue was dissected from flies 4–6 days old in 1X PBS (phosphate buffered saline). The tissue was stored in RNAlater (Life Technologies, Carlsbad, CA, USA) at -80° until RNA extraction. Total RNA extraction and DNase I digestion was performed using the MasterPure RNA Purification Kit (Epicentre, Madison, WI, USA). One microgram of total RNA was reverse transcribed using random primers and SuperScript II reverse transcriptase (Life Technologies) following the manufacturer’s instructions. TaqMan gene expression assays (Applied Biosystems) were used for MtnA (Dm02362764_s1) and RpL32 (Dm02151827_g1). qRT-PCR was performed using a Real-Time thermal cycler CFX96 (Bio-Rad, Hercules, CA, USA). Two biological replicates, each with two technical replicates, were processed for each sample. The ΔΔCt method was used to compute the normalized expression of MtnA using the ribosomal protein gene RpL32 as the reference [59].
The paralogous genes AttacinA (AttA) and AttacinB (AttB) were used as positive controls for CNV assays, because they share 97% nucleotide identity [21] and can be co-amplified with the same primer set. The sequences for AttA and AttB were downloaded from FlyBase [60] and aligned using the ClustalW2 algorithm implemented in SeaView (version 4) [61]. Primers were designed for the second coding exon, where the nucleotide identity of AttA and AttB is 100%. The primer sequences were as follows: forward (5’-GGTGCCTCTTTGACCAAAAC-3’) and reverse (5’-CCAGATTGTGTCTGCCATTG-3’). The ribosomal protein gene RpL32, which is not known to show CNV, was used as a negative control. The RpL32-specific primers were: forward (5’-GACAATCTCCTTGCGCTTCT-3’) and reverse (5’-AGCTGGAGGTCCTGCTCAT-3’). The primers specific for MtnA were: forward (5’-CACTTGACCATCCCATTTCC-3’) and reverse (5’-GGTCTGCGGCATTCTAGGT-3’). CNV was assessed among 12 lines from the Netherlands and 11 lines from Zimbabwe. Individual DNA extractions were performed separately for three flies of each line and copy number was assessed individually for each fly. Genomic DNA was extracted using the MasterPure DNA Purification Kit (Epicentre). The assessment of CNV from genomic DNA was done with iQ SYBR Green Supermix (Bio-Rad) following the manufacturer’s instructions. CNV assays were performed using a Real-Time thermal cycler CFX96 (Bio-Rad). The relative copy numbers of MtnA and AttA/AttB were obtained by the ΔCt method using RpL32 as the reference gene.
Approximately 6 kb of the MtnA genomic region, spanning from the second intron of CG12947 to the 3’ UTR of CG8500 (genome coordinates 3R: 5,606,733–5,612,630), were sequenced in 12 Dutch, 11 Zimbabwean and 12 Swedish lines (Fig 1). The following primer pairs were used (all 5’ to 3’): GATGGTGGAATACCCTTTGC and AAAGCGGGTTTACCAGTGTG; GTTGGCCTGGCTTAATAACG and ACTGGCACTGGAGCTGTTTC; GCTCTTGCTAGCCATTCTGG and AGAACCCGGCATATAAACGA; GATATGCCCACACCCATACC and GTAGAGGCGCTGCATCTTGT; CACTTGACCATCCCATTTCC and CAAGTCCCCAAAGTGGAGAA; CTTGATTTTGCTGCTGACCA and ATCGCCACGATTATGATTGC; CAGGACAATCAAGCGGAAGT and TTATGAAGCGCAGCACCAGT; GACCCACTCGAATCCGTATC and TGCTTCTTGGTGTCCAGTTG. PCR products were purified with ExoSAP-IT (Affymetrix, Santa Clara, CA, USA) and sequenced using BigDye chemistry on a 3730 automated sequencer (Applied Biosystems). Trace files were edited using Sequencher 4.9 (Gene Codes Corporation, Ann Arbor, MI, USA) and a multiple sequence alignment was generated using the ClustalW2 algorithm in SeaView (version 4) [61]. All sequences have been submitted to GenBank/EMBL under the accession numbers KT008059–KT008093.
For individual flies of the isofemale lines described above, the presence or absence of the MtnA 3' UTR deletion was assessed by performing a two-step PCR (35 cycles of 98° for 5 sec. and 60° for 10 sec.) using the following primers: forward (5’-GCCGCAGACCAATTGATTA-3’) and reverse (5’-TTCTTTCCAGGATGCAAATG-3’). The frequency of the deletion was estimated on an allelic basis, as heterozygous individuals were detected in some populations. Binomial 95% confidence intervals were calculated for the frequency of the deletion using the probit method implemented in R [62]. The strength and significance of the correlation between the frequency of the deletion and latitude was determined using linear regression.
To determine the frequency of the MtnA 3’ UTR deletion on other continents, raw pool-seq reads from North America [30] and Australia [31] were downloaded from the National Center for Biotechnology Information (NCBI) short read archive (SRA). The reads were mapped to either the ancestral or derived (with 49-bp deletion) version of the MtnA 3’ UTR using NextGenMap [63]. Only reads spanning the site of the indel were considered informative. The deletion frequency was estimated as the proportion of informative reads that matched the deletion allele. The 95% confidence interval was estimated using the probit method in R [55].
To test whether the indel polymorphism found in the MtnA 3’ UTR can account for the difference in expression observed between the Dutch and the Zimbabwean populations, we constructed transgenic flies using the phiC31 transgenesis system [23]. Two expression vectors containing a green fluorescent protein (GFP) reporter gene were constructed. MtnA 3’ UTR sequences from the Netherlands (line NL20) and Zimbabwe (line ZK84), corresponding to chromosome arm 3R coordinates 5,607,448–5,611,691, were PCR-amplified with forward (5’-TTTCCTCGAACTTGTTCACTTG -3’) and reverse (5’- GCCCGATGTGACTAGCTCTT -3’) primers and cloned into the pCR2.1-TOPO vector (Invitrogen). The promoter region of MtnA (corresponding to genome coordinates 3R: 5,607,983–5,612,438), which is identical in the Dutch and the Zimbabwean populations, was also PCR amplified and cloned separately into the pCR2.1-TOPO vector using forward (5’-GCCGCAGACCAATTGATTA-3’) and reverse (5’-TTCTTTCCAGGATGCAAATG-3’) primers. To generate the GFP expression construct, the MtnA promoter was excised with EcoRI and ligated into the EcoRI site at the 5’ end of GFP in the plasmid pRSET/EmGPP (Invitrogen). Using AvaI and XbaI, the fragment containing the MtnA promoter and GFP was excised from the pRSET/EmGPP plasmid and ligated into the AvaI–XbaI sites proximal to the MtnA 3’ UTR in the pCR2.1-TOPO vector. The whole construct (promoter + GFP + 3’ UTR) was then excised with XbaI and KpnI and ligated into the XbaI–KpnI sites of the pattB integration vector [23]. For the lacZ constructs, the MtnA promoter was excised from the pCR2.1-TOPO vector with EcoRI and ligated into the EcoRI site 5’ of the lacZ coding sequence in the pCMV-SPORT-βgal plasmid (Life Technologies). PCR primers with overhangs containing restriction sites for XhoI and XbaI (forward 5’- GGTCCGACTCGAGGCGAAATACGGGCAGACATG -3’ and reverse 5’- GGTGCTCTAGAGCTCCATAGAAGACACCGGGAC -3’) were used to amplify the MtnA promoter/lacZ fragment and the product was ligated into the XhoI–XbaI sites just upstream of the MtnA 3’ UTR fragment in the pCR2.1-TOPO vector. Finally, the whole construct was excised using XbaI and KpnI and ligated into the XbaI–KpnI sites of the pattB vector (Fig 5). PhiC31 site-specific transgenesis was used to generate flies that differed only in the presence or the absence of the 49–bp sequence in the 3’ UTR of the reporter gene. The M{vas-int.Dm}ZH-2A, M{3xP3-RFP.attP}ZH-51D line was used for embryo microinjections. Microinjection and screening for transformants were carried out by Fly Facility (Clermont-Ferrand Cedex, France) and Rainbow Transgenic Flies (Camarillo, CA, USA). The successfully transformed flies were crossed to a yellow, white (yw) strain for two generations to eliminate the integrase.
Brain tissue was dissected in ice-cold 1X PBS and fixed with PLP (8% paraformaldehyde in NaOH and PBS with lysine (1)-HCl) for one hour at room temperature as described in [65]. After fixation, the tissue was washed twice for 15 minutes with PBS-0.5% Triton X and then incubated for one hour in blocking solution (20% donkey serum, 0.5% Triton X in PBS) at room temperature. The primary antibody, mouse anti-disclarge (Developmental Studies Hybridoma Bank, University of Iowa, USA) was used at a 1:200 dilution and incubated overnight at 4° Celsius in blocking solution. After washing twice with PBS-0.5% Triton X, the tissue was incubated with the secondary antibody, 1:200 anti-rat-CY3 (Dianova, Hamburg, Germany). The brains were mounted in Vectashield mounting medium (Vector Laboratories, Burlingame, CA, USA) and scanned using confocal microscopy with a Leica SP5-2. The images were analyzed using the StackGroom plugin in ImageJ [66].
Summary statistics, including the number of segregating sites (S), number of haplotypes and Tajima’s D [34] were calculated using DnaSP v.5.10.1 [67]. The mean pairwise nucleotide diversity (π) [33], Watterson’s [32] estimate of nucleotide diversity (θ) and Fst [68] were calculated as described in [5]. Hudson’s haplotype test (HHT) was carried out using ms [69] to perform coalescent simulations and psubs [70] to calculate the probability of observing a subset of n sequences containing p or fewer polymorphic sites. The demographic models of Thornton and Andolfatto [35], Duchen et al. [5], and Werzner et al. [6] were used to simulate the out-of Africa bottleneck.
To test for a selective sweep, a SweepFinder analysis was performed using the SweeD software [38]. The background site frequency spectrum (SFS) was calculated for the entire 3R chromosome arm using 11 whole genome sequences from the Netherlands population and one whole genome sequence from the French (Lyon) population [8]. The French sequence was included in order to have a constant sample size of 12 sequences for the calculation of the SFS. This approach did not bias the background, as the French sequence did not differ more from the Netherlands sequences than the Netherlands sequences did from each other (S2 Table, S2a Fig). Furthermore, the inclusion of a French line did not lead to a skew in the background SFS (S2b Fig). For the Swedish population, the background SFS of chromosome arm 3R was determined from 12 whole genome sequences from the Umeå population (S3 Table). In order to increase the power of the test, the invariant sites in the alignment were also included [37]. To assess the significance of the composite likelihood ratio (CLR) statistic, neutral simulations were performed using ms [69]. In the neutral simulations three demographic models were taken into account [5,6,35]. These models differ in several parameters, including: the timing of the out-of-Africa bottleneck, the current effective population sizes of the European and African populations, and the ancient demographic history of the African population. For our analyses, it is the estimated time of the out-of-Africa bottleneck that has the largest impact on the results. Duchen et al. [5] infer this bottleneck to have occurred around 19,000 years ago, Thornton and Andolfatto [35] around 16,000 years ago, and Werzner et al. [6] around 13,000 years ago. However, the 95% confidence intervals of the estimates are very wide, ranging from 7,359–43,000 years ago. Thus, the three estimates are not incompatible with each other. The recombination rate of the MtnA genomic region was obtained from the D. melanogaster recombination rate calculator [71]. A total of 10,000 simulations were performed. For each simulation, the maximum value of the CLR statistic was extracted and used to determine the 5% significance threshold. Linkage disequilibrium was calculated between all pairs of SNPs present using Lewontin’s r2 = D2/p1q1p2q2, where D is the frequency of the haplotypes and p and q represent the allele frequencies [39]. A fragment of ~100 kb flanking the MtnA locus (3R: 9,732,746..9,835,406) was analyzed, with singletons excluded. A Fisher’s exact test was used to assess significance of the r2 values.
Copper sulfate and hydrogen peroxide tolerance assays were performed using five D. melanogaster lines containing the MtnA 3’ UTR deletion (two Dutch and three Malaysian lines) and three lines without the deletion (two Malaysian and one Dutch line), as well as an MtnA knockdown line (RNAi-MtnA/Act5C-GAL4 and its control (control/Act5C-GAL4). Assays were performed at 25°C in tolerance chambers consisting of a plastic vial (diameter = 25 mm, height = 95 mm) with compressed cotton at the bottom containing 2.5 ml copper sulfate (Sigma Aldrich) or hydrogen peroxide (Sigma Aldrich) solution supplemented with 5% sucrose and sealed with a cork. Four to six day-old flies were separated by sex and tested in groups of 20. For each assay, one concentration of copper sulfate (50 mM) or two concentrations of hydrogen peroxide (5 or 10%) were tested with 5–7 replicates per sex and concentration. A control solution containing only sucrose was also tested with 3–5 (10–15 for Act5C-GAL4 background) replicates per sex for each assay. Mortality was recorded as the number of dead flies after 48 ± 1 hours. To determine the effect of the deletion, lines with and without the deletion were compared within each population or background. For copper sulfate assay analysis, t-tests were performed to assess significance. In order to account for potential differences in mortality inherent among the lines, proportional mortality data was scaled by mortality at 0 mM using the formula mortality/(1 + mean mortality at 0mM). For hydrogen peroxide assay analysis, the data for each assay and population was fit to a generalized linear model (GLM) using concentration, line, sex, and presence of the deletion as factors and a quasibinomial distribution using the glm function in R [62]. The tolerance results for each sex (S3 Fig) and the GLM coefficients (S4–S12 Tables) are provided as supporting information.
|
10.1371/journal.pgen.1008042 | Whole genome sequencing of experimental hybrids supports meiosis-like sexual recombination in Leishmania | Hybrid genotypes have been repeatedly described among natural isolates of Leishmania, and the recovery of experimental hybrids from sand flies co-infected with different strains or species of Leishmania has formally demonstrated that members of the genus possess the machinery for genetic exchange. As neither gamete stages nor cell fusion events have been directly observed during parasite development in the vector, we have relied on a classical genetic analysis to determine if Leishmania has a true sexual cycle. Here, we used whole genome sequencing to follow the chromosomal inheritance patterns of experimental hybrids generated within and between different strains of L. major and L. infantum. We also generated and sequenced the first experimental hybrids in L. tropica. We found that in each case the parental somy and allele contributions matched the inheritance patterns expected under meiosis 97–99% of the time. The hybrids were equivalent to F1 progeny, heterozygous throughout most of the genome for the markers that were homozygous and different between the parents. Rare, non-Mendelian patterns of chromosomal inheritance were observed, including a gain or loss of somy, and loss of heterozygosity, that likely arose during meiosis or during mitotic divisions of the progeny clones in the fly or culture. While the interspecies hybrids appeared to be sterile, the intraspecies hybrids were able to produce backcross and outcross progeny. Analysis of 5 backcross and outcross progeny clones generated from an L. major F1 hybrid, as well as 17 progeny clones generated from backcrosses involving a natural hybrid of L. tropica, revealed genome wide patterns of recombination, demonstrating that classical crossing over occurs at meiosis, and allowed us to construct the first physical and genetic maps in Leishmania. Altogether, the findings provide strong evidence for meiosis-like sexual recombination in Leishmania, presenting clear opportunities for forward genetic analysis and positional cloning of important genes.
| Leishmania promastigotes are able to undergo genetic exchange during their growth and development in the sand fly vector, however, it is still not known if they have a true sexual cycle involving meiosis. Here, we used whole genome sequencing to follow the chromosomal inheritance patterns of 44 experimental hybrids generated between different strains of L. major, L. infantum, and L. tropica. In almost every case the number of chromosomes and the allele contributions from each parent matched the inheritance patterns expected under meiosis. Rare instances of hybrid chromosomes that did not fit Mendelian expectations were observed, including gain or loss of somy, and loss of heterozygosity. Strong evidence for a meiotic-like process was also obtained from the genome wide patterns of recombination observed in the offspring generated from backcrosses involving an experimental or natural hybrid, consistent with crossing over occurring between homologous chromosomes during meiosis. The frequency and position of the recombination breakpoints observed on each chromosome allowed us to construct the first physical and genetic maps in Leishmania. The results demonstrate that forward genetic approaches are possible for positional cloning of important genes.
| Protozoan parasites of the genus Leishmania present a remarkable epidemiologic and clinical diversity, producing a spectrum of human and veterinary diseases ranging from localized, self-limiting cutaneous lesions, to more chronic and destructive mucocutaneous involvement, to disseminating, visceral infection that is fatal in the absence of treatment. Leishmania have a dimorphic life cycle consisting of extracellular promastigotes that multiply within the alimentary tract of the sand fly vector, and intracellular amastigotes that multiply within host mononuclear cells. The diversity of clinical outcomes, as well as reservoir host range and vector species compatibilities, have distinct parasite species associations, with over 20 species associated with human infections. The origins of this diversity, whether by gradual accumulation of mutations through mitotic cell division, and/or by sexual recombination producing admixtures of divergent genomes, remain a matter of considerable debate [1].
Hybridization, defined as reproduction between members of genetically distinct populations and producing offspring of mixed ancestry [2], is common in nature and has wide-ranging effects on speciation and the evolution of populations. The isolation of Leishmania strains that have been characterized as hybrids is by now well described. Multi locus genotyping using a variety of techniques identified hybrids between closely related New World species [3–7], between closely related Old World species [8–10], and between two very divergent species, L. infantum and L. major [11]. Using more discriminatory genotyping approaches, mainly whole genome sequencing, natural hybridization has also been reported at the intraspecific level for L. infantum, L. donovani, and L. tropica [12–14]. Some L. tropica strains in particular show high levels of allelic diversity and heterozygosity consistent with full genome-hybridization due to natural outcrossing. Experimentally, we and others have demonstrated that inter- and intraspecific hybrids can be generated in the sand fly vector, formally demonstrating that promastigote stages of Leishmania possess the machinery for genetic exchange [15–18]. Using pairwise combinations of parental lines expressing distinct drug resistant markers, double drug resistant lines could be recovered from sand flies co-infected with different strains of L. major, or with L. major and L. infantum, that in every case appeared to be full genomic hybrids based on their bi-parental inheritance of a limited number of allelic markers distributed across the nuclear genome. The majority of the experimental hybrids were close to diploid, though triploid and tetraploid offspring were also observed. Mating competency was confined to promastigote stages developing in the fly, and both Old and New World vector species could support hybrid formation. Based on the experimental outcrosses performed so far in which only a low frequency of co-infected flies yielded hybrids (2–20%), mating must be considered a non-obligatory part of the parasite life-cycle. Overall, the current debate regarding Leishmania reproductive strategies reflects mainly the mode of genetic exchange, its frequency and impact on population structure, not whether or not it occurs [19, 20].
Among kinetoplastid protists, the most well studied mating system is that of Trypanosoma brucei, for which a meiotic process is well supported based on the identification of a haploid parasite stage in the vector [21], and on the patterns of allele inheritance and recombination observed in experimental hybrids [22, 23]. While genome hybridization is one of the signatures of meiosis, it can also be explained by a parasexual process, as observed in some fungi [24] and proposed for Trypanosoma cruzi [25] and Leishmania [26], involving fusion of cells from both parents with generation of a transient polyploid state, followed by chromosome shuffling and random loss. True sex, incorporating meiosis with generation of haploid gametes or gamete-like cells, cell fusion or syngamy, and fusion of haploid nuclei, has not been directly observed in Leishmania, although in vitro cell fusion events were recorded in 1990 for two species, L. infantum and L. tropica [27]. Each of these processes, if they occur at all, may be difficult to detect because sex does not appear to be an obligatory stage of the life cycle, there is no obvious sexual dimorphism, and the mating competent forms are so far confined to promastigote stages developing in vivo, i.e. the sand fly midgut. While sex might be inferred from the presence and expression of meiotic gene orthologues in Leishmania [28], these orthologues can have other functions and are known to be maintained even in asexual species [29]. We have therefore turned to a genetic analysis of experimental hybrids, for which chromosome inheritance patterns expected under meiosis might be revealed, including balanced parental contributions, and recombination between homologous chromosomes. Importantly, in so far as variation in chromosome copy number is thought to be a constitutive, well tolerated, and potentially adaptive mechanism in Leishmania [30–32], then analysis of the genome structures of experimental hybrids can also reveal somy inheritance patterns and the extent to which genome hybridization is a source of aneuploid variation.
In the current studies of 3 different species of Leishmania, including L. major, L. infantum, and L. tropica, we have used whole genome sequencing to reveal the genome structures, chromosome inheritance patterns, and recombination events present in experimental intra- and interspecies hybrids. The highly predictable somy and allele inheritance patterns, and especially the genome wide recombinations observed in backcrosses involving experimental and natural hybrids, provide strong evidence for a meiotic-like sexual cycle in Leishmania.
We have previously described the recovery of hybrids from P. duboscqi and Lu. longipalpis sand flies co-infected with different pairwise combinations of L. major strains originating from across the geographic range of this species [15, 16]. For the whole genome sequencing analysis, we selected for comparison all of the hybrids previously described, plus two additional hybrids (fl6b and fl5b in Table 1A), that were generated between LmFV1/SAT, originating in Israel, and LmLV39/HYG, originating from southern Russia. We further confined the initial analysis to the hybrids that were generated in P. duboscqi, a natural vector of L. major transmission, and that had an approximate 2n DNA content, for which assessing parental inheritance is more tractable compared to the polyploid hybrids. The parental and 16 progeny clone sequences were aligned to L. major Friedlin FV1 genome Version 6 (LmjFV1.06)(http://tritrypdb.org) using Novoalign (http://www.novocraft.com/), and yielded an average 60x coverage per sample. The mapped reads were processed to obtain total read depth, reference, and alternate allele frequencies using AGELESS software (http://ageless.sourceforge.net/). The analysis identified 76103 homozygous SNPs in the LmLV39/HYG parent in comparison to LmFV1/SAT, or 0.11 SNPs/Kb, and only a relatively low number (3573) of heterozygous SNPs (Table 1A). By contrast, each hybrid possessed a high number of heterozygous SNPs (72,449–77,069; Table 1A), reflecting their bi-allelic inheritance of the approximate 76,000 SNPs that were homozygous and different between the parents.
The somy values of the parents and progeny clones were rounded off to the closest 0.5 value and depicted using a heatmap (Fig 1A). The actual somy values are not absolute integers (S1 Table), which may be attributed to the tendency of Leishmania chromosomes to show somy differences amongst cells in culture, referred to as mosaic aneuploidy [33]. Overlaid on the heatmap are the proportionate values for each parental contribution, rounded off to the closest 0.1. The profiles indicate that both parents were mostly disomic, with the exception of chromosome 31, which was pentasomic in both the parents, and chromosomes 5 and 23, which were trisomic in LmFV1/SAT. Of the chromosomes that were disomic in both parents (16x33 = 528 chromosome copies), 98% were disomic in the hybrids, and of these, 99% showed a variant allele frequency of roughly 0.5, having inherited an equal contribution from both parents, consistent with a meiosis-like process. For the two chromosomes that were trisomic in the LmFV1/SAT parent (16x2 = 32 chromosomes), we found disomic and trisomic hybrid chromosomes 41% and 53% of the time, respectively, close to expected frequencies for a meiotic process in which gametes have a roughly 50% chance of receiving either one or two copies of each trisomic chromosome. Of the trisomic chromosomes, all but 2 inherited their extra copy from the LmFV1/SAT parent, as expected. For chromosome 31, the hybrid somies ranged from 4–5, with each parent contributing two copies in most cases.
The main exceptions to the expected chromosome inheritance patterns were the 2% of chromosomes in which a new trisomic chromosome was contributed from a disomic parent (12/576 = 2%), and the 0.7% of the chromosomes for which a partial or total loss of heterozygosity (LOH) was inferred (highlighted by blue boxes). The LOH events were visualized using bottle brush plots in which the allele count at each SNP position is displayed. Bottle brush plots of representative chromosomes from hybrids showing either balanced contribution of parental alleles, gain of somy, or complete or partial LOH are shown in Fig 1B. Each partial LOH is thought to have arisen from a single crossover event that likely occurred following meiosis.
L. tropica is the causative agent of zoonotic and anthroponotic cutaneous leishmaniasis (ACL), which is endemic throughout the Middle East and in some areas of Africa and the Indian sub-continent. Prior population genetic studies have identified discreet geographic regions where L. tropica isolates possess only low levels of heterozygous SNPs, and other regions where they display extensive heterozygosity, consistent with genetic exchange [13, 34, 35]. Experimentally, the mating competency of L. tropica strains has not been demonstrated, so far as we are aware. For the generation of experimental hybrids in L. tropica, we introduced stable drug resistance markers into two different strains, LtMA37/NEO originating from Jordan, and LtL747/HYG from Israel. In comparison to one another, these strains possess 157085 homozygous SNPs, or 4.9 SNPs/Kb, and only 5756 heterozygous SNPs (Table 1C). These strains were then used to co-infect Lu. longipalpis, a non-natural vector that is permissive to L. tropica development in our laboratory colonized flies. Out of a total of 143 co-infected flies, double drug resistant hybrids were recovered from 48 flies (34%), formally demonstrating the mating competency of members of this species. Ten of the 48 double drug resistant lines recovered from 10 different flies were cloned and tested by PCR to confirm inheritance of both parental antibiotic resistance markers (labeled as LtHLMA in S1 Fig).
Two diploid clones (a and b) generated from each of five different hybrid lines were selected for whole-genome sequencing. In each case the two clones presented nearly identical genotypes, and are likely derived from the same hybridization event. All of the hybrids were equivalent to F1 progeny, heterozygous at positions where each of the parents was homozygous for a different nucleotide, resulting in each of the hybrids possessing a high density of 155428–166612 heterozygous SNPs (4.9–5.2 SNPs/Kb) (Table 1C). In the somy analysis of the parents and each of the hybrid clones (Fig 2B), all appeared to be near-diploid. Of the chromosomes that were disomic in both parents (5x33 = 165), 99% were disomic in the hybrids, and all showed variant allele frequencies of close to 0.5. Chromosome 23 was trisomic in the LtMA37/NEO parent and all but one of the hybrids, with LtMA37/NEO contributing two copies to the trisomic chromosome, as expected. Chromosome 31 was tetrasomic in all samples, and showed variant allele frequencies of 0.5 in all of the hybrids. Thus, the parental somy contributions were non-random and 99% of the time met expectations of chromosomal segregation during a meiosis-like process. The exceptions were the hybrid clones LtHLM4a/b that were trisomic at chromosome 4 despite being disomic in both parents, possibly a result of chromosomal non-disjunction. No LOH was observed in any of the hybrids.
Altogether, the whole genome sequencing of the experimental hybrids generated within and between three different Old World Leishmania species demonstrate that the progeny clones are near full genomic hybrids, with highly predictable somy and allele inheritance patterns that are strongly consistent with a meiotic-like process. Rare instances of chromosomes showing a gain of somy or loss of heterozygosity were also observed.
To generate the first experimental backcross hybrids in Leishmania we chose a L. major hybrid, 1.16.A1, generated between LmFV1/SAT and LmLV39/HYG with ploidy close to 2n, and that demonstrated robust growth in culture and in flies. The mating studies involved co-infection of P. duboscqi flies with 1.16.A1 and L. major lines stably transfected with a third antibiotic resistance marker, blasticidin-S deaminase (BSD), and selection for midgut promastigotes that were doubly drug resistant to either SAT + BSD or to HYG + BSD. A series of 4 independent backcross experiments involving 1.16.A1 and LmFV1/BSD resulted in a low frequency of flies yielding hybrids (Table 2, expts 1–4). Of a total of 316 midguts from co-infected flies that could be evaluated for hybrid recovery, double drug resistant lines could be recovered from 2 flies (0.6%). PCR tests confirmed that both progeny clones had in fact inherited all three drug resistance markers (S2A Fig). In two experiments there were sufficient flies to include for comparison co-infections involving the parental lines, which in prior studies have produced an average of 11.3% hybrid recovery [15, 16]. In this case, co-infections with LmFV1/BSD and LmLV39/HYG produced recoverable hybrids in a total of 10 of 75 flies (13.3%). PCR tests confirmed their inheritance of both selectable markers (S2B Fig).
Since intraspecies F1 hybrids have so far been generated by outcrossing L. major strains of discrete geographic origin, we considered the possibility that the mating efficiency of the F1 hybrid might be improved if outcrossed with a third L. major strain unrelated to the original parents. We have previously confirmed the mating competency of a L. major strain originating in Senegal, West Africa, LmSd/BSD [16]. In the two outcross experiments involving flies co-infected with 1.16.A1 and LmSd/BSD, hybrids were recovered from 3 of 201 flies (1.5%) (Table 2, expts 5&6), all of which were PCR positive for SAT and BSD (S2C Fig). Co-infections with LmSd/BSD and LmLV39/HYG yielded a higher frequency of hybrid recovery from 7 of 31 flies (22.6%) (Table 2, expts 6, S2D Fig). Together, the backcross and outcross mating attempts indicate that while not sterile, the intraspecies F1 hybrid had reduced fertility compared to the parents used for their generation.
The availability of 2 backcross and 3 outcross progeny allowed us for the first time to test for the presence of recombination events characteristic of meiosis in other organisms. To extract recombination patterns in outcrosses, we considered a total of 37368 markers that were common to the LmFV1/SAT and LmSd/BSD parental lines, but homozygous different from the LmLV39/HYG line, effectively treating the outcross progeny as backcrosses. The parental inheritance patterns were depicted as bottle brush plots and the recombination loci were determined visually. The schematic in S5A Fig shows the possible inheritance profiles for the backcrosses and outcrosses, assuming a maximum of 2 recombinations based on random assortment, crossovers and selection. The actual profiles of the backcrosses and outcrosses indicate that each of the expected profiles was observed (representative examples shown in S5B Fig). We incorporated the bottle brush plots into circos plots [36] depicting the allelic contributions genome wide, revealing regions of homozygosity and heterozygosity that allowed us to visualize the recombination patterns (Fig 3A). Based on these plots, we compiled the recombination loci on each chromosome to generate a physical map (S2 Table, Fig 3B). Each of the backcrosses and outcrosses had between 17 and 25 recombination events across 36 chromosomes, 21 recombinations per genome on average, or 1 recombination / 1.54 Mb. Backcrosses 1 and 2 each had 17 chromosomes with a single cross-over, and 1 or 3 chromosomes, respectively, that had two cross-overs. Outcrosses 1,2, and 3 had 17 single and 1 double, 15 single and 1 double, and 23 single and 1 double recombinations, respectively. We found a total map size of 1840 centimorgan (cM) across the 32 Mb genome which translates to 1 cM per 17,391 bp. Physical length correlated with genetic length using a linear regression model (p = 0.003). Too few backcrosses and outcrosses were available for sequencing to draw meaningful conclusions regarding recombination hotspots or coldspots.
To test the mating competency of natural hybrids in L. tropica, we introduced stable drug resistance markers into two strains, LtKub/SAT from Syria and LtRup/HYG from Afghanistan, that in our prior studies were each found to possess high levels of heterozygous SNPs (approximately 100,000 SNPs in comparison to the L. tropica L590 reference genome) in patterns consistent with these isolates being naturally occurring full genome hybrids [13]. Both strains showed robust growth in culture and in Lu. longipalpis flies (S6 Fig). Hybridization of these lines with each other or with LtMA37/NEO or LtL747/HYG, was tested in 3 independent experiments (Table 3, expts 1–3). Of the 251 flies that were co-infected with LtRup/HYG and either LtKub/SAT, LtL747/HYG, or LtMA37/NEO, no hybrids were recovered (Table 3, expt 1). By contrast, co-infection of the same population of flies with LtMA-37/NEO and LtL747/HYG, again yielded a high rate of hybrid recovery in 29 of 69 flies (42%). Ten of these hybrids were cloned and genotyped by PCR to confirm their hybrid genotypes (labeled as LtHLMB in S1 Fig). Mating attempts involving the other natural hybrid, LtKub/SAT, in two experiments yielded a total of 56 hybrid lines when paired with either LtL747/HYG or LtMA37/NEO (Table 3, expts 2&3), all of which were cloned and their hybrid genotypes confirmed (labeled respectively as LtHKLA/B hybrids in S3B Fig, and LtHKMA/B hybrids in S3A Fig). A high rate of hybrid recovery was again obtained from the same population of flies co-infected with LtMA37/NEO and LtL747/HYG (65%), 10 of which were cloned and their hybrid genotypes confirmed (labeled as LtHLMC in S1 Fig). Thus, the exceptional outcrossing efficiency in reciprocal matings of two predominantly homozygous L. tropica strains was reinforced, while the ability of two natural hybrids to mate was strain dependent, with one strain essentially sterile and the other showing good mating compatibility when tested in outcrosses with the homozygous strains.
Eight progeny clones generated from the crosses between Lt/Kub/SAT and LtL747/HYG, and 9 clones generated between LtKub/SAT and LtMA37/NEO, designated LtHKM or LtHKL, respectively, were submitted for whole genome sequencing. This allowed us to study genome wide patterns of chromosome segregation and recombination involving a natural hybrid. The sequencing identified a high number of both heterozygous and homozygous SNPs in LtKub/SAT that were different in comparison to either Lt747/HYG or LtMA37/NEO and that were passed on to the progeny clones (Table 1D & 1E). When we enumerated the somies and the parental contributions of the chromosomes in the 567 hybrid chromosomes where each of the three parents were approximately disomic (estimated somy between 1.6 and 2.5), we found that 97% were disomic with an equal contribution from both parents, as expected under meiosis (S1 Table, S7 Fig). However, we also found a few trisomic chromosomes (14 or 3%) where an extra copy was contributed by one of the disomic parental lines, and 4 chromosomes where only the homozygous parent contributed to the hybrid (blue boxes). These 4 chromosomes were monosomic in each case, and we speculate that the chromosome contribution from the LtKub/SAT parent was lost as a consequence of a non-disjunction event during meiosis. Aneuploidy was observed in the chromosomes where one or both the parents had somies greater than 2, although of the 36 chromosomes for which the LtKub/SAT parent was trisomic (chromosomes 12 & 5) and should have had an equal opportunity to contribute a double copy, a single copy contribution was observed 92% of the time. The skewed single copy contribution of the trisomic chromosomes might have occurred by chance, or perhaps by haplotype selection during adaptation to clonal growth in culture or in the fly [32].
We next constructed the zygosity profiles of the hybrids between LtKub/SAT and LtL747/HYG or LtMA-37/NEO. We identified the SNPs against the L. tropica L590 reference genome using SAMtools utility [37] and filtered out all the markers with coverage less than 10. We tagged the SNPs as homozygous if the major allele frequency was greater than 90% and as heterozygous if both the major and minor allele frequencies were between 15% and 85%. We divided the genome into blocks of 5kb and enumerated the heterozygous and homozygous SNP counts within each window. We colored the block red if the heterozygous proportion within the block was greater than 90%, blue if the homozygous proportion was greater than 90% and yellow otherwise. LtKub/SAT was mostly heterozygous while the other parental strains, LtMA37/NEO and LtL747/HYG were homozygous, as expected (Fig 4). By contrast, the outcrosses contained blocks of long runs of homozygosity, heterozygosity, or sequences that were neither homozygous nor heterozygous, similar to the patterns that are observed in the backcross progeny (see Fig 3). This suggested that the LtKub hybrid shares haplotypes with those present in LtMA37 and LtL747, and therefore might be a natural cross between strains containing genotypes similar to these strains. To test this possibility, we used a new population genetics software (https://popsicle-admixture.sourceforge.io) and performed the analysis as follows: we leveraged 159900 homozygous SNPs in LtMA37/NEO against the reference strain (TritrypDB L590 V.33) and removed 134846 homozygous SNPs that were common with LtL747/HYG. The SNPs at the remaining 23775 markers were homozygous and different between LtMA37/NEO and LtL747/HYG. We created separate Circos plots for the two groups of outcross progeny, and redrew the plots by coloring the SNPs as green if they matched Lt747/HYG, red if they matched LtMA37/NEO, and yellow if they were heterozygous (Fig 5). As expected, Lt747/HYG and LtMA37/NEO were both homozygous and contained alternate genotypes. LtKub/SAT was mostly heterozygous due to contributions from both Lt747- and LtMA37- like genotypes. The outcrosses between Lt747/HYG and LtKub/SAT contained longs runs of heterozygous and homozygous SNPs similar to backcrosses, and the homozygous regions matched Lt747/HYG. Similarly, the outcrosses between LtMA37/NEO and LtKub/SAT contained long runs of heterozygous and homozygous SNPs for which the homozygous regions matched LtMA37/NEO. These results strongly support the hypothesis that LtKub is a hybrid between strains that contained alternate genotypes matching Lt747 and LtMA37.
The transitions between the long runs of heterozygous and homozygous regions are the recombination breakpoints (S2 Table), as highlighted on the Circos plots in which both single, double, and triple crossovers were observed. The hybrids between LtKub/SAT and LtL747/HYG, and between LtKub/SAT and LtMA37/NEO recorded an average of 22 recombinations each (S2 Table), which translates to 1 recombination every 1.45Mb. Double and triple crossovers were routinely observed in the hybrids between LtKub/SAT and LtMA37/NEO in comparison to the hybrids between LtKub/SAT and LtL747/HYG (S2 Table; p-value of 0.0082 using t-test). The recombinations observed across the 17 hybrids plotted by chromosome indicated that although the crossover points were distributed throughout the genome, certain hotspots (more than 2 recombinations in a 50kb window) were observed in 18 chromosomes (Fig 6A). The larger chromosomes on an average contained more recombinations in comparison to the smaller chromosomes as evaluated by linear regression (p-value of 0.009). We translated the observed recombinations into cM distances by calculating probability of finding recombinations across the genome in sliding non-overlapping blocks of 20Kb (see methods) and drew a genetic map based on the blocks that contained recombinations (Fig 6B). We found a total map size of 2091.4 cM across the 32 Mb genome, which translates to 1 cM per 15,300 bp, very close to the recombination frequency that we recorded for L. major.
Lastly, a series of backcross and outcross mating attempts involving the interspecies hybrids were undertaken in Lu. longipalpis flies. In 8 independent experiments (Table 4), the flies were co-infected with a pool of the 4 near diploid F1 hybrids (H2, H4, H7, H6), or with each of the hybrids individually, and paired with either LmFV1/BSD or LmSd/BSD. The midgut promastigotes were selected for growth in either SAT + BSD or HYG + BSD. From a total of 420 co-infected flies that could be evaluated for hybrid recovery, no backcross or outcross progeny were obtained (Table 4). Thus, the interspecies hybrids appear to be sterile under the mating conditions employed. By comparison, when the flies were co-infected with the parental lines LiL/HYG and LmSd/BSD, a total of 10 of 140 flies (7.1%) yielded hybrids, with recovery rates in the different experiment ranging from 3% to 23%. PCR tests confirmed that the progeny clones generated from these 10 lines contained both parental selectable markers (hybrids labeled as LimHLS, S4 Fig).
We present here the first comprehensive analysis of experimental intra- and interspecies hybrids in Leishmania by analyzing high-resolution whole genome sequencing data. We determined the chromosomal somy and studied the parental inheritance of 44 hybrids generated within and between different Old World species of Leishmania, including L. major, L. infantum, and L. tropica, and compared them against the inheritance patterns expected under meiosis. The somies and parental chromosomal contributions matched the expected inheritance patterns 97%-99% of the time, supporting a predominant meiotic-like process in Leishmania, which we believe is the most parsimonious interpretation of the genome-wide inheritance patterns presented in this report. The hybrids appeared equivalent to F1 progeny, heterozygous throughout most of the genome for the homozygous alleles that were different between the parents. The majority of the hybrid clones that we have generated and analyzed in this report were near diploid, showing balanced segregation of the chromosomes, the majority of which were disomic in the parents. Trisomic chromosomes in the parents were passed on to the progeny in single or double copy, never in their original trisomic state, and in frequencies expected by Mendelian segregation. Tetrasomic chromosomes were passed on in double copy and not in quadruple copy. Such predictable, balanced allotments of parental chromosomes during hybridization seem highly unlikely to have arisen by a random, parasexual process. While it is true that the meiotic intermediates in aneuploid lines of Leishmania would not be strictly haploid, aneuploidy is not incompatible with meiosis. Meiotic chromosome segregation is well described in triploid strains of S. cerevisiae which accurately produce viable tetrads containing 2 spores with 2 copies and 2 spores with 1 copy of each homolog [38]. Furthermore, all 3 copies of a trisomic chromosome in a close to diploid strain of S. cerevisiae were shown to undergo recombination in a single meiosis [39]. In Drosophila, although the fertility of triploid females is reduced, viable offspring between diploids and triploids can be readily obtained [40]. Finally, it is worth noting that some women with non-mosaic trisomy 21 are fertile and pass on the extra chromosome to a high proportion of their offspring [41].
Polyploid, predominantly triploid hybrids, have been routinely recovered from experimental matings in Leishmania [15–17], and are observed across organisms capable of sexual reproduction, such as amphibians [42], plants [43], and fruit flies [40]. The genomic analysis of 4 interspecies hybrids with close to 3n DNA content showed parental contributions consistent with syngamy between a parental ‘2n’ cell that failed to undergo meiosis and a ‘1n’ cell from the other parent, similar to what has been suggested for T. brucei [44] for which triploid progeny are also common. Their inheritance profiles again argue against a parasexual process involving random chromosome loss from a tetraploid intermediate that would seem highly unlikely to produce progeny for which almost all of the chromosomes that were disomic in each parent were allotted one extra copy, with the extra copies within each clone always contributed by the same parent.
Our recovery of a few hybrids with close to 4n DNA content, one of which was analyzed here, is interesting because they suggest that fusion of diploid cells producing a tetraploid hybrid can indeed occur in Leishmania. The progeny clone analyzed remained close to tetrasomic for the majority of chromosomes. While it might be argued that this hybrid represents a parasexual intermediate that has not yet experienced chromosome loss, a stronger case might be made for a tetraploid meiotic cycle, originally described for Saccharomyces [45], for which diploid cells of two different mating types fuse and then undergo meiosis followed by fusion of haploid nuclei to produce diploid progeny. The triploid and tetraploid hybrids in Leishmania might be the result of one or both of the parental nuclei failing to undergo meiosis following fusion of the diploid cells. While the tetraploid meiotic cycle does not involve the generation of gametes, it is still referred to in the context of a sexual reproductive cycle in yeast [46]. Sexual reproduction can be generalized to mean all forms of meiotic reproduction in protists, which frequently retain the ability to reproduce asexually via mitosis.
The instances of unbalanced chromosome inheritance that might support a parasexual process in Leishmania were the exceptions, not the rule. Specifically, they were manifested as a gain of somy in comparison to either parent, observed approximately 2% of the time, and uniparental inheritance, observed approximately 1% of the time, that we interpret as LOH subsequent to the original hybridization event. Non-disjunction of one parental chromosome at meiosis or during subsequent mitotic generations in the fly or in culture seem more likely to explain the instances of trisomy than a parasexual process, which would be expected to lead to high chromosome dosage genome-wide. An error-prone meiosis would not be unique to Leishmania genome biology. For example, the fungus Candida lusitaniae has a well defined sexual cycle during which aneuploid hybrids frequently arise [47]. Alterations in chromosome copy number are also known to arise during in vitro cultivation of Leishmania promastigotes [48, 49], so may have occurred during mitotic division in the fly or during the in vitro selection and cloning procedures. The emergence of new aneuploidies during clonal growth may reflect a process of haplotype selection [32], driven by adaptation to the growth conditions in the fly or in culture. Mosaic aneuploidy can lead to LOH during mitotic divisions [26], or a partially homozygous state should crossing over events also occur.
Meiosis also involves frequent reciprocal crossover events between homologous chromosomes, and is considered an essential process for the two homologues to establish the physical connections needed to orient properly during the first meiotic division [50]. Whole genome sequencing of the progeny generated from crosses involving an experimental F1 hybrid of L. major, or a natural hybrid of L. tropica, bearing in each case a high number and genome wide distribution of heterozygous alleles, provides the first clear evidence for recombination events in Leishmania. The very high levels of interhomolog recombination observed have not been reported in parasexual organisms, so far as we are aware. For example, while recombination between homologous chromosomes was observed during the parasexual cycle in C. albicans, only 3 of 13 progeny strains showed evidence of mitotic crossing-over [24]. By comparison, of the backcross and outcross hybrids that were generated in the present study and for which homologous recombination could be properly assessed, all 22 revealed genome wide crossover events. This analysis allowed us to calculate the best estimate to date of recombination frequencies, and to construct the first physical maps of recombination break points in Leishmania. The backcross and outcross progeny generated from the L. major hybrid revealed discrete blocks of homozygous and heterozygous SNPs throughout the genome that would be expected from meiotic recombination and random segregation of the homologous chromosomes in the hybrid parent. The haplotypes of the heterozygous alleles in LtKub/SAT implied a relationship to both the LtMA37/NEO and LtL747/HYG parents with which it was crossed, reflected in the progeny genotypes showing discrete blocks of homozygosity and heterozygosity that would be expected from backcross matings. The recombination frequencies for the L. major and L. tropica backcrosses were similar and averaged 1 cross-over per 1.5Mb. The only other estimate of recombination frequency in Leishmania, 1 cross-over event per 2.3Mb per generation, was calculated from the levels of recombination observed in natural sand fly and human isolates of L. infantum obtained from a transmission focus in Turkey [14]. For our experimental crosses, the average recombination unit for L. major and L. tropica was 1 cM per 17,391 bp and 15,300 bp, respectively. These metrics compare well with that calculated for Trypanosoma brucei [23], which has a map length of 733 cM across the 17.89 Mb genome, or 1 cM per 24,406 bp.
Our studies of the mating compatibilities of experimental F1 hybrids suggested that the intraspecies L. major hybrid had reduced mating competency in comparison to the parents used for its generation, while each of the interspecies L. major x L. infantum hybrids were mating incompetent, at least under the mating conditions employed. Reproductive barriers arising as a result of hybrid sterility are well described across phylogeny, and are the most common form of postzygotic reproductive isolation in plants [51–53]. In Saccharomyces yeast, F1 hybrids can reproduce normally by asexual budding, however, the spores they make are inviable and sexually sterile [51]. A variety of underlying mechanisms have been described to account for yeast hybrid sterility, including chromosomal rearrangement [54], aneuploidy [55], and sequence divergence acted on by mismatch repair [56], where pairing, recombination and segregation fail because the parental homologous chromosomes are not sufficiently alike. It is interesting that for the cross-species hybrids, which are so far ‘sterile', the sequence divergence between the hybridizing genomes of the parents responsible for their generation is approximately 10-fold greater than the sequence divergence between the parental genomes of the intraspecies hybrids, which remain mating competent, though at reduced efficiency. It is notable, however, that the two presumed natural L. tropica hybrids, LtKub/SAT and LtRup/HYG, showed vastly different mating potentials despite their similar levels of heterozygosity. In T. brucei, the success of experimental backcrosses was also variable, suggesting a range of mating compatibilities among F1 progeny [57]. Our experience with LtRup/HYG aside, the successful crosses involving 3 other L. tropica strains has provided the first direct demonstration of genetic exchange between members of this species, and strongly supports the conclusion that the hybrid genotypes found especially prevalent in natural populations of L. tropica are in fact products of meiotic outcrossing events [13, 34, 35].
In summary, the whole genome sequencing analysis of experimental hybrids generated within and between different Old World species of Leishmania, provides the clearest evidence to date that the system of genetic exchange is Mendelian and involves meiosis-like sexual recombination. These findings, along with the demonstration that it is possible to backcross Leishmania in the laboratory, show that the forward genetic tools are available for linkage studies and positional cloning of important genes.
A summary of the parental lines that were used to generate the experimental hybrids analyzed in this report, including their heterozygous drug resistance loci, is provided in S3 Table. The following lines of L. major were used: LmFV1/BSD (MHOM/IL/80/Friedlin) was originally isolated from a patient with cutaneous leishmaniasis (CL) acquired in the Jordan Valley, and is heterozygous for an allelic replacement of LPG5B on chromosome 18 by a blasticidin S-resistance (BSD) marker [58]; LmSd/BSD was previously described [16] and is derived from a strain isolated from a patient with CL acquired in Senegal (MHOM/SN/74/SD) [59] and contains the same allelic replacement of LPG5B; LmLV39/HYG [15] is derived from a strain originally isolated from a reservoir rodent host in southern Russia (MRHO/SU/59/P-strain) [60] and is heterozygous for an allelic replacement of on chromosome 24 by a hygromycin B–resistance cassette [58]; Lm1.16.A1 is a previously described intraspecies hybrid [15] generated by a cross between LmFV1/SAT and LmLV39/HYG. LiL/HYG [17] is derived from L. infantum (MHOM/ES/92/LLM-320; isoenzyme typed MON-1) [61], isolated from a human case of visceral leishmaniasis (VL) in Spain, and is heterozygous for a hygromycin B resistance cassette integrated into the 18S rDNA locus. Hybrids LimH2, 4, 6, 7 and 10 were previously generated [17] interspecies hybrids from crosses between LiL/HYG and LmFV1/SAT. The following L. tropica lines were generated from strains described previously [13]: L. tropica LtMA37 (MHOM/JO/94/MA37) was originally isolated from a patient with CL in Jordan and is heterozygous for either a G418 (LtMA37 /NEO) or a hygromycin B (LtMA-37/HYG) resistance cassette integrated into the 18S rDNA locus on chromosome 27 using the constitutive expression vectors pLEXY neo2.1 or pLEXY hyg2.1, respectively (Jena Bioscience EGE 273 and EGE-272); LtL747/HYG (MHOM/IL/02/LRC-L747) was originally isolated from a patient with CL in Israel and is heterozygous for a hygromycin B resistance cassette, using the pLEXY hyg2.1 vector; LtKub/SAT is derived from a strain isolated from a patient with CL in Syria (MHOM/SY/?/Kub) and is heterozygous for a Nourseothricin resistance cassette (SAT) also integrated into the 18S rDNA locus on chromosome 27 using the pLEXY sat2.1 vector (Jena Bioscience EGE-274); LtRup/HYG and LtRup/NEO are derived from a strain originally isolated from a patient with CL in Afghanistan (MHOM/AF/87/RP) and are heterozygous for either a hygromycin B or G418 resistance cassette using the pLEXY hyg2.1 or pLEXY neo2.1 vectors, respectively. All lines were grown at 26o C in medium M199 as described previously [16] with or without 25 μg/ml hygromycin B (HYG), 100 μg/ml nourseothricin (SAT), 50 μg/ml G-418 (NEO), 25 μg/ml blasticidin S or a combination of these drugs as necessary. The hybrid clones underwent approximately 25–28 generations in vitro prior to sequencing.
Lutzomia longipalpis and Phlebotomus duboscqi sand flies, collected from field specimens in Brazil and Mali, respectively, were infected by artificial feeding through a chick skin membrane on heparinized mouse blood as previously described [16]. The mouse blood was obtained in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals at the NIH. The protocols were approved by the Animal Care and Use Committee of the NIAID (protocol LPD 68E).
For the generation of hybrids, the blood was seeded with an equal mixture of the two parental lines containing independent drug resistant markers at 4–8 x106 logarithmic phase promastigotes /ml of blood. Midguts were dissected after 7–11 days, and hybrid parasites were recovered by double drug selection as described previously [16]. For the generation of backcross (BC) or outcross hybrids (OC) using an F1 hybrid as one of the parents, flies were coinfected with the F1 hybrid clone (already double drug resistant) and one of the parental lines or an unrelated parent, each harboring a third antibiotic resistance marker. Flies were dissected and the guts were homogenized in M199 and plated in 96-well culture plates without selection. One day following plating, each well was subdivided into two wells containing either one of the drugs for which the F1 parent is resistant, and the third drug for which the other parental line is resistant (e.g. HYG-BSD and SAT-BSD). Notably, some of the BC or OC lines were resistant to all three drugs. Double/triple drug resistant BC and OC lines were cloned as described previously [15].
Total DNA content and inferred ploidy was determined on the different parental lines and hybrid progeny clones (see previous section) using flow cytometry following staining of permeabilized RNAse-treated cells with propidium iodide, as previously described [16]. Genomic DNA was extracted from the parental lines and hybrid clones using DNeasy blood & Tissue kit (Qiagen #69506) according to the manufacturer instructions. PCR amplification of the drug selectable markers was carried out as previously described [17] using high fidelity PCR mix MyFi 2X (Bioline #25050). PCR products were cleaned using ExoSAP-IT, PCR cleanup reagent (ThermoFisher #78200). Amplified products were verified by electrophoresis on a 1.5% (wt/vol) agarose gel and visualized by ethidium bromide staining.
DNAs extracted from most of the lines (see previous section) were submitted to whole genome sequencing in Rocky Mountain Laboratory’s Research Technologies Section-Genomics Unit, Division of Intramural Research, NIAID, NIH. DNAs were sheared using the Covaris LE220 for a 200bp insert target size and the fragmented genomic DNA was purified using AMPure XP beads. Libraries were generated with the ThruPLEX DNA-Seq kit (Rubicon Genomics) and quantified using the Kapa SYBR FAST Universal qPCR kit for Illumina sequencing (Kapa Biosystems, Boston, MA). The libraries were diluted to 2 nM stocks and pooled equally. The 2 nM pooled stock was prepared for sequencing by denaturing and diluting to 11 pM for clustering to a Rapid flow cell. On-board cluster generation and paired-end sequencing was completed on the HiSeq 2500 (Illumina, Inc, San Diego, CA) using a Rapid Paired End cluster kit and 200 cycle sequencing kit. The cluster density averaged 990 k/mm2 per lane resulting in 163 M reads passing filter per lane with 95% > Q30. The samples were sequenced to target 60x coverage, and achieved median coverages across the genome per sample of between 19 and 38x coverage.
Similar methods were used for DNA extracted from the 17 intraspecies L. tropica progeny and parental clones which were sequenced in DNA pipelines at Wellcome Sanger Institute (WSI). Differences were that DNA was sheared to a 400bp insert target, and libraries were generated with the NEBNext DNA Library Prep kit (New England BioLabs). These libraries were diluted to 4 nM stocks, and the 4nM pooled stock was diluting to 8 pM for clustering. The cluster density averaged 550 k/mm2 per lane resulting in at least 100 M reads passing filter per lane with 95% > Q30. The raw data was deposited in the European Nucleotide Archive (ENA) in study ERP111791. These samples were sequenced to target 30x coverage, and after removal of duplicates, achieved median coverages across the genome per sample of between 19 and 38x coverage.
All samples were subjected to whole genome sequencing using the Illumina Hi-Seq 2500 platform. The 100bp paired end reads were aligned to the closest available reference genome (Tritrypdb, L. major FV1 Version 6, L. infantum JPCM5 Version 6, L. tropica L590 V. 33) using novoalign software, and using the parameters -F (format) ILMFQ for Illumina 1.3 pipeline, -H enables hard clipping that removes low quality bases from the 5’end, -g 40 for gap opening penalty, -x 6 for gap extend penalty, -R 5 for difference in score between the best alignment and the second best alignment, -r none ignores the reads with similar scores to multiple locations in the genome, -e 1000 for maximum number of alignments for a single read.
Somies were determined using AGELESS software (http://ageless.sourceforge.net/) by dividing the chromosomes into blocks of 5kb and average coverage within each block was calculated. Punctate regions with coverage greater than twice the average coverage for each chromosome or less than half the average coverage for each chromosome were ignored for somy determination. These regions are associated with repeats, gene families and other genomic noise that alter the somy values. The blocks with zero coverage were ignored from further analyses as they indicate noisy/repeat regions filtered in the previous step. The average of block coverages was scaled to the ploidy of the organism to determine the somies.
The alleles in each sample were also determined using AGELESS software. The nucleotide compositions at loci with coverage greater than or equal to 5 were considered and the rest were reported as missing data. The loci that met the minimum coverage criteria were queried for allele frequencies. Only those alleles with frequencies between 0.15 and 1 were reported; alleles with frequencies less than 0.15 were ignored as noise. The alleles with frequencies greater than 0.90 were tagged homozygous and those with frequencies between 0.15 and 0.85 were tagged heterozygous.
Parental chromosomal inheritance patterns determined using AGELESS indicate how many chromosomes were inherited from each parent. The homozygous SNP differences between the parents were used as markers and frequencies of alleles inherited from each parent were calculated in the hybrids. These parental allele frequencies when multiplied by somy reveal the number of chromosomes inherited from each parent.
The analysis of the interspecies hybrids was limited to genomic regions that were conserved between the species. The conserved regions were determined by a read mapping-based strategy, and represented approximately 92% of the genomes of each species. The reads from one parental line were aligned to the reference genomes of both the species, L. major FV1 Version 6 and L. infantum JPCM5 Version 6. The reads that aligned to both the genomes were retained, and independently, the reads from the other parental line were also aligned to the genomes of both species. The conserved regions from this alignment were compared against the conserved regions from alignments to the other parent to retain the regions common to both the alignments.
We constructed the genetic map based on the recombinations observed in L. tropica hybrids. We divided the genome into segments of 20kb and counted the number of recombinations in each segment. We calculated the probability of observing a recombination in each block by dividing the recombination counts with total hybrids employed in our analysis. Since a centimorgan (cM) is 1% probability of finding a recombination, the calculated probabilities were converted to cM distances by multiplying by 100.
|
10.1371/journal.ppat.1001221 | CD4+ Natural Regulatory T Cells Prevent Experimental Cerebral Malaria via CTLA-4 When Expanded In Vivo | Studies in malaria patients indicate that higher frequencies of peripheral blood CD4+ Foxp3+ CD25+ regulatory T (Treg) cells correlate with increased blood parasitemia. This observation implies that Treg cells impair pathogen clearance and thus may be detrimental to the host during infection. In C57BL/6 mice infected with Plasmodium berghei ANKA, depletion of Foxp3+ cells did not improve parasite control or disease outcome. In contrast, elevating frequencies of natural Treg cells in vivo using IL-2/anti-IL-2 complexes resulted in complete protection against severe disease. This protection was entirely dependent upon Foxp3+ cells and resulted in lower parasite biomass, impaired antigen-specific CD4+ T and CD8+ T cell responses that would normally promote parasite tissue sequestration in this model, and reduced recruitment of conventional T cells to the brain. Furthermore, Foxp3+ cell-mediated protection was dependent upon CTLA-4 but not IL-10. These data show that T cell-mediated parasite tissue sequestration can be reduced by regulatory T cells in a mouse model of malaria, thereby limiting malaria-induced immune pathology.
| Severe malaria can kill people via complications such as cerebral malaria. The number of malaria parasites in the body is a major determinant of whether a patient will develop severe disease. T cells are thought to help control parasite numbers, but regulatory T cells, which are known to dampen immune responses, are present at a greater frequency in the blood of malaria patients with the highest parasitemia, suggesting that these cells might impair parasite control. Our experiments in a mouse model of cerebral malaria show for the first time that regulatory T cells can contribute to protection against disease. Specifically, our data shows that accumulation of parasites in host tissues can be promoted by anti-parasitic T cell responses, and that regulatory T cells can reduce this parasite tissue sequestration and protect against experimental cerebral malaria if their numbers are sufficiently elevated. These results suggest that regulatory T cells can help reduce pathogenic T cell responses during experimental infection and protect against malaria induced immune pathology.
| Severe malaria syndromes, including cerebral malaria (CM), claim the lives of approximately 900,000 people annually, mostly children under the age of 5 living in sub-Saharan Africa [1]. The mechanisms of CM pathogenesis remain poorly understood, since studies in humans are often restricted to post-mortem examinations. In particular, the roles played by the host immune response in either driving or preventing CM are unclear. It is possible that the immune response could be over-exuberant in some CM patients or lethargic in others, the balance of which may depend on the patient's and the parasite's genetic background. Several studies in malaria patients have reported associations between higher frequencies of peripheral blood regulatory T (Treg) cells and increased parasitemia [2], [3], [4]. However, these studies provided limited mechanistic insight into the role of Treg cells in severe malarial disease.
Under homeostatic conditions, Treg cells limit potentially aberrant T cell responses, thus preventing autoimmunity [5]. However, they can also impair effective pathogen clearance [6], [7], [8], while potentially playing a beneficial role in preventing immune-pathology during infection. The molecular mechanisms by which Treg cells perform these functions are incompletely understood, but have been reported to involve production of cytokines such as TGFβ and IL-10, and increased expression of the negative regulatory molecule CTLA-4 [9], [10], [11]. Furthermore, it is not known whether Treg cells act directly upon conventional T cells or on accessory cells such as antigen-presenting cells. Nevertheless, Treg cells are often viewed as detrimental during infection, since they may impede the generation of effective pathogen-specific T cell responses. Thus, an emerging paradigm is that Treg cells block T cell-mediated clearance of malaria parasites in humans, facilitating an increase in parasitemia.
The direct study of immune mechanisms in malaria patients is problematic for obvious practical and ethical reasons. Therefore, mouse models of severe and non-severe malaria have been employed to study the immune response to infection. Studies in an experimental model of cerebral malaria (ECM) caused by infection of C57BL/6 mice with P. berghei ANKA (PbA) have highlighted the important role played by various immune cells in disease pathogenesis, including CD4+ T cells, CD8+ T cells, conventional dendritic cells and Natural Killer (NK) cells [12], [13], [14], [15], [16], [17], [18]. In mice that succumb to ECM, parasite biomass is poorly controlled and there is clear evidence of immune-mediated parasite tissue sequestration [19]. Until recently, the deleterious role proposed for Treg cells in studies of human malaria has been difficult to test in mice, due to the lack of appropriate reagents [20], [21]. Our initial studies indicated a detrimental role for Treg cells because depletion of CD25hi cells prior to infection, the majority of which were Treg cells, protected mice from ECM and was associated with increased antigen-specific CD4+ T cell responses [20]. Recently however, specific depletion of FoxP3+ Treg cells did not protect against ECM, bringing into question the role for these cells in mediating disease [22]. Although the effect of Treg cell depletion on T cell responses and pathogen burden was not studied [22], given that ECM is mediated by pathogenic T cells that promote parasite tissue sequestration [19], we hypothesized that under certain conditions, Treg cells can suppress deleterious T cell responses and protect against ECM.
One approach to manipulate Treg cell numbers in vivo has been to use IL-2/anti-IL-2 antibody complexes to potentiate IL-2 signalling and drive expansion of FoxP3+ Treg cells [23]. Certain monoclonal antibodies (mAbs) bind to IL-2 in such a way that its signalling capacity is preserved, while its in vivo half-life is dramatically extended [24]. Moreover, different mAbs against IL-2 bind to different regions of the molecule, thus skewing its signalling capacity [25]. For example in mice, IL-2 bound to S4B6 mAb is not capable of interacting with the high affinity, heterotrimeric IL-2 receptor, but does interact with the lower affinity heterodimeric receptor. In contrast, IL-2 bound to JES6-1A12 mAb retains the ability to interact with the higher affinity receptor. IL-2/anti-IL-2 complexes profoundly alter lymphocyte dynamics during homeostasis, autoimmunity and vaccination [23], [25], [26], [27], [28]. Recently, IL-2/JES6-1A12 was shown to expand Treg populations, prevent auto-immunity and induce long term graft tolerance [23]. Here, we show for the first time that while removal of naturally-occurring Treg cells minimally affects the course of disease, increasing their numbers in vivo throughout the course of infection via IL-2/anti-IL-2 antibody complexes allows these cells to protect against ECM.
The foxp3-DTR transgenic (DEREG) mouse was recently used to deplete Foxp3+ cells prior to and over the course of PbA infection, with no impact on susceptibility to ECM [22]. We employed the same system here to study the effect of Foxp3+ cell depletion on T cell responses and pathogen burden during ECM. Consistent with the published data, in our hands DEREG mice depleted of Foxp3+ cells the day prior to, and over the course of infection (Figure 1A), remained as susceptible to ECM as Foxp3+ cell replete DEREG mice (data not shown). Furthermore, we observed no change in whole body parasite burden (Figure 1B); with a trend towards an increase in the splenic IFNγ+ CD4+ T cell response (Figure 1C). These data show that the depletion of Foxp3+ cells had little effect on pathogen burden or disease outcome during ECM.
Since removal of Foxp3+ cells had no effect on disease progression, we next examined whether increasing numbers of Foxp3+ cells would impact upon ECM development. Therefore C57BL/6 mice were infected with PbA and immediately treated with a single dose of IL-2/JES6-1A12 (hereafter referred to as IL-2Jc) or IL-2/S4B6 (IL-2Sc) complexes. The IL-2Jc complex binds the high affinity heterotrimeric IL-2 receptor to drive Treg cell expansion, while the IL-2Sc complex binds the lower affinity heterodimeric IL-2 receptor resulting in the expansion of activated CD8+ T cells and NK cells [25]. Control mice that received rat IgG displayed clinical signs of illness from day 6 post-infection (p.i), and succumbed to infection with neurological symptoms typical of ECM with a Median Survival Time (MST) of 8 days (Figure 2A). Mice treated with IL-2Sc were also susceptible to ECM (MST: 7 days), demonstrating that IL-2Sc afforded no protection against infection (Figure 2A). In stark contrast, IL-2Jc treated, infected mice, rarely displayed ECM symptoms and were protected from ECM-related morbidity, dying instead from hyperparasitemia with an MST of 28 days (Figure 2A). Mice treated with S4B6 alone, JES6-1A12 alone or recombinant IL-2 alone, were as susceptible to ECM as control mice (Figures 2B & 2C). Together, these data demonstrate a specific capacity for IL-2Jc, but not its component parts in isolation or an alternative IL-2 antibody complex (IL-2Sc), to protect against ECM.
To investigate the timing and dosing requirements for IL-2Jc-mediated protection, PbA-infected mice were treated on days 0 or 2 p.i., or on both days with either a standard IL-2Jc dose (1.5ug cytokine: 50ug antibody) (Figure 2D left) or a ten-fold lower dose (Figure 2D right). Control mice displayed clinical signs of disease from day 6 p.i., with 100% of mice succumbing to ECM by day 8 p.i. Only mice that had received a standard IL-2Jc dose on day 0 were protected from ECM. Importantly, mice receiving a delayed IL-2Jc dose on day 2 p.i. were completely susceptible to ECM. Thus, IL-2Jc protects against ECM only when administered at the time of infection.
C57BL/6 mice typically display ECM symptoms when blood parasitemia reaches ∼7–10% parasitized red blood cells (pRBCs) (Figure 3A). Blood parasitemia in IL-2Jc-treated mice was similar to control, infected mice on days 4 & 5 p.i. (Figure 3A). However, from day 6 p.i. onwards, when clinical symptoms appeared in control mice, IL-2Jc treated mice displayed significantly lower blood parasitemia for the following 3 days, only rising again from day 10 p.i. onwards (Figure 3A). While blood parasitemia has been routinely used to monitor disease progression, it is now recognised that measurements of total parasite biomass in the whole body offer a better correlate of the disease status of malaria patients [29]. To assess parasite biomass in infected mice, we used a transgenic PbA strain engineered to constitutively express firefly luciferase (PbA-luc) [20]. The bioluminescence generated by PbA-luc parasites at any given time is directly proportional to the sum of parasites in the tissues and circulating blood of the infected animal [19], [20], [30] (Figure 3B). IL-2Jc-treated, PbA-luc-infected mice harboured significantly lower parasite biomass compared to control mice on day 6 p.i., when control animals displayed severe ECM symptoms. Moreover, following whole body perfusion to remove circulating RBCs, brains from IL-2Jc-treated mice also exhibited significantly lower parasite sequestration than brains from control animals (p<0.05) (Figure 3C). These data demonstrate that IL-2Jc-mediated protection against ECM was associated with lower parasite biomass and reduced pRBC brain sequestration.
ECM is associated with the recruitment of CXCR3+ leukocytes to the brain [13], [31], [32], [33], [34], including T cells responsible for disease pathology [16], [18]. On day 6 p.i., the recruitment of CD8+ and CD4+ T cells, but not NK cells, to the brain was significantly reduced by IL-2Jc treatment, compared with mice receiving either IL-2Sc or control treatment (Figure 4). Furthermore, Treg cell numbers were significantly higher in IL-2Jc-treated mice compared to all other groups studied (Figure 4). These data indicated that IL-2Jc-mediated protection was associated both with a specific blockade of conventional T cell recruitment to the brain, and also an increase in the number of Treg cells in this tissue site.
CD8+ T cells play a key role in ECM pathology [16], [18]. A previous study using a transgenic PbA strain expressing model T cell epitopes showed that antigen-specific CD8+ T cells are primed in the spleen [35]. We employed this experimental system to assess the fate of antigen-specific CD8+ T cells in IL-2Jc-treated mice during ECM. OVA-specific, congenic (CD45.1) CD8+ T (OTI) cells were transferred into mice prior to infection with OVA-transgenic PbA (PbTG) or a non-OVA-expressing control parasite (PbG). On day 6 p.i., splenic OTI cell numbers and activation status, via Granzyme B (GzmB) expression, were assessed (Figure 5A). OTI cells were not detected in the spleens of naïve mice or mice infected with PbG (Figure 5A). The expression of GzmB was detected in ∼30% of endogenous (CD45.1 negative) CD8+ T cells in PbG-infected mice, indicating dramatic activation of CD8+ T cells at the onset of ECM. In control treated mice infected with PbTG, OTI cells were readily detected, indicating antigen-specific activation and proliferation of these cells had occurred. Moreover, nearly all of these cells expressed GzmB, at a level similar to activated endogenous CD8+ T cells. IL-2Jc treatment at the time of infection dramatically impaired, though did not abrogate, the OTI CD8+ T cell response (Figure 5A). This effect was not apparent in mice treated on day 2 p.i. with IL-2Jc. Furthermore, mice treated with IL-2Sc displayed a trend towards an enhanced OTI T cell response compared to control mice, which is consistent with reports of the stimulatory effect of IL-2Sc on CD8+ T cells [25], [26], [27], [28], [36]. Together these data demonstrate that IL-2Jc, when administered on the day of infection, potently inhibits pathogenic, antigen-specific CD8+ T cell responses during ECM.
To be sure that IL-2Jc did not stimulate NK cells or NKT cells, we examined their expression of the activation markers CD69, GzmB and IFNγ, 24 hours after infection and treatment with IL-2Jc. As expected, no further activation of NK or NKT cells in IL-2Jc-treated mice was detected relative to control-treated, infected mice; while in contrast, IL-2Sc-treatment clearly stimulated both NK and NKT cells (Figure S1A). Furthermore, neither depletion of NK cells with anti-NK1.1 antibody, nor the absence of invariant chain NKT cells in B6.Jα18−/− mice, impeded either CD4+ Foxp3+ T cell expansion (Figure S1B) or control of parasite burden (Figure S1C) by IL-2Jc. Together these data indicate that IL-2Jc does not protect against ECM via activation of NK cells or NKT cells.
PbA infection induces a potent pro-inflammatory cytokine response in C57BL/6 mice that is strongly associated with ECM pathogenesis. IFNγ is absolutely critical for disease onset [37], [38], [39], possibly by promoting PbA tissue sequestration [19]. We found that IL-2Jc administered on the day of infection resulted in lower serum IFNγ levels by day 4 p.i., whereas neither IL-2Sc nor delayed IL-2Jc treatment had any significant effect (Figure S2). Examination of the antigen-specific splenic CD4+ T cell response indicated an impaired ex vivo proliferative and IFNγ recall response from IL-2Jc treated mice (Figure S3), suggesting that in vivo CD4+ T cell responses were impaired by IL-2Jc treatment. Therefore, we enumerated splenic IFNγ-producing CD4+ T cells and Treg cells over the course of PbA-infection in mice treated with IL-2Jc or IL-2Sc. IFNγ-producing CD4+ T cells were detectable from day 4 p.i. onwards in infected, but not naïve mice (Figure 5B). Control saline-treated and IL-2Sc-treated infected mice had very similar numbers of IFNγ+ CD4+ T cells on day 4 p.i.. In contrast, IL-2Jc treatment suppressed the number of IFNγ+ CD4+ T cells (p<0.001). Interestingly, the number of IFNγ+ CD4+ T cells was greatly enhanced if IL-2Jc was administered on day 2 p.i. (p<0.001)(Figure 5B), possibly indicating increased expression of high affinity IL-2 receptor on these cells by day 2 p.i.. Treg cells expanded in control, infected mice and numbers peaked on day 4 p.i., before declining by day 6 p.i. (Figure 5C), consistent with previous reports [20], [21], [40]. Infected mice treated with IL-2Sc exhibited almost identical Treg cell expansion kinetics to that of control treated mice (Figure 5C), consistent with the notion that IL-2Sc has little impact on Treg cell numbers [25]. Strikingly, IL-2Jc treatment at the time of PbA infection resulted in a dramatic expansion in Treg cell numbers with a >6-fold increase by day 2 p.i., peaking at day 4 p.i. (∼3.5-fold greater numbers than in control treated mice), before retracting somewhat by day 6 p.i., although numbers still remained ∼4-fold greater than in IL-2Sc and control treated groups. In contrast, delaying IL-2Jc treatment until day 2 p.i., resulted in very little enhanced Treg cell expansion. Thus, the protection afforded by IL-2Jc treatment at the time of PbA infection was associated with a dramatic and sustained elevation of Treg cell numbers over the course of infection. Taken together, these data show that IL-2Jc-mediated protection against ECM was associated with an expansion of CD4+ Treg cells and an accompanying impairment of the conventional CD8+ and CD4+ T cell responses.
We next determined whether the increase in Treg cell numbers caused by IL-2Jc treatment during infection was the result of natural Treg cell expansion, or de novo conversion of naïve, Foxp3− CD4+ T cells into Treg cells. Foxp3+ CD4+ natural Treg cells and Foxp3− CD4+ T cells were sorted from the spleens of naïve foxp3gfp/gfp mice [41], and transferred into C57BL/6 mice. On the same day, these mice were infected, and treated either with IL-2Jc or saline. At the peak of Treg cell expansion (4 days later), splenic GFP+ Foxp3+ CD4+ Treg cells were enumerated. These cells were readily detected in mice that received GFP+ natural Treg cells, and indeed their numbers were boosted by IL-2Jc treatment (Figure 6A). However, in mice receiving GFP− non-Treg CD4+ T cells, we observed no evidence of their conversion into Foxp3+ Treg cells either spontaneously during infection or after stimulation with IL-2Jc. These data indicate that IL-2Jc treatment during ECM triggers natural Treg cell expansion, but not conversion of Foxp3− CD4+ T cells to a Foxp3+ phenotype.
We next examined whether natural Treg cell expansion caused by IL-2Jc was dependent upon infection. Naïve and infected groups of mice were treated with IL-2Jc or control saline. Four days later, the number of splenic Foxp3+ CD4+ T cells was determined (Figure 6B). Consistent with previous reports [23], [25], substantial Treg cell expansion was observed in both naïve and infected mice, demonstrating that this phenomenon is not dependent on infection. Importantly, however, when we assessed direct ex vivo production of cytokines by Treg cells (by intracellular cytokine staining with no in vitro stimulation) (Figure 6B), we noted that while expanded Treg cells in naïve mice made little IL-10, those in IL-2Jc treated, infected mice, made significantly higher amounts of this cytokine than those from control, infected mice. A small number of Treg cells from IL-2Jc-treated, infected mice, also appeared to make IFNγ, but this response was much lower than the IL-10 response (Figure 6B). These data indicate that IL-2Jc triggers the expansion of IL-10-producing Treg cells during PbA infection. We further analysed the effects of IL-2Jc on Treg cells, and observed that their expression of CD25, Foxp3 and CTLA-4 was substantially elevated by IL-2Jc treatment compared to control saline treated, infected mice (Figure 6C). Taken together, these data demonstrate that IL-2Jc treatment triggers the expansion of natural CD4+ Treg cells, which then express higher levels of Foxp3, IL-10 and CTLA-4 in response to PbA infection.
To determine if Treg cells were important for IL-2Jc mediated protection against ECM, we employed the DEREG mice [5]. C57BL/6 mice and DEREG mice were infected with PbA, and treated with IL-2Jc or saline. DT or saline was administered to IL-2Jc treated DEREG and C57BL/6 mice from day 3 p.i., around the peak expansion of Treg cells. The following day (day 4 p.i.), while Treg cell expansion was evident in DEREG mice given IL-2Jc, depletion of Treg cells (>95% efficacy in this study) was confirmed in mice that had received DT (Figure 7A). C57BL/6 mice were completely protected from ECM when treated with IL-2Jc, either with or without DT treatment (Figure 7B), indicating no side-effects of DT treatment in C57BL/6 mice during PbA-infection over this time-frame. DEREG mice were equally susceptible to ECM as C57BL/6 mice, and were protected by IL-2Jc treatment (Figure 7B). Crucially, IL-2Jc-mediated control of parasite burdens and protection from disease was completely abrogated when DEREG mice were treated with DT (Figure 7B & 7C). These data formally demonstrate that Foxp3+ cells are responsible for IL-2Jc-mediated protection against ECM.
Since IL-2Jc treatment increased IL-10 and CTLA-4 expression by Foxp3+ CD4+ T cells during infection (Figure 6B & 6C), we hypothesized that protection was dependent upon these two molecules. To test this, IL-2Jc-treated, PbA-infected C57BL/6 mice received anti-CTLA-4 or anti-IL-10R blocking antibodies, or control IgG from day 3 p.i.. Anti-CTLA-4 significantly reduced IL-2Jc-mediated protection, with >60% of IL-2Jc-treated mice succumbing to infection with pathogen burdens similar to control infected mice (Figure 8A & 8B). Anti-IL-10R blockade, on the other hand, only partially reversed IL-2Jc-mediated protection, with >60% survival (Figure 8A), and interestingly, further reduced pathogen burdens in IL-2Jc treated mice (Figure 8B). Both antibody blockade treatments restored the splenic IFNγ CD4+ T cell response that had been impaired by IL-2Jc treatment (Figure 8C). Since we could detect only a modest role for IL-10 in IL-2Jc mediated protection of wild-type C57BL/6 mice, we further examined the effect of IL-2Jc treatment in IL-10−/− mice, and found that these animals were significantly protected against ECM in a CTLA-4-dependent manner (Figure S4). These data demonstrate that IL-10 is not essential for IL-2Jc-expanded natural Treg cells to protect against ECM.
In addition, non-IL-2Jc treated mice were also treated with anti-CTLA-4 or anti-IL-10R blocking antibodies to study their effects alone on the course of ECM. Anti-CTLA-4 treatment did not alter disease outcome in saline treated mice (Figure 8A), despite partially reducing parasite burdens (Figure 8B), while anti-IL-10R treatment significantly accelerated ECM onset (Figure 8A) (MST 6.5 days vs. 8 days; p<0.01), with a partial reduction in parasite burden (Figure 8B). Of note, when anti-IL-10R mAb was administered from the start of infection, a more substantial reduction on parasite burden is observed [19].
Taken together, these data demonstrate that when CTLA-4 was blocked in IL-2Jc treated mice, pathogenic CD4+ T cell responses were restored, pathogen burdens were poorly controlled and protection from ECM was reversed. Thus IL-2Jc mediated protection against ECM is strongly dependent on Foxp3+ cells, and CTLA-4, but not IL-10.
T cell responses to infection are generally required for pathogen control, but can also contribute to disease. The roles of T cells in the pathogenesis of severe malaria syndromes, including CM, are unclear. Leukocytes have been observed in the brains of patients who have died from CM [42], [43], but their contribution to CM pathogenesis is not known. Studies in the ECM model show that T cells play a critical role in disease pathogenesis [16], [18], although the role of Treg cells in this model remains the subject of debate [20], [21], [22]. Data from this study and others [22] suggest that Treg cells do little to impact on ECM onset, and may in some cases exacerbate disease [20], [21]. Higher Treg cell frequencies have been associated with elevated blood parasitemia in human malaria patients [4], [44], [45], suggesting that Treg cells impair pathogen clearance during malaria. However, one report demonstrated that anti-CD25 treatment of ECM-resistant BALB/c mice increased the incidence of neurological symptoms during secondary PbA challenge [46], suggesting Treg cells might be protective against ECM. Here, we report that Treg cells can protect against ECM following their expansion in vivo. Thus, while Treg cell responses during ECM are usually insufficient to control pathogenic T cells, and Treg cell ablation has no effect on pathogen burden or disease outcome, if present in large enough numbers, Treg cells can prevent disease. This is the first report to clearly show that CD4+ Foxp3+ Treg cells can play a direct protective role during experimental malaria infection. However, it is important to bear in mind that Treg-mediated protection was only achieved by treatment with IL-2Jc from the start of infection, and this therapeutic opportunity will not exist in human malaria patients. Thus, alternative approaches to rapidly expand Treg cell numbers in a clinical setting would have to be considered for therapeutic effect.
Previous data from our laboratory showed that anti-CD25 (PC61) monoclonal antibody treatment, partially depleted/blocked Treg cells (i.e.,, affecting only those cells expressing high levels of CD25), enhanced anti-parasitic CD4+ T cell responses, and reduced both parasite burden and ECM incidence [20]. These data were interpreted to mean that natural Treg cells normally impair pathogen clearance, and thus help to promote ECM. However, it is now clear from this work, and from another recent report, that total depletion of natural Treg cells does not protect against ECM [22]. The discrepancy between the outcome of partial and total Treg cell depletion in ECM are unresolved at present, but could be linked to the fact that anti-CD25 mAb-treated mice retain a population of CD25lo Foxp3+ CD4+ T cells that can display plasticity in vivo [47], and might therefore contribute to protection from disease.
Pathogen control is clearly inefficient during ECM, with little evidence that T cells provide any protection against infection. To date, only NK cells have been reported to mediate some pathogen control during ECM [13]. However, NK cell depletion in IL-2Jc-treated mice did not prevent protection from ECM, indicating that these cells were not targets for IL-2Jc and did not contribute to enhanced parasite control. We recently showed that T cells promote the accumulation of pRBC in multiple tissue sites during PbA infection, and that depletion of either CD4+ or CD8+ T cells to protect from ECM dramatically reduced parasite tissue sequestration [19]. Furthermore, lymphocyte-deficient B6.RAG1-deficient mice failed to develop ECM and had markedly reduced parasite burdens compared to control animals and B cell-deficient mice following PbA infection [19]. Therefore, in C57BL/6 mice, PbA tissue sequestration is promoted by host T cell responses, possibly by the conditioning host tissue endothelial cells via cytokines to allow binding of pRBC, as described by others [48].
Our earlier studies on anti-CD25 mAb treatment of PbA-infected mice indicated that early blockade/depletion of CD25hi Treg cells allowed the generation of an enhanced anti-parasitic CD4+ T cell response that was accompanied by recovery and expansion of CD25hi Treg cells during the course of infection [20]. Furthermore, our current data indicates that the most likely explanation for the protective effects of in vivo expanded Treg cells is the suppression of pathogenic T cell expansion that would otherwise promote parasite tissue sequestration. Thus, we propose a model whereby naturally occurring CD25hi Treg cells suppress the development of potent anti-parasitic immunity early during PbA infection and their depletion/blockade results in enhanced anti-parasitic CD4+ T cells responses, reducing both parasite burdens and the risk of severe pathology. However, there also appears to be an important role for the remaining CD25lo Foxp3+ CD4+ T cells in achieving a balance between emerging anti-parasitic immunity and immune-pathology in anti-CD25mAb-treated mice, as indicated by the failure of DT-mediated Treg cell depletion in DEREG mice to protect against ECM. A role for IL-10-producing inducible regulatory T cells [40] and/or IL-10/IFNγ-producing Th1 cells [4] identified in P. yoelii-infected mice and malaria patients, respectively, may also contribute to this latter process. Our data reported in this study shows that if Treg cells can be expanded sufficiently via IL-2Jc during PbA infection, they can suppress normally pathogenic T cell responses, and prevent parasite tissue sequestration and ECM. Hence, Treg cells could have two potentially important roles during PbA infection (Figure 9). First, they could suppress the early generation of anti-parasitic CD4+ T cells responses that are detrimental to the host, and second, they can modulate pathogenic T cell responses to reduce parasite tissue sequestration later during infection. This latter effect may allow better clearance of parasites in the spleen, thus lowering parasite burden. Again, we do not rule out the possibility that inducible regulatory T cells may also play a role in protecting against disease later in infection, and even restricting parasite tissue sequestration.
Treg cell-mediated protection against ECM was exquisitely sensitive to the timing of IL-2Jc treatment. Delaying treatment by 48 hours caused the selective expansion of conventional CD4+ T cells rather than Treg cells, presumably due to infection-induced expression of high affinity IL-2 receptor by the former cell population. However, this expansion of conventional CD4+ T cells was unable to control parasite growth and failed to protect from ECM. Treg cell depletion from day 3 p.i., completely abrogated IL-2Jc-mediated protection, showing that IL-2Jc-expanded Treg cells were responsible for protection from ECM. Clearly, there is a fine balance between the activation and expansion of anti-parasitic T cell responses and the emergence of disease-protective Treg cells that determines the outcome of PbA infection. Whether such an intimate relationship between these two different types of T cells exists during human malaria remains to be determined. Nevertheless, our data suggests a more complex temporal and spatial relationship between emerging anti-parasitic T cell responses required for control of parasite growth and the functions of Treg cells that may protect against disease, than has previously been recognised.
Treg cells function via multiple mechanisms, including CTLA-4 [11], IL-10, TGFβ and IL-2-deprivation [10]. We found that Treg cell-mediated protection against ECM required CTLA-4, but was only modestly affected by IL-10 blockade. Previous reports have demonstrated that a murine AIDS infection induces IL-10 expressing Treg cells, and that enhanced IL-10 levels protect against ECM [49], [50]. Our data is consistent with a moderate therapeutic role for IL-10 in ECM, but demonstrates that CTLA-4 is a more potent regulator of pathogenic T cell responses. We attempted to determine the antigen specificity of the IL-10 response made by Tregs cells in IL-2Jc treated mice, by sorting these cells, stimulating with APC and parasite antigen, and looking for IL-10 production at both the protein and mRNA level. However, these experiments are notoriously difficult to perform [51], [52], [53], and we were unable to assess the antigen specificity of IL-10 producing Tregs. There may also be different roles for the major regulatory molecules produced by Treg cells in our study (CTLA-4) and inducible regulatory T cells identified by others (IL-10; [4], [40]), by acting on different cellular/tissue targets during malaria. However, there is no direct evidence for this as yet.
In conclusion, we and others have shown that Treg cells numbers expand during malaria infection, but are unable to protect against T cell mediated immune pathology [20], [21], [40]. Here we show for the first time that Treg cells can protect against T cell-mediated immune pathology in malaria if their numbers are sufficiently expanded at the appropriate time during the immune response. Thus, while increased Treg cell frequencies may contribute to increased parasitemia in malaria patients [4], [44], [45], a further possibility is that these cells expand in an attempt to protect against disease caused by parasite sequestration.
Female C57BL/6 mice and congenic CD45.1+ C57BL/6 mice aged 6–8 weeks were purchased from the Australian Resource Centre (Canning Vale, Perth, Western Australia) and maintained under conventional conditions. DEREG mice [5], OTI [54], C57BL/6 il10−/−, and C57BL/6 Jalpha18−/− mice were bred and maintained in house. foxp3gfp/gfp mice [41] were backcrossed ten times onto the C57BL/6 background, bred and maintained in house.
All animal procedures were approved and monitored by the Queensland Institute of Medical Research Animal Ethics Committee. This work was conducted under QIMR animal ethics approval number A02-633M, in accordance with the “Australian code of practice for the care and use of animals for scientific purposes” (Australian National Health & Medical Research Council).
P. berghei ANKA (PbA) strains were used in all experiments after one in vivo passage in mice. A transgenic PbA (231c1l) clonal line expressing luciferase and green fluorescent protein under the control of the EF1-α promoter (PbA-luc) was used for all experiments unless stated otherwise [20]. Transgenic PbA strains expressing model T cell epitopes, and control strains, PbTG and PbG, were obtained from Prof. William R. Heath, University of Melbourne, Australia, and were maintained and used as previously reported [35]. All mice were infected with 105 pRBCs intravenously (i.v.) via the lateral tail vein. Blood parasitemia was monitored by examination of Diff-Quick (Lab Aids, Narrabeen, NSW, Australia) stained thin blood smears obtained from tail bleeds.
Mice were monitored twice daily after day 5 p.i., and clinical ECM evaluated. Clinical ECM scores were defined by the presentation of the following signs: ruffled fur, hunching, wobbly gait, limb paralysis, convulsions, and coma. Each sign was given a score of 1. Animals with severe ECM (accumulative scores = 4) were sacrificed by CO2 asphyxiation according to ethics guidelines, and the following timepoint given a score of 5 to denote death.
Allophycocyanin (APC) or Pacific Blue (PB)-conjugated anti-TCRβ chain, phycoerythrin(PE)-Cy5- or PE-conjugated anti-CD4, PE-Cy5-conjugated anti-CD8, PE or fluorescein isothiocyanate-conjugated anti-CD45.1, APC or PE-conjugated anti-IFNγ, and PE-conjugated anti-IL-10 were purchased from Biolegend (San Diego,CA) or BD Biosciences (Franklin Lakes, NJ). Alexa-647-labelled anti-mouse Foxp3 mAb was purchased from eBioscience (San Diego, CA). PE-conjugated anti-human Granzyme B (GzmB), with mouse cross reactivity, was purchased from Invitrogen (Mount Waverley, Vic., Australia). Anti-CTLA-4 (UC10-4F10-11) and control IgG was purchased from BioXCell, (West Lebanon, NH, USA). Anti-IL-10R (1B1.3a), anti-CD4 (YTS191), anti-IL-2 (S4B6 and JES6-1A12), anti-NK1.1 (PK136), and isotype control mAb (MAC49; ratIgG1) were purified from culture supernatants by protein G column purification (Amersham, Uppsala, Sweden) followed by endotoxin removal (Mustang Membranes; PallLife Sciences, East Hills, NY). Purified control rat IgG were also used in some experiments and purchased from Sigma-Aldrich (Castle Hill, NSW, Australia). Diphtheria toxin (DT) was purchased from Sigma-Aldrich, diluted in saline, and 1µg doses injected via the intraperitoneal route.
1.5µg of recombinant murine IL-2 (eBioscience, San Diego, CA) was incubated with 50µg of either S4B6 or JES6-1A12 (prepared as detailed above) in saline, for 30 minutes at 37°C prior to intraperitoneal administration to each mouse in a volume of 200µl.
Blood mononuclear cells were analysed in heparinised blood after 2 rounds of red cell lysis using hypotonic red cell lysis buffer according to the manufacturer's instructions (Sigma-Aldrich). Spleen cells were isolated by passing tissue through a 100-µm sieve in RPMI-1640 tissue culture medium supplemented with 2% (v/v) fetal calf serum (Wash Buffer). Red blood cells were lysed as above (Sigma-Aldrich) and washed once more with Wash Buffer. Brain mononuclear cells were isolated by digesting tissue in collagenase type 4 (1 mg/ml; Worthington Biochemical Corp., Lakewood, NJ) and deoxyribonuclease I (0.5 mg/ml; Worthington Biochemical) at room temperature for 40 minutes, before passing through a 100-µm sieve and washing twice with Wash Buffer. The cell pellet was resuspended in 33% (v/v) Percoll in PBS and centrifuged at 693×g for 12 minutes at room temperature. Supernatant containing debris was removed, and the leukocyte pellet was washed once in Wash Buffer, red blood cells lysed as described above, and washed and resuspended in RPMI-1640 medium supplemented with 5% (v/v) fetal calf serum.
For the staining of cell surface antigens, cells were incubated with fluorochrome-conjugated mAbs on ice for 20 minutes. Intracellular staining for Foxp3 was performed on fixed/permeabilized cells using Alexa647-labeled anti-mouse Foxp3 kit (eBioscience), according to the manufacturer's instructions. Intracellular cytokine staining for IFNγ, CTLA-4 and GzmB was performed using a BD Fixation/Permeabilisation kit (BD Biosciences) according to manufacturer's instructions. Data were acquired on a FACSCanto II flow cytometer (BD Biosciences) and analysed using FlowJo software (Treestar, Ashland, OR, USA). Cell populations in the blood, spleen and brain were defined as follows: CD4+ T cells (CD4+TCRβ+), CD8+T cells (CD8α+TCRβ+), NK cells (NK1.1+TCRβ−), CD4+ Treg cells (CD4+Foxp3+TCRβ+). Cytokines in tissue culture supernatants and serum samples were quantified using the cytometric bead array flexsets (BD Biosciences) on a FACSarray equipped with BD Flexset analysis software (BD Biosciences).
Splenic CD4+ T cells (5×104 cells/well), either bulk populations purified to >85% purity by magnetic bead positive selection techniques (Miltenyi Biotec; North Ryde, NSW, Australia), or cell sorted to isolate Treg cells from foxp3gfp/gfp mice to a purity >99%, were stimulated with 2.5×105 PbA-parasitized RBC (pRBC) or naive RBC (nRBC), and 1×106 irradiated naive C57BL/6 spleen cells at 37°C in 5% (v/v) CO2. Cell culture supernatant was collected after 24h or 72 h and cytokines were measured as above (BD Biosciences). After 72 h of culture, cells were pulsed with 1 µCi [3H]thymidine for 18 h, before measuring thymidine incorporation using a Betaplate Reader (Wallac).
Luciferase-expressing PbA pRBCs were visualized by imaging whole bodies or dissected organs with an I-CCD photon-counting video camera and in vivo imaging system (IVIS 100; Xenogen, Alameda, CA). Mice were anesthetized with isofluorane and injected intraperitoneally with 0.1 ml of 5 mg/ml D-luciferin firefly potassium salt (Xenogen). 5 minutes afterwards, images were captured on the IVIS 100 according to the manufacturer's instructions. Parasites were visualized in the brain after removal from mice that had been perfused with 15ml of saline via the heart. Bioluminescence generated by luciferase transgenic PbA in mice or brain tissue was measured according to the manufacturer's instructions. The unit of measurement was photons/second/cm2/steer radiant (p/sec/cm2/sr).
Differences in survival of treatment groups were analysed using the Kaplan-Meier log-rank test. All other analyses of differences in parasitemia, cytokine levels, cell numbers, bioluminescence etc. were performed using the Mann-Whitney nonparametric test. For all statistical tests, p<0.05 was considered significant. In all figures, *, **, *** denote p values of p<0.05, p<0.01 & p<0.001 respectively.
|
10.1371/journal.pgen.1007458 | Reduced monocyte and macrophage TNFSF15/TL1A expression is associated with susceptibility to inflammatory bowel disease | Chronic inflammation in inflammatory bowel disease (IBD) results from a breakdown of intestinal immune homeostasis and compromise of the intestinal barrier. Genome-wide association studies have identified over 200 genetic loci associated with risk for IBD, but the functional mechanisms of most of these genetic variants remain unknown. Polymorphisms at the TNFSF15 locus, which encodes the TNF superfamily cytokine commonly known as TL1A, are associated with susceptibility to IBD in multiple ethnic groups. In a wide variety of murine models of inflammation including models of IBD, TNFSF15 promotes immunopathology by signaling through its receptor DR3. Such evidence has led to the hypothesis that expression of this lymphocyte costimulatory cytokine increases risk for IBD. In contrast, here we show that the IBD-risk haplotype at TNFSF15 is associated with decreased expression of the gene by peripheral blood monocytes in both healthy volunteers and IBD patients. This association persists under various stimulation conditions at both the RNA and protein levels and is maintained after macrophage differentiation. Utilizing a “recall-by-genotype” bioresource for allele-specific expression measurements in a functional fine-mapping assay, we localize the polymorphism controlling TNFSF15 expression to the regulatory region upstream of the gene. Through a T cell costimulation assay, we demonstrate that genetically regulated TNFSF15 has functional relevance. These findings indicate that genetically enhanced expression of TNFSF15 in specific cell types may confer protection against the development of IBD.
| Crohn’s disease and ulcerative colitis, characterized by gut inflammation, are the two main subtypes of Inflammatory bowel disease (IBD). Over two hundred genetic loci have been identified that contribute to risk for IBD. However, functional studies are required to determine the mechanisms by which these genetic changes might affect disease risk. To date, only a few IBD loci have been functionally characterized. We focused on the IBD risk locus at TNFSF15, encoding the cytokine also known as TL1A. Previous work has shown that TNFSF15 enhances lymphocyte activation and promotes inflammation in animal models of disease. However, we here call into question the assumption that TNFSF15 is always a pro-inflammatory molecule by demonstrating that the IBD risk allele at TNFSF15 is associated with decreased production of TNFSF15 by monocytes and macrophages at rest and after stimulation. In addition, we narrow the list of potential variants at TNFSF15 responsible for controlling its expression, and we demonstrate that genetically controlled TNFSF15 can have functional impact on responding T cells. These findings both demonstrate a mechanism by which this IBD risk locus might drive disease predisposition and suggest a new protective role for TNFSF15 in maintaining the intestinal barrier.
| Maintaining immune homeostasis in the microbiome-rich environment of the intestine is a complex process, mediated by numerous mechanisms including the physical epithelial barrier, mucin secretion, antimicrobial peptides, anti-inflammatory cytokines, regulatory cells, and IgA responses [1, 2]. Inflammatory bowel disease (IBD, including Crohn’s disease, CD, and ulcerative colitis, UC) results from a breakdown in mucosal immune homeostasis [3], but the precise mechanisms by which barrier dysfunction begins remain largely unknown. To date, genome-wide association studies (GWASs) have identified approximately 200 distinct susceptibility loci for IBD, the majority of which are associated with both CD and UC [4, 5]. A fine-mapping study used Bayesian statistical methodology to find the most probable causal variants underlying the association signal at each locus [6]. However, to fully realize the benefit of GWAS discoveries, functional studies are required to move beyond statistical associations with genetic loci and uncover the biological mechanisms behind genetic predisposition to disease. Functional studies of IBD-associated genetic variants have been performed for several loci, demonstrating that risk variants can lead to a breakdown in the intestinal barrier through both reducing (e.g NOD2 [7–10] and ATG16L1 [11–13]) and enhancing (e.g. IL23R [14–17]) gene function. Expression quantitative trait locus (eQTL) studies have found genes whose expression may be increased or decreased by IBD risk-associated variants and have highlighted the impact of cell type on the effects of a genetic variant [18–20]. In a complementary approach, investigation of gene expression patterns in monocytes and monocyte-derived macrophages revealed enrichment for genes near IBD loci among those upregulated during macrophage differentiation or stimulation [21], suggesting the importance of this cellular lineage in IBD. Investigating how IBD-associated variants influence disease susceptibility should provide insight into disease etiology and enable design of improved therapeutic or preventive strategies.
The genomic locus encoding the tumor necrosis factor superfamily member TNFSF15 (also known as TL1A) is associated with both CD and UC in populations of multiple ethnic backgrounds [4, 5, 22, 23]. TNFSF15 is produced by a variety of tissues, the most studied of which include myeloid lineage cells, activated T cells, and endothelial cells [24]. At the cellular level, TNFSF15 costimulates both T cells and innate lymphoid cells (ILC) and promotes differentiation of IL-9-producing T cells [25–31]. Increased TNFSF15 expression has been observed systemically and at the site of inflammation in IBD and is particularly associated with active disease [32–37]. Such observations have led to the hypothesis that the allele associated with risk for IBD at TNFSF15 might increase its expression and thereby promote inflammation. Several studies have found associations between genetic variants tagging the IBD-associated locus at TNFSF15 and TNFSF15 mRNA and protein expression [18, 19, 38–47]. However, the population studied, the disease status of the subjects, the cell type considered and the observed direction of effect differ between studies, leaving the mechanism by which this genetic locus confers susceptibility to IBD unclear.
To shed light on the functional consequences of the IBD susceptibility locus at TNFSF15 we examined the association of single nucleotide polymorphism (SNP) genotype with mRNA expression in specific immune cell types in multiple cohorts of healthy individuals and patients with inflammatory diseases. We found that the IBD-associated locus is an eQTL for monocyte TNFSF15. The genetic signals underlying the associations with TNFSF15 expression and IBD colocalize, suggesting that disease risk is mediated though regulation of gene expression. Importantly, we show that the IBD-protective allele at TNFSF15 is strongly associated with increased monocyte TNFSF15 mRNA in both healthy individuals and patients with inflammatory diseases. To further investigate the mechanism of this protective haplotype, we used a “recall-by-genotype” bioresource of healthy individuals from the United Kingdom. This analysis demonstrated that the IBD-protective allele is associated with increased TNFSF15 mRNA and protein expression under a variety of stimulation contexts, as well as monocyte costimulatory capacity in acute lymphocyte activation. Through allele-specific expression measurements in individuals with breaks in the associated haplotype block, we functionally fine-mapped the expression-associated locus to upstream of the gene. Importantly, we found that association of the protective allele with increased TNFSF15 expression was maintained in monocyte-derived macrophages, which play an important role in mucosal immunology. Thus, our findings suggest that genetically elevated TNFSF15 from monocytes or monocyte-derived cells may protect healthy individuals from the development of IBD.
We previously identified an association between genotype at the TNFSF15 IBD susceptibility locus and TNFSF15 mRNA expression in peripheral blood monocytes in both healthy individuals and newly diagnosed IBD patients from the UK [18, 19] (Fig 1A, S1 Fig). We also confirmed this association in a cohort of British patients with another immune-mediated disease, anti-neutrophil cytoplasmic antibody (ANCA)-associated vasculitis [18] (S1 Fig). The SNP most associated with TNFSF15 expression was rs6478109 [19], which is located 358 base pairs upstream of TNFSF15 and has a minor allele frequency (MAF) of 32.5% in individuals of European descent. An IBD GWAS trans-ancestry analysis by Liu et al indicated that the SNP most strongly associated with IBD in individuals of European descent at this locus is rs7848647 [4], which is 280 base pairs upstream of rs6478109 and in near complete linkage disequilibrium (LD) with it (r2 = 0.995 in 1000 Genomes Phase 3 European cohort). Comparison of the regional association plots for the TNFSF15 eQTL and IBD susceptibility revealed similar patterns of association (Fig 1A). In order to formally evaluate whether the eQTL and the disease association signal are driven by the same causal variant or two distinct variants in LD, we performed colocalization testing [48]. Our results indicated that the colocalization was highly likely (posterior probability of a shared underlying causal variant was 0.998).
rs6478109 genotype is also associated with TNFSF8 expression in whole blood [49] and monocytes [50], and additional monocyte TNFSF8 eQTLs have been found in this genomic region [18–20, 51]. However, rs6478109 is not the SNP most associated with TNFSF8 expression in any of these studies. Comparison of the patterns of genetic association across this locus with IBD and TNFSF8 expression in our previous monocyte eQTL data [19] demonstrated stark differences (S2 Fig), suggesting that TNFSF8 is unlikely to be the causal gene at this locus. Colocalization testing confirmed that the IBD association and TNFSF8 eQTL signals were very unlikely to be driven by a shared causal variant (posterior probability 0.092).
These analyses indicate that IBD susceptibility may be mediated by changes in TNFSF15 expression. We therefore performed further functional examination of the locus using a recall-by-genotype study design in an independent bioresource of healthy individuals recruited from the region around Cambridge, UK. We confirmed the association of rs6478109 with TNFSF15 mRNA in peripheral blood monocytes (p 5.63 x 10−6, Fig 1B, S1 Table). Importantly, in both cohorts of healthy volunteers and the cohorts of IBD and ANCA-associated vasculitis patients, the IBD protective allele (A, the minor allele in the European population) was consistently associated with increased expression of TNFSF15 (Fig 1, S1 Fig). In contrast to previous speculation that TNFSF15 predisposes to inflammatory disease, this suggests a novel protective role for this cytokine in preventing the development of human IBD.
TNFSF15 expression is rapidly upregulated in myeloid cells by stimulation via pattern recognition receptors, such as toll-like receptor ligands, and Fc receptors (FcRs) [29, 52, 53]. Although resting T cells minimally express TNFSF15, T cell receptor (TCR) stimulation results in robust expression [29, 54]. To examine the effect of cellular stimulation on the association of genotype with TNFSF15 expression, we stimulated monocytes and T cells for 4 and 24 hours, respectively (time courses depicted in S3 Fig). In monocytes stimulated with immune complex and intracellular poly(I:C), the protective allele was associated with higher levels of TNFSF15 mRNA (p 1.33 x 10−3 and p 0.0155, respectively), similar to results in unstimulated cells, while LPS-stimulated monocytes showed a comparable trend but with more variability (p 0.273, Fig 2A). A formal comparison of the eQTL in unstimulated versus stimulated monocytes revealed no significant change in the magnitude of the effect of genotype on gene expression following stimulation (Methods, S4 Fig, S1 Table). In T cells stimulated via the TCR with anti-CD3 and anti-CD28, there was no association between rs6478109 genotype and TNFSF15 expression (CD4+ T cells p 0.861, CD8+ T cells p 0.627, Fig 2B). Thus rs6478109 is a monocyte eQTL in which the IBD risk allele is associated with reduced TNFSF15 expression both at baseline and after stimulation.
Inter-individual differences apart from SNP genotype at the TNFSF15 locus might influence gene expression and thereby obscure genotype-dependent differences. Heterozygotes present a unique opportunity to control for this variability as the ratio of allelic expression can be measured within each individual using allele-specific expression (ASE) assays. rs6478109 is in LD with several intronic SNPs measurable in amplified pre-mRNA, two of which we used to examine ASE (S5 Fig). Due to low TNFSF15 pre-mRNA expression in unstimulated cells, allelic ratio measurements were only feasible in stimulated cells. In accordance with the allelic dosage effect observed in Fig 2A, such that TNFSF15 expression increased with more copies of the IBD-protective rs6478109:A allele, heterozygous monocytes stimulated with immune complex or intracellular poly(I:C) showed significant allelic imbalance favoring the protective allele (p 1.09 x 10−6, p 0.0215, respectively, Fig 2C left and middle panels). Although TNFSF15 expression in LPS-stimulated monocytes was not associated with allelic dosage by standard eQTL analysis (Fig 2A), in the internally controlled environment of heterozygous individuals, we did find significant ASE (p 1.46 x 10−5, Fig 2C right panel). In contrast, once again, stimulated T cells showed no allelic imbalance (CD4+ T cells p 0.175, CD8+ T cells p 0.424, Fig 2D and S2 Table), confirming the monocyte-specific nature of the TNFSF15 eQTL.
The most direct mechanism by which TNFSF15 mRNA expression could influence IBD susceptibility would be through controlling TNFSF15 protein levels. Stimulated monocytes express transmembrane TNFSF15 protein, but it is rapidly cleaved from their surfaces. TNFSF15 expression is thus measurable both on the cell surface and in the supernatant. To determine if the TNFSF15 eQTL extended to the protein level, we first measured surface TNFSF15 expression in immune-complex stimulated monocytes from homozygous individuals. Individuals homozygous for the IBD protective allele exhibited increased cell surface TNFSF15 (p 0.0145, Fig 3A). We next looked at soluble TNFSF15 in the supernatants of monocytes stimulated with immune complex, intracellular poly(I:C), or LPS. In the absence of stimulation, soluble TNFSF15 levels in monocyte supernatants were low (around the limit of detection) regardless of genotype. In contrast, under all stimulation conditions, soluble TNFSF15 was significantly increased in supernatants of cells from IBD protective allele homozygotes (immune complex p 6.56 x 10−4, intracellular poly(I:C) p 4.16 x 10−3, LPS p 3.42 x 10−3, Fig 3B). Thus, the levels of TNFSF15 both on the cell surface and in solution were associated with genotype. Of note, total serum TNFSF15 protein levels were not associated with genotype (p 0.962, S6A Fig), suggesting that the effect of genotype on monocyte TNFSF15 levels is relevant primarily in local cellular contexts. Secretion of other inflammatory cytokines from these monocytes was not associated with rs6478109 genotype at the same time-point, demonstrating the cis-specificity of the TNFSF15 eQTL (S6B–S6C Fig). To confirm that secreted TNFSF15 originated from newly synthesized protein, we stimulated monocytes with immune complex in the presence of either actinomycin D to block transcription (Fig 3C) or cycloheximide to block translation (Fig 3D). Both treatments resulted in loss of detectable protein, indicating that the effects of genotype on TNFSF15 protein expression were due to differences in de novo transcription and translation of TNFSF15.
Fine-mapping of genotype-phenotype associations generally requires a large number of samples to achieve the power necessary to statistically infer probable causality for one SNP over another in high LD. In order to solve this problem with a limited number of samples, we again leveraged the power of the controlled environment within heterozygous individuals. We examined ASE in immune complex-stimulated monocytes from healthy volunteers recruited specifically for having genetic cross-over events in the TNFSF15 haplotype block, such that they were heterozygous for certain SNPs but homozygous for others (Fig 4A). An earlier IBD meta-analysis by Jostins et al identified rs4246905 in the third intron of TNFSF15 as the tag SNP for disease association at this locus in European individuals [5], whereas the more recent trans-ancestry analysis described above by Liu et al [4] identified the upstream SNP rs7848647 as the most associated in the same population. To narrow the location of the eQTL causal variant to either the upstream or downstream portion of the gene, we first examined ASE in individuals heterozygous at rs6478109 in the promoter and rs4263839 in the first intron but homozygous at rs4246905 in the third intron. These individuals maintained ASE of TNFSF15 (p 1.67 x 10−4, Fig 4B), indicating that rs4246905 is not causal and suggesting that the SNP influencing TNFSF15 expression is in greater LD with the upstream SNPs than with rs4246905. Confirming this finding, individuals heterozygous at rs4246905 but homozygous at the two upstream loci showed no ASE (p 0.752, Fig 4C).
Examining the LD structure of variants in the TNFSF15 locus, we found 17 variants that are in greater LD with the upstream SNPs rs6478109 and rs4263839 than with the downstream SNP rs4246905, and that are in higher LD with these upstream SNPs than is rs4246905 (S3 Table). To further distinguish between these potential causal variants, we identified individuals homozygous at the rs6478109 promoter SNP but heterozygous at rs4263839 in the first intron and measured TNFSF15 ASE in immune-complex-stimulated monocytes. These samples lacked ASE (p 0.512, Fig 4D), indicating that rs4263839 is not the causal variant. Only three variants identified in the 1000 Genomes Phase 3 European cohort in high LD with rs6478109 (r2 > 0.8) are in higher LD with rs6478109 than with the eliminated SNP rs4263839. These are rs6478109 itself, rs7848647 (the most significant IBD-associated variant in Europeans in Liu et al [4]) and rs10817678, all of which are located upstream of TNFSF15. Two of these, rs6478109 and rs7848647, are located close to the promoter of TNFSF15, in a region containing a cluster of transcription factor binding sites, active chromatin marks and enhancers, making them the leading candidates for the causal variant (S7 Fig). These two SNPs were completely linked in all individuals recruited for ASE measurement in Figs 2 and 4.
To examine the phenotypic consequences of variation in TNFSF15 expression, we performed comprehensive immunophenotyping of T cells, B cells, monocytes, dendritic cells, and NK cells from peripheral blood of individuals homozygous for the rs6478109 polymorphism. We found no association between genotype and cell population frequencies (S8 and S9 Figs).
To test whether genetically driven variation in TNFSF15 expression under acute stimulation would result in differences in responding cell populations, we measured the effects of endogenous TNFSF15 on T cell activation. TNFSF15 costimulation promotes T cell proliferation and upregulation of the IL-2 receptor alpha chain (CD25), particularly under low levels of TCR stimulation [26, 29, 30]. We therefore examined CD4+ T cell proliferation in stimulated peripheral blood mononuclear cells from individuals homozygous for the TNFSF15 expression-associated variant rs6478109. CD8+ T cells make up a highly variable proportion of peripheral blood cells across individuals and these cells respond to TCR stimulation by making and consuming IL-2, which can affect CD4+ T cell proliferation. To avoid this confounding factor, we depleted peripheral blood mononuclear cells of CD8+ T cells before culturing the remaining cells (including CD4+ T cells and monocytes) with low level TCR stimulation for two days. Blocking TNFSF15 signaling with an antagonistic antibody resulted in decreased CD25 expression and proliferation of CD4+ T cells (Fig 5A), demonstrating the impact of endogenous TNFSF15 in this setting. Samples from individuals homozygous for the IBD protective allele that is associated with increased monocyte TNFSF15 exhibited significantly increased CD25 expression and proliferation of CD4+ T cells compared to individuals homozygous for the IBD risk allele (p 9.32 x 10−3, p 0.0289, respectively, Fig 5B and 5C). The differences in CD25 expression and proliferation were reduced with addition of anti-TNFSF15 such that these activation markers were no longer statistically significantly elevated in protective allele homozygotes (CD25 expression p 0.0939, proliferation p 0.0541, Fig 5B and 5C). While this system does not recapitulate the complex intestinal environment of the pre-morbid at-risk individual, it demonstrates that genetically regulated TNFSF15 expression can influence the ability of monocytes to costimulate responding lymphocytes, confirming the functional relevance of this TNFSF15 eQTL.
Intestinal CX3CR1+ mononuclear phagocytes derived from circulating monocytes [55–57] are likely to be the most relevant producers of TNFSF15 in the context of gut immune homeostasis [31]. To test whether phagocytic cells derived from peripheral blood monocytes maintain genotype-dependent TNFSF15 expression, we differentiated monocyte-derived macrophages (MDMs) from heterozygous individuals and measured ASE. MDM differentiated in the presence of M-CSF and GM-CSF both exhibited significant allelic imbalance favoring the minor, IBD-protective allele (p 0.0178 and p 1.39 x 10−3, respectively, Fig 6A and 6B). ASE was also maintained after stimulation of M-CSF-derived MDM with LPS for 4 hours (p 0.0251, Fig 6C). To confirm these findings in a second cohort, we mined publicly available data from an MDM eQTL study by Nedelec et al [46]. These data demonstrated a significant TNFSF15 eQTL at rs6478109 in resting MDM and after Salmonella infection (S10 Fig). The direction of effect was concordant with our monocyte and MDM data, such that expression of TNFSF15 increased with more copies of the IBD-protective allele rs6478109:A. These results demonstrate that the IBD-protective allele increases TNFSF15 expression in macrophages as well as monocytes, and is therefore likely to be relevant to the gut environment.
Understanding the mechanisms by which disease susceptibility variants influence disease risk is a key challenge of the post-GWAS era. Exploring the consequences of disease risk variants in healthy individuals allows examination of their effects in the pre-disease state and avoids the potential confounding effects of treatment and disease itself. Use of a “recall-by-genotype” bioresource enables investigation of specific polymorphisms with balanced experimental designs and facilitates in-depth investigation of causal quantitative trait loci by specific recruitment of individuals with breaks in common LD blocks. Here we used such a bioresource to investigate the functional consequences of an IBD-associated genetic variant and perform functional fine-mapping. We have demonstrated that the IBD-associated locus at TNFSF15 harbors an expression-associated polymorphism in which the rs6478109:A IBD protective allele is associated with increased monocyte TNFSF15 expression and lymphocyte costimulatory activity. Interestingly, TNFSF15 expression in stimulated T cells was not significantly associated with SNP genotype at this locus (discussed further in S1 Text), suggesting the importance of cellular context in utilization of the particular cis regulatory element that this polymorphism affects. Through ASE assays in cells from individuals specifically recruited for haplotype cross-over events, we refined the location of the causal variant for gene expression to the upstream region of the gene. Colocalization testing indicated that the eQTL and IBD association are very likely to be due to the same causal variant. This suggests that the mechanism by which SNP genotype at TNFSF15 influences IBD susceptibility is through altering TNFSF15 expression. Our work provides an example of how functional studies not only uncover the phenotypic effects of genetic variation but can also complement statistical methods for mapping disease association.
The IBD fine-mapping study by Huang et al examined the TNFSF15 locus, detailing technical difficulties with their genotyping of an indel in the region [6]. The three typed variants that interrogated the indel (chr9:117571294, chr9:117571293, and rs59418409) were assigned posterior probabilities of being causal 0.40, 0.40, and 0.11, respectively. On the basis that these three variants in fact represent a single indel (now annotated as rs35396782), the authors then summed the probabilities and concluded that this indel is the likely causal variant for IBD risk. The rs35396782 indel is 2885 base pairs upstream of TNFSF15 and in LD (r2 = 0.817) with rs6478109 (S3 Table and S11 Fig). Our analysis cannot exclude the possibility that this indel is causal for the eQTL, but the LD patterns in our functional fine-mapping and the presence of active chromatin marks in the region of the promoter SNPs rs6478109 and rs7848647 favor these candidates. Robust genotyping of the indel will be necessary to draw conclusions about its association with both IBD and TNFSF15 expression.
Previous studies describing gene expression association with genotype at TNFSF15 have yielded conflicting results [38–45] (discussed further in S1 Text). Our data unequivocally show that the IBD risk allele is associated with decreased monocyte TNFSF15 expression. Use of three genotyping platforms, two mRNA measurement technologies, and two protein measurement methods ensures the robustness of our results. In support of our findings, mining of the GTEx project database [45, 47] reveals a significant TNFSF15 eQTL at rs6478109 in whole blood with the same direction of effect that we observed. Of relevance to IBD, the GTEx database also includes a nominally significant eQTL (p <10−3) with the same direction of effect in sigmoid colon.
In the gut, peripheral blood monocytes differentiate into macrophages, which are critical for maintaining mucosal immune homeostasis [58]. A recent study by Baillie et al posited that monocyte maladaptation to macrophage differentiation and activation in the gut environment is an important driver of IBD [21]. Examination of the supporting information for their study reveals that TNFSF15 is strongly upregulated during this process. A key question is whether the TNFSF15 monocyte eQTL is maintained in macrophages and is therefore likely to be relevant to IBD pathogenesis. A previous study by Hedl et al reported association of the TNFSF15 risk haplotype SNP rs6478108:A (in phase with rs6478109:G, LD r2 0.917, S1 Fig) with increased TNFSF15 expression in M-CSF-differentiated MDM [41]. In contrast, we observed the opposite result in these cells (Fig 6), demonstrating ASE that was directionally concordant with our findings in monocytes. Through further ASE assays in macrophages from multiple differentiation and stimulation conditions, we demonstrated that the IBD protective allele is consistently preferentially expressed. We corroborated our findings by mining publicly-available data from a recent eQTL study in MDM with and without Salmonella infection, confirming that the IBD risk allele is associated with lower MDM TNFSF15 expression [46]. We thus clearly establish that, in both monocytes and macrophages, genetic predisposition to lower TNFSF15 expression is associated with IBD risk.
The association we have identified may at first seem counterintuitive given the known functions of TNFSF15. In IBD, TNFSF15 is generally considered an inflammatory marker, with TNFSF15 expression levels increasing with IBD activity [32–37]. However, such observational studies cannot distinguish causal effects from associations arising from confounding factors or the consequences of IBD (reverse causation). In animal models of inflammatory disease including colitis, asthma, arthritis, and experimental autoimmune encephalomyelitis, genetic or antibody-mediated disruption of TNFSF15 signaling through its cognate receptor TNFRSF25 (also known as DR3) generally leads to reduced pathology [29, 54, 59–63], but it is important to remember that these animal studies usually measure disease course and are poor models for disease susceptibility. At the cellular level, TNFSF15 generally promotes cytokines associated with inflammation, such as IL-2 and IFNγ from T cells, and IL-13 and IL-5 from ILC2 [26–28, 52, 64]. Indeed, we find that genetically-driven TNFSF15 enhances T cell activation in our in vitro assay. Despite this inflammatory role, studies have highlighted that TNFSF15 may be more pleiotropic than originally thought, costimulating lymphocytes that control both pro- and anti-inflammatory activities. Jia et al demonstrated a protective role for TNFSF15-TNFRSF25 interaction in acute DSS colitis and clearance of gut Salmonella enterica infection via maintenance of regulatory T cells [65], suggesting that TNFSF15 may be protective in certain contexts of intestinal inflammation. Additionally, as well as T cells and ILC2, TNFSF15 can also costimulate group 3 innate lymphocytes (ILC3), which reside in the gut and respond to TNFSF15 with enhanced IL-22 production [31, 66]. IL-22 promotes gut barrier maintenance in both infectious and non-infectious contexts [67–69]. Thus, there is also the potential for TNFSF15 to play a protective role in the gut through costimulation of ILC3.
Reduced gene expression driven by the IBD risk allele at TNFSF15 is in line with several other IBD risk variants that reduce protein function and lead to intestinal barrier disruption. For example, the T300A variant of ATG16L1 reduces autophagy in intestinal Paneth cells, dampening antimicrobial activity [11–13]. Likewise, while multiple mechanisms have been posited for the association of variants in the NOD2 locus with CD [70], disease-associated coding polymorphisms were found to reduce cellular responsiveness to peptidoglycan ligands [9] and were associated with decreased expression of Paneth cell antimicrobial peptides [10] and defective anti-bacterial responses by dendritic cells [8]. The reduction in monocyte and macrophage expression of TNFSF15 that we find associated with the IBD risk allele suggests that TNFSF15 may also promote intestinal homeostasis in the pre-morbid state.
Additional methods are included in S1 Text.
We present peripheral blood monocyte eQTLs for TNFSF15 from a previous study using samples from 39 healthy individuals and 80 patients with IBD [19], as well as a study using 45 patients with anti-neutrophil cytoplasmic antibody-associated vasculitis [18]. Gene expression data was measured on the Affymetrix Human Gene 1.1 ST Array. Microarray mRNA expression was normalized by robust multiarray averaging using the oligo package [71] and adjusted with PEER [72]. IBD patient and healthy control genotypes were measured using the Illumina Human OmniExpress12v1.0 BeadChip, and vasculitis patients were genotyped using the Affymetrix SNP6.0 Array. eQTL testing was performed using a score test implemented in the GGtools Bioconductor package [73]. Genomic locus plots were generated using the Gviz Bioconductor package [74] with RefSeq annotation for genes in the hg19 genome build.
Colocalization testing was performed using the coloc R package v2.3–1 [48]. We used the coloc.abf function and the default priors (prior probability that a SNP is associated with trait 1 = 1 x 10−4, prior probability that a SNP is associated with trait 2 = 1 x 10−4, prior probability that a SNP is associated with both traits = 1 x 10−5). Summary statistics for association of TNFSF15 and TNFSF8 expression with genotype (regression coefficients and variances) were calculated through linear regression (lm function in R) using the genotype and expression data from our previous eQTL study of healthy controls and patients with IBD [19]. For the IBD GWAS data, we used summary statistics for the European cohort from the GWAS plus Immunochip trans-ancestry MANTRA meta-analyses by Liu et al [4] (downloaded from the International Inflammatory Bowel Disease Genetics Consortium’s website, url https://www.ibdgenetics.org/, link “Latest combined GWAS and Immunochip trans-ancestry summary statistics”, file “IBD_trans_ethnic_association_summ_stats_b37.txt.gz”).
Peripheral blood samples from healthy volunteers were obtained through the Cambridge BioResource. Ninety-six blood samples were taken from 90 separate volunteers recruited based on relevant genotype at rs6478109 and/or rs4246905. All recruited individuals were Caucasian, between 18 and 65 years of age. Volunteers self-declared that they were free from autoimmune disease, cancer and human immunodeficiency virus. No individuals took regular systemic immunomodulatory therapy for at least one year before recruitment. During the recruitment process, volunteer samples were grouped by genotype, and investigators were blinded as to which group corresponded to which genotype. Derived numeric data from experiments utilizing Cambridge BioResource samples are included in S1 Data.
Whole blood was collected in CPDA tubes and passed over a Histopaque-1077 (Sigma Aldrich) gradient to separate peripheral blood mononuclear cells (PBMC). Where indicated, cell types were separated first into the CD14+ monocyte fraction and CD14- fraction by positive selection with human CD14 MicroBeads and LS columns (Miltenyi Biotec), according to the manufacturer’s protocol. The negative fraction was then enriched for CD4+ T cells or CD8+ T cells with human CD4 or CD8 MicroBeads (Miltenyi Biotec) in the same manner. For CD8-depleted PBMC, cells were separated with CD8 MicroBeads (Miltenyi Biotec) and the negative fraction collected. Purity of separated cell subsets was examined by flow cytometry as described in the “Cell subset purity QC” section. For eQTL and ASE measurements, eighty monocyte samples were sorted for use in various assays, all with over 60% purity and a median purity of 74%; thirty-five CD4+ T cell samples were sorted, all with over 92% purity and a median purity of 97%; thirty-four CD8+ T cell samples were sorted, all with purity over 70% and a median purity of 90%; sixteen PBMC samples were depleted of CD8+ T cells, all demonstrating less than 7% CD8+ T cells remaining. One CD8-depleted PBMC sample was excluded on the basis of a CD8-intermediate CD3+ population composing 13% of the PBMC population after depletion.
Purity of cell subsets was determined by flow cytometry. Cells were blocked with FcR Blocking Reagent (Miltenyi) and stained with anti-human CD14-PE (BD Biosciences); anti-human CD3-AmCyan, -FITC or–PE (clone SK7 or UCHT1, BD Biosciences); anti-human CD4-FITC (clone RPA-T4, BD Biosciences); and/or anti-human CD8-APC (clone RPA-T8, BD Biosciences) or -eFluor 450 (clone SK1, eBioscience). Flow cytometry was performed on a BD LSR Fortessa (BD Biosciences) and data analyzed in FlowJo (FlowJo, LLC).
All stimulations took place in complete medium, composed of RPMI with 10% FCS, 10 mM HEPES, 1x MEM non-essential amino acids, 1 mM sodium pyruvate, 1x GlutaMAX, 100 U/mL penicillin and 0.1 mg/mL streptomycin (Sigma or Gibco). Monocytes were stimulated with 100 ng/mL LPS (Sigma), plate-bound immune complex (as previously described [52]), or 100 μg/mL intracellular polyinosinic:polycytidylic acid (poly(I:C)), prepared with 1 mg/mL high molecular weight poly(I:C) mixed 1:1 with LyoVec transfection reagent (Invivogen) at RT for 15 minutes. Where indicated, 1 μg/mL cycloheximide or 5 μg/mL actinomycin D (Sigma) was added to cell cultures. T cells were stimulated with Dynabeads Human T-Activator CD3/CD28 (Life Technologies) at a 1:1 ratio of beads:cells. For the PBMC proliferation assay, five million CD8- cells were stained with Cell Proliferation Dye eFluor 670 (eBioscience), according to the manufacturer’s instructions. These cells were stimulated for 48 hours with 1 μg/mL anti-CD3 (OKT3), with addition of 10 μg/mL blocking anti-TNFSF15 monoclonal antibody (1A9, described under Soluble cytokine measurements) where indicated.
Monocytes were differentiated into monocyte-derived macrophages in the presence of M-CSF or GM-CSF. For M-CSF macrophages, cells were grown in 10 ng/mL recombinant human M-CSF (R and D Systems) for 7 days, adding half the volume of media with 30 ng/mL M-CSF to replenish the cytokine on day 5. For GM-CSF macrophages, monocytes were grown in 50 U/mL recombinant human GM-CSF (Peprotech) for 5 days. Where indicated, macrophages were stimulated with 10 ng/mL LPS for 4 hours. One sample recruited for monocyte-derived macrophage studies was excluded due to poor RNA yield before any measurements were made.
RNA and DNA from ex vivo and cultured cells was extracted using the AllPrep DNA/RNA Mini Kit (or RNeasy Mini Kit for extracting RNA only), using on-column DNase digestion with the RNase-Free DNase Set (Qiagen). RNA was reverse-transcribed to cDNA using the SuperScript VILO cDNA Synthesis Kit (Life Technologies). qPCR reactions were performed with Taqman Gene Expression Master Mix and Taqman Gene Expression Assays for TNFSF15 (Hs00270802_s1) or Beta-2-Microglobulin (B2M, Hs00984230_m1) (Life Technologies) on an Applied Biosystems 7900HT Fast Real-Time PCR System (Life Technologies) or CFX384 Touch Real-Time PCR Detection System (BioRad). All reactions were performed in triplicate, and the median TNFSF15 Ct value was subtracted from the median B2M Ct value for each sample to generate ΔCt values representing log expression relative to B2M.
Linear regression test statistics were calculated in R to estimate the effect of each additional copy of the minor (IBD protective) allele on gene expression. To test for an interaction between genotype and stimulation condition, a linear model with coefficients for genotype, stimulation and genotype x stimulation was fitted.
Samples were genotyped using Taqman SNP Genotyping Assays (rs6478109, C___1305297_10; rs4263839, C____120268_10; rs4246905, C____363307_20; rs7848647, C__11277159_10; rs6478108, C____170492_10) and Taqman Genotyping Master Mix (Life Technologies) according to the manufacturer’s protocol on an Applied Biosystems 7900HT Fast Real-Time PCR System (Life Technologies) or CFX384 Touch Real-Time PCR Detection System (BioRad).
ASE in heterozygous genomic DNA or reverse-transcribed cDNA from the same individual was measured by qPCR. Where ASE was measured in multiple conditions using cells from the same individuals, one genomic DNA sample from each individual was measured concurrently with cDNA samples from multiple conditions. First, the intronic region containing the target SNP was amplified from DNA using the primers and PCR conditions listed in S4 Table. Amplified regions were gel purified. A standard curve was constructed by mixing amplified genomic DNA samples from homozygous individuals 8:1, 4:1, 2:1, 1:1, 1:2, 1:4, and 1:8. Samples were then measured by qPCR in triplicate reactions using Taqman SNP Genotyping Assays for rs4263839 (C____120268_10) or rs4246905 (C____363307_20) and Taqman Gene Expression Master Mix on an Applied Biosystems 7900HT Fast Real-Time PCR System (Life Technologies) or CFX384 Touch Real-Time PCR Detection System (BioRad). Allelic ratios were calculated from VIC and FAM Ct values as described in S1 Text. Unpaired Welch's t-test statistics were calculated using GraphPad Prism Software or R. We used two intronic SNPs, rs4246905 and rs4263839 (the latter of which is in LD r2 = 0.977 with rs6478109 and is the most linked of all transcribed SNPs) to examine ASE of TNFSF15. Examination of the effect size estimates obtained from assays with each of the two SNP probes on a subset of samples from Fig 2C revealed similar detection of ASE but a slightly lower estimate by the rs4263839 probe (S5B Fig). For this reason, we do not recommend directly comparing effect sizes between samples measured with different probes.
A mouse monoclonal antibody against human TNFSF15 (clone 1A9) was generated through immunizing mice with human TNFSF15 and screening supernatants for binding TNFSF15 transfected 293T cells. The 1A9 clone also blocks soluble TNFSF15 binding to TNFRSF25 and TNFSF15 costimulation of human T cell activation ex vivo [75].
Soluble TNFSF15 in supernatants of stimulated cells was measured by custom Bio-Plex assay. Capture beads were created by conjugating anti-human TNFSF15 antibody (clone 1A9) to Bio-Plex Pro Magnetic COOH beads, region 27, using the Bio-Plex Amine Coupling Kit (BioRad) with 10 μg antibody per reaction, according to the manufacturer’s instructions. Cell culture supernatant assays were performed with 2500 beads per well for capture and 1 μg/mL biotinylated polyclonal rabbit anti-human TNFSF15 (Peprotech) for detection, using the Bio-Plex Pro Reagent Kit (BioRad) and following the manufacturer’s protocol. Data was collected on a Bio-Plex 200 or Bio-Plex MAGPIX Multiplex Reader (BioRad). Cell culture supernatant TNFSF15 concentrations were calculated using a standard curve of recombinant human TNFSF15 (Peprotech).
Serum TNFSF15 levels were also measured by custom Bio-Plex assay. Capture beads were created exactly as described for supernatant cytokine measurements but polyclonal rabbit anti-human TNFSF15 antibody (Peprotech) was used for conjugation to Bio-Plex Pro Magnetic COOH beads, region 27, with 8.5 μg antibody per reaction. Assays were run with samples diluted 1 in 4 in Sample Diluent from the Bio-Plex Pro Reagent Kit (BioRad). Values below the detection limit of the assay were set to 0. Values above the detection limit of the assay were excluded, and this is indicated in the figure legend. For genotype association testing, serum TNFSF15 protein levels were inverse-rank normalized prior to linear regression.
Inflammatory cytokines in the supernatants of stimulated monocytes were measured using the Human ProInflammatory 9-Plex Tissue Culture Kit (MesoScale Diagnostics) at the Core Biochemical Assay Laboratory in Cambridge University Hospitals or using Bio-Plex Pro reagents (BioRad), according to the manufacturer’s protocol. For plotting on a log-scale graph, 2 undetectable IL-10 measurements were set to positive values below the lowest value detected.
Mann-Whitney test statistics were calculated using GraphPad Prism Software.
Monocyte surface TNFSF15 expression was measured by flow cytometry. Cells were treated with human FcR Blocking Reagent (Miltenyi) for 5 minutes in PBS and then stained with LIVE/DEAD Fixable Aqua Dead Cell Stain Kit (Invitrogen) and either anti-TNFSF15 (clone 1A9) Alexa-Fluor 647 or mouse IgG2a Alexa-Fluor 647 (BD Biosciences) as an isotype control. Cells were analyzed on a BD LSRFortessa. Mann-Whitney test statistics were calculated using GraphPad Prism Software.
The use of genotyped human peripheral blood samples in this study was approved by the National Research Ethics Service, Cambridgeshire 2 Research Ethics Committee (08/H0308/176). All subjects provided written informed consent prior to inclusion in the study.
The data used for corroborating eQTL evidence described in this manuscript were obtained from the GTEx Portal on 10 February 2018 and from the IMMUNPOP browser (http://132.219.138.157/nedelec/eQTL/ and [46]) on 7 March 2018. Regulatory information about the TNFSF15 promoter region was visualized in the UCSC genome browser (http://genome.ucsc.edu/ and [76], hg19 build) on 17 Feb 2018. Tracks included transcription factor binding sites from ENCODE transcription factor ChIP-seq of 161 factors across a variety of cell types, DNase hypersensitivity signal measured by ENCODE/UW, and CTCF and histone ChIP-seq profiles from ENCODE/BROAD [77]. CAGE data for enhancer identification were downloaded from the ZENBU browser (http://fantom.gsc.riken.jp/zenbu/gLyphs/#config=dXO5cTaJBZiiw73fJq2oGD;loc=hg19::chr9:117566249..117571251+ and [21, 78]), on 17 February 2018.
|
10.1371/journal.pcbi.1007381 | LOTUS: A single- and multitask machine learning algorithm for the prediction of cancer driver genes | Cancer driver genes, i.e., oncogenes and tumor suppressor genes, are involved in the acquisition of important functions in tumors, providing a selective growth advantage, allowing uncontrolled proliferation and avoiding apoptosis. It is therefore important to identify these driver genes, both for the fundamental understanding of cancer and to help finding new therapeutic targets or biomarkers. Although the most frequently mutated driver genes have been identified, it is believed that many more remain to be discovered, particularly for driver genes specific to some cancer types. In this paper, we propose a new computational method called LOTUS to predict new driver genes. LOTUS is a machine-learning based approach which allows to integrate various types of data in a versatile manner, including information about gene mutations and protein-protein interactions. In addition, LOTUS can predict cancer driver genes in a pan-cancer setting as well as for specific cancer types, using a multitask learning strategy to share information across cancer types. We empirically show that LOTUS outperforms five other state-of-the-art driver gene prediction methods, both in terms of intrinsic consistency and prediction accuracy, and provide predictions of new cancer genes across many cancer types.
| Cancer development is driven by mutations and dysfunction of important, so-called cancer driver genes, that could be targeted by specific therapies. While a number of such cancer genes have already been identified, it is believed that many more remain to be discovered. To help prioritize experimental investigations of candidate genes, several computational methods have been proposed to rank promising candidates based on their mutations in large cohorts of cancer cases, or on their interactions with known driver genes in biological networks. We propose LOTUS, a new computational approach to identify genes with high oncogenic potential. LOTUS implements a machine learning approach to learn an oncogenic potential score from known driver genes, and brings two novelties compared to existing methods. First, it allows to easily combine heterogeneous sources of information into the scoring function, which we illustrate by learning a scoring function from both known mutations in large cancer cohorts and interactions in biological networks. Second, using a multitask learning strategy, it can predict different driver genes for different cancer types, while sharing information between them to improve the prediction for every type. We provide experimental results showing that LOTUS significantly outperforms several state-of-the-art cancer gene prediction software.
| In our current understanding of cancer, tumors appear when some cells acquire functionalities that give them a selective growth advantage, allowing uncontrolled proliferation and avoiding apoptosis [1, 2]. These malignant characteristics arise from various genomic alterations including point mutations, gene copy number variants (CNVs), translocations, inversions, deletions, or aberrant gene fusions. Many studies have shown that these alterations are not uniformly distributed across the genome [3, 4], and target specific genes associated with a limited number of important cellular functions such as genome maintenance, cell survival, and cell fate [5]. Among these so-called driver genes, two classes have been distinguished in the literature: tumor suppressors genes (TSGs) and oncogenes (OGs) [6, Chapter 15]. TSGs, such as TP53 [7], participate in defense mechanisms against cancer and their inactivation by a genomic alteration can increase the selective growth advantage of the cell. On the contrary, alterations affecting OGs, such as KRAS [8] or ERBB2 [9], can be responsible for the acquisition of new properties that provide some selective growth advantage or the ability to spread to remote organs. Identifying driver genes is important not only from a basic biology point of view to decipher cancer mechanisms, but also to identify new therapeutic strategies and develop precision medicine approaches targeting specifically mutated driver genes. For example, Trastuzumab [10] is a drug given against breast cancer that targets the protein precisely encoded by ERBB2, which has dramatically improved the prognosis of patients whose tumors overexpress that OG.
Decades of research in cancer genomics have allowed to identify several hundreds of such cancer genes. Regularly updated databases such as the Cancer Gene Census (CGC) [11], provide catalogues of genes likely to be causally implicated in cancer, with various levels of experimental validations. Many cancer genes have been identified recently by systematic analysis of somatic mutations in cancer genomes, as provided by large-scale collaborative efforts to sequence tumors such as The Cancer Genome Atlas (TCGA) [12] or the International Cancer Genome Consortium (ICGC) [13]. Indeed, cancer genes tend to be more mutated than non-cancer genes, providing a simple guiding principle to identify them. In particular, the COSMIC database [14] is the world’s largest and most comprehensive resource of somatic mutations in coding regions. It is now likely that the most frequently mutated genes have been identified [15]. However, the total number of driver genes is still a debate, and many driver genes less frequently mutated, with low penetrance, or specific to a given type of cancer are still to be discovered.
The first methods to identify driver genes from catalogues of somatic mutations simply compared genes based on somatic mutation frequencies, which was proved to be far too basic [16]. Indeed, mutations do not appear uniformly on the genome: some regions of the genome may be more affected by errors because they are more often transcribed, so that some studies actually overestimated the number of driver genes because they were expecting lower mutation rates than in reality. Mathematically, they were formulating driver prediction as a hypothesis testing problem with an inadequate null hypothesis [17]. Several attempts have been made to adequately calibrate the null hypothesis, like [16] or [18], where it is assumed that mutations result from a mixture of several mutational processes related to different causes.
A variety of bioinformatics methods have then been developed to complete the list of pan-cancer or cancer specific driver genes. Globally, they fall into three main categories. First, a variety of “Mutation Frequency” methods such as MuSiC [19] or ActiveDriver [20] identify driver genes based on the assumption that they display mutation frequencies higher than those of a background mutation model expected for passenger mutations. However, this background rate may differ between cell types, genome positions or patients. In order to avoid such potential bias, some methods like MutSigCV [21] derive a patient-specific background mutation model, and may take into account various criteria such as cancer type, position in the genome, or clinical data. Second, “Functional impact” methods such as OncodriveFM [22] assume that driver genes have higher frequency of mutations expected to impact the protein function (usually missense mutations) than that observed in passenger genes. Third, “Pathway-based” methods consider cancer as a disease in which mutated genes occupy key roles in cancer-related biological pathways, leading to critical functional perturbations at the level of networks. For example, DriverNet [23] identifies driver genes based on their effect in the transcription networks. Although these methods tend to successfully identify the most frequently mutated genes, their overall prediction overlap is modest. Since they rely on complementary statistical strategies, one could recommend to use them in combination, as CompositeDriver allows us to do [24]. The results of some of these tools are available at the Driver DB database [25].
Some methods integrate information on mutation frequency and functional impact of mutations, or other types of data such as genome position, copy number variations (CNVs) or gene expression. The underlying idea is that combining data should improve the prediction performance over tools that use a single type of information. For example, TUSON [26] or DOTS-Finder [27] combine mutation frequencies and functional impact of mutations to identify OGs and TSGs. Also in this category, the 20/20+ method [28] encodes genes with features based on their frequency and mutation types, in addition to other biological information such as gene expression level in difference cancer cell lines [29] or replication time. Then, 20/20+ predicts driver genes with a random forest algorithm, which constitutes the first attempt to use a machine learning method in this field. In [28], the authors benchmark 8 driver gene prediction methods based on several criteria including the fraction of predicted genes in CGC, the number of predicted driver genes and the consistency. Three methods proved to perform similarly on all criteria, and better than the five others: TUSON, MutSigCV, and 20/20+, validating the relevance of combining heterogeneous information to predict cancer genes.
In the present paper, we propose a new method for cancer driver gene prediction called Learning Oncogenes and TUmor Suppressors (LOTUS). Like 20/20+, LOTUS is a machine learning-based method, meaning that it starts from a list of known driver genes in order to “learn” the specificities of such genes and to identify new ones. In addition, LOTUS presents two unique characteristics with respect to previous work in this field. First, it combines three types of features likely to contain relevant information to predict cancer genes (mutation frequency, functional impact, and pathway-based informations). This integration of heterogeneous information is carried out in a unified mathematical and computational framework thanks to the use of kernel methods [30], and allows in principle to integrate other sources of data if available, such as transcriptomic or epigenomic information. More precisely, in our implementation, we predict cancer driver genes based not only on gene mutations features like “Mutation Frequency” and “Functional Impact” methods do, but also on known protein-protein interaction (PPI) network like “Pathway-based” methods do. Indeed, the use of PPI information is particularly relevant since it has been reported that proteins encoded by driver genes are more likely to be involved in protein complexes and share higher “betweenness” than a typical protein [26]. Moreover, it has been successfully used by HotNet2 [31] to detect gene pathways enriched in driver genes, and in [32] for cancer driver prediction. Second, LOTUS can predict cancer genes in a pan-cancer setting, as well as for specific cancer types, using a multitask learning strategy [33].
Although many efforts are devoted to identify cancer-specific genes based on experimental approaches, in in-silico approaches, the pan-cancer setting has been adopted by most available prediction methods, since more data are available to train models when gathering all cancer types. Prediction of drivers for specific cancer types has been less explored so far, because the number of known driver genes for a given cancer is often too small to build a reliable prediction model, and because the amount of data such as somatic mutations to train the model is smaller than in the pan-cancer setting. However, the search for cancer specific driver genes is relevant, because cancer is a very heterogeneous disease: different tumorigenic processes seem to be at work in different tissue types, and consequently, each cancer type probably has its own list of driver genes [15]. LOTUS implements a multitask algorithm that predicts new driver genes for a given cancer type based on its known driver genes, while also taking into account the driver genes known for other types of cancer according to their similarities with the considered type of cancer. Such approaches are of particular interest when the learning data are scarce in each individual tasks: they increase the amount of data available for each task and thus perform statistically better. To our knowledge, while a similar approach was used to predict disease genes across general human diseases [34], this is the first time a multitask machine learning algorithm is used for the prediction of cancer driver genes.
We compare LOTUS to five state-of-the art cancer prediction methods. We show that LOTUS outperforms the state-of-the-art in its ability to identify novel cancer genes, and clarify the benefits of heterogeneous data integration and of the multitask learning strategy to predict cancer type-specific driver genes. Finally, we provide predictions of new cancer genes according to LOTUS, as well as supporting evidence that those predictions are likely to contain new cancer genes.
We propose LOTUS, a new method that predicts cancer driver genes. LOTUS is a machine learning-based method that estimates a scoring function to rank candidate genes by decreasing probability for them to be OGs or TSGs, given a training set of known OGs and TSGs. The score of a candidate gene is a weighted sum of similarities between the candidate gene and the known driver genes, where the weights are optimized by a one-class support vector machine (OC-SVM) algorithm. The similarities between genes are calculated based on gene features that are derived from the analysis of somatic mutation patterns in the genes (see Materials ans methods section for a description of these features), or from the relative positions of genes in a PPI network, or from both; the mathematical framework of kernel methods allows to simply combine heterogeneous data about genes (i.e., patterns of somatic mutations and PPI information) in a single model.
Another salient feature of LOTUS is its ability to work in a pan-cancer setting, as well as to predict driver genes specific to individual cancer types. In the later case, we use a multitask learning strategy to jointly learn scoring functions for all cancer types by sharing information about known driver genes in different cancer types. We test both a default multitask learning strategy, that shares information uniformly across all cancer types, and a new strategy that shares information across cancer types according to their degree of similarity. More details about the mathematical formulation and algorithms implemented in LOTUS are provided in the Material and Methods section.
In the following, we assess the performance of LOTUS first in the pan-cancer regime, i.e. in the single task setting, and compare it to five state-of-the-art methods (TUSON, MutSigCV, 20/20+, MUFFINN and DiffMut), and second in the cancer type specific regime, where we illustrate the importance of the multitask learning strategies.
We first study the pan-cancer regime, where cancer is considered as a single disease, and where we search for driver genes involved in at least one type of cancer. Several computational methods have been proposed to solve this problem in the past, and we compare LOTUS with five well-known state-of-the-art methods [28]: MutSigCV [21], which is a frequency-based method, TUSON [26], 20/20+ [28], which combines frequency and functional information, MUFFINN [32] and DiffMut [35], which takes the mutation patterns of genes into account.
As explained in the Materials and methods section, to perform a fair comparison between different methods we use different training databases for LOTUS adapted to each competing methods. We measure the performance of each method in terms of consistency error (CE), which estimates the mean number of non-driver genes that a given prediction method ranks higher than known driver genes; hence the smaller the CE the better the method. The results in terms of cross-validated CE for OG and TSG prediction are presented in Table 1 for TUSON and MUFFINN, in Table 2 for 20/20+ and DiffMut, and in Table 3 for MutSigCV. When analyzing these results, one should keep in mind that the total number of cancer driver genes is still a subject of debate, but it is expected to be much lower than the size of the test set (which depends on the method but is of the order of 18, 000), and it should rather be in the range of a few hundreds. Therefore, CE above a few thousand can be considered as poor performance results.
These results show that LOTUS strongly outperforms all other algorithms in term of CE, for both TSG and OG predictions. More precisely, for OG predictions, LOTUS is about three times better than MutSigCV and TUSON, twice better than 20/20+ and MUFFINN, and five times better than DiffMut, in terms of CE. For TSG predictions, the reduction in CE with LOTUS is two-fold compared to MutSigCV and 20/20+, and five-fold compared to TUSON, MUFFINN DiffMut. Note that the performance is overall much better in the first two experiments, which are also easier because they provide larger mutational data.
It is interesting to note that, for all methods except in the MutSigCV experiment, the performances obtained for OG do not reach those obtained for TSG, suggesting that OG prediction is a more difficult problem than TSG prediction. This reflects the fundamental difference between TSG mutations and OG mutations: the first lead to loss-of-function and can pile up, while the second are gain-of-function mutations and have a much more subtle nature. In addition, gain-of-function can also be due to overexpression of the OG, which can arise from other mechanisms than gene mutation. One way to improve the OG prediction performance may be to include descriptors better suited to them, such as copy number. Moreover, as mutations affecting OGs are not all likely to provide them with new functionalities, many mutations on OGs present in the database and used here might not bear information on OGs. Therefore, relevant information on OGs is scarce, which makes OG prediction more difficult. In addition, the data themselves might also contribute to difference in performance between TSG and OG prediction. For example, in the case of the TUSON train set, although the TSG and OG train sets both contain 50 genes, the mutation matrix that we used to build the gene features contains 13, 525 mutations affecting TSGs and 7, 717 mutations affecting OGs. Therefore, the data are richer for TSG, which might contribute to the difference in prediction performance.
LOTUS, 20/20+, MutSigCV, DiffMut, MUFFINN and TUSON differ not only by the algorithm they implement, but also by the type of data they use to make predictions: in particular, TUSON and 20/20+ use only mutational data while LOTUS uses PPI information in addition to mutational data. To highlight the contributions of the algorithm and of the PPI information to the performance of LOTUS, we ran LOTUS with Kgenes equal to Kmutation or KPPI, i.e., with only mutation information, or only PPI information.
The results are presented in the first two columns of Tables 4 and 5, respectively for OG and TSG. The last column of these tables recalls the performance obtained when mutation and PPI information are both used (values reported from Tables 1, 2 and 3).
These results show that, both for OG and TSG, using both mutation and PPI information dramatically improves the prediction performance over using only one of them. This underlines the fact that mutation and PPI bear complementary information that are both useful for the prediction tasks. The performances obtained with only PPI information are similar for OG and TSG, which seems to indicate that this information contributes similarly to both prediction tasks. On the contrary, the performances obtained using only mutation information are much better for TSG than for OG. This is consistent with the above comment that mutation information is more abundant in the database and more relevant in nature for TSG than for OG. It is also consistent with the fact that using Kmutation alone outperforms using KPPI alone for TSGs, while the opposite is observed for OGs.
A possible issue with PPI network-based models is that they may be subject to study bias, in the sense that cancer genes tend to be more studied and thus have higher degree in the PPI network. Hence learning an association between cancer genes and degree in a PPI network may be problematic since this is not purely a biological feature, and we may not be able to extrapolate the model to new, less studied cancer genes. This concern is supported by the observation that the gene rank computed by LOTUS with PPI kernel significantly correlates with the gene degree in the PPI network (Fig 1). This correlation only slowly decreases when one removes the top genes in the list, which indicates that this phenomenon can not only be attributed to a few genes (as opposed to, e.g., supplementary Figure 22 in [31]).
To assess whether the good performance of LOTUS using PPI is only due to the gene degree information, we performed two further experiments. First, we ran LOTUS with the kernel Kdegree defined by:
( K d e g r e e ) i , j = d i d j ,
where di is the degree of i in the PPI network. This allows LOTUS to capture in its model a linear function of the degree. We see in Tables 4 and 5 that the degree kernel has in almost all cases worse performance than the PPI kernel, which confirms that the number of neighbors alone contains less relevant information in relation with the driver prediction problem than the PPI kernel does. Second, we trained LOTUS with a PPI kernel derived from a randomized PPI network, where we kept the network structure but randomly shuffled the genes while approximately preserving their degree. For that purpose, we binned the genes in 20 groups of roughly equal sizes, by decreasing degree in the network, and randomly shuffled the genes within each group. After this random shuffling, each gene has approximately the same degree as in the initial PPI network, but not the same neighborhood. We repeated the random shuffling 100 times and computed the performance of LOTUS with the corresponding PPI kernels. We observed that the performance was worse with the randomized PPI network than with the original PPI network in 95% of the cases, confirming that LOTUS with the PPI kernel uses more than the mere degree information to predict cancer genes.
Furthermore, we examined the first predictions (excluding already known driver genes) of LOTUS with the 20/20 datasets, when both the mutation and the PPI kernels are used. Among the first 50 TSGs (described by the number of frameshift, LOF and splice mutations), 26 have less than 3 mutations of each kind, 4 predicted TSGs even having no mutation at all. This demonstrates that LOTUS predictions strongly benefit from the PPI information, and that some of these genes would have never been detected using mutation data only.
Finally, note that gene lengths and PPI-network degrees are not used explicitly as features by LOTUS, although these characteristics correlate with our predictions. Hence, LOTUS retrieves implicit relevant characteristics of cancer driver genes from the mutations and the PPI-network.
We now evaluate the generalization properties of the different methods on new unseen data as external test set. This could mitigate the potential bias in the evaluation of the performance of TUSON, DiffMut and 20/20+ based on cross-validation experiments, as in the previous paragraph. For that purpose, we evaluate the performance of the different methods when predicting supposedly “difficult” new cancer genes (an independent test set), which have only been added recently in CGCv86. We train on the one hand LOTUS, MUFFINN and TUSON with the TUSON mutation database and driver gene train sets, and on the other hand LOTUS, DiffMut and 20/20+ with the 20/20 mutation database and driver gene train sets. Then, we make predictions on the remaining genes in COSMIC, and count how many driver genes in CGCv86 appear among the 20, 50 and 100 first predictions. Note that the driver genes from the train sets were excluded from the predictions. The results are shown in Tables 6–9 and are illustrated by corresponding receiver operating characteristic (ROC) curves in Figs 2 and 3.
First, we observe that TUSON outperforms LOTUS in almost all these experiments. Second, LOTUS outperforms DiffMut and MUFFINN in all experiments. Third, LOTUS is better than 20/20+ for TSG detection, and the contrary holds for OGs. Generally speaking, the first predictions of TUSON and 20/20+ are more reliable than LOTUS’s, but, as shown in Figs 2 and 3, LOTUS outperforms all the methods when all genes are considered, and not only the first 20 to 100 genes.
The good performance of TUSON and 20/20+ for the top ranked genes compared to those of LOTUS could be explained by the fact that, all genes in CGCv86 so far have been reported through analysis of mutation data (cf. CGC web page: ‘The Cancer Gene Census (CGC) is an ongoing effort to catalogue those genes which contain mutations that have been causally implicated in cancer’). This interpretation would also explain why LOTUS hardly agrees with the other methods when comparing the top ranked genes. Indeed, for the 20 top predictions of LOTUS (excluding training sets), the intersection with TUSON consists only in two TSGs and one OG, the intersection with DiffMut consist of one TSG and two OGs, and the intersection with 20/20+ consists in four TSGs and one OG. Since LOTUS and MUFFINN are the only methods, among those considered here, that use network information in addition to mutation data, this could explain less overlap between their predictions and those of the other methods, and a lower overlap with CGC for the top ranked genes, since no PPI information was used to establish the CGC database.
To underpin this hypothesis, we computed the total number of non-silent mutations for the 20, 50 and 100 first predictions for all method, i.e., the top-ranked genes not belonging to the training sets. The result in Tables 10 and 11 show that predictions from LOTUS have lower mutations rates than those of TUSON, 20/20+ and DiffMut, especially for TSGs. This demonstrates that LOTUS relies less on mutation data than these methods. Similarly, MUFFINN, that also uses network information, tends to rank genes with fewer mutations on top of the list.
We tested the ability of LOTUS to make new driver gene predictions. We trained LOTUS with the CGCv86 train set, made predictions over the complete COSMIC database (19,320 genes including the training sets). The complete results are given in S1 Table. Complete analysis of the predicted OGs and TSGs rankings is out of the scope of this paper. However, we considered the 22/21 best ranked TSGs and OGs, and made bibliographic search in order to look for independent information that could validate these predictions.
For the 22 best ranked predicted OGs, we found abundant literature reporting their implication in various cancers. It is not possible to make a full review for each of these genes, but we group them based on their global functions and focus on some examples.
Twelve out of the 22 best ranked genes are related to transcription regulation, a mechanism that is invariably perturbed in cancer. These 12 genes are involved at various levels of transcription regulation: chromatin remodeling (PYGO1, PYGO2, EP300, DOT1L), transcription factors or repressors (MSEI1, MSEI2, MSEI3, TFEC, NKX2-2, ZIK1), transcription regulation via miRNA (DROSHA, ELF1). For all these genes, a corpus of publications confirm their role in promoting diverse types of cancers. For example, PYGO1 is involved in colorectal cancer [36], while PYGO2 was shown to be a tumor promoter in mice [37]. MSEI1, MSEI2, and MSEI3 are involved in the etiology, progression and metastatic evolution of some cancer types such as prostate cancer [38], or leukemia [39, 40]. DROSHA is involved in the miRNA depletion observed in lung cancer, and alterations in this gene was shown to have a remarkable impact in lung cancer [36].
Eight other best ranked OGs belong to signaling pathways known to play a role in cancer. Among them, FGF6 and FGF5 are growth factors from the fibroblast growth factors signaling pathway, which are well known players contributing to tumor progression [41]. Similarly MOB3B (or MOB1) is a pivotal kinase player in the Hippo tumor suppressor pathway, and mutations in this gene is associated to prostate cancer susceptibility and agressive tumors [42].
Although not exhaustive, these findings indicate that the best ranked oncogenes predicted by LOTUS are realistic OG for some cancer types.
As for OGs, for the 21 best ranked TSGs, we found many publications indicating their role in cancer, and strikingly, APOM is actually a known TSG for hepatocellular carcinoma [43]. We will group some of the other TSGs based on their function, and discuss a few examples. Five genes are involved in DNA repair, a role closely related to genome maintenance and cancer [44, 45], and were shown to have a protective role in various types of cancer (only one paper is cited by gene, but many others confirm this role): EXO1 [46], ERCC1 [47], GTF2H1 [48], MDC1 [49], and DGCR8 [50].
Three genes are involved in immune system response to cancer, which clearly indicates that they bear a protective role: PDCD1 [50], KLRG1 [51], and MUC16 [52].
For six other genes of various functions, we found recent publications indicating that they could potentially act as TSGs: SPTA1, GALNT5, PIWIL1, PIWIL4, SNX5, ADAM6. Mutations in SPTA1 was found to play a role in glioblastoma [53]. The expression of the 5 other genes were found to be repressed by over-expressed non-coding RNA or by aberrant methylation of the promoter: GALNT5 in gastric cancer [54], ADAM6 in lung adenocarcinoma [55], PIWIL1 and PIWIL4 in lung adenocarcinoma [56] and renal cell carcinoma [57]. Loss or decreased expression SNX5 promotes thyroid cancer progression [58].
Intriguingly, 3 of the best 21 ranked predicted TSGs, bibliographic search provided clues that they indeed play a role in cancer, but that they would rather behave as OG. These genes are CENPU [59], FXYD2 [60], ANXA9 [61]. In fact, the literature provides other examples of genes able to switch from oncogenes to tumor suppressor genes, depending on the context [62], which could be the case for these genes. In most cases, the cited papers (and others) observe that over-expression of these genes are observed in various types of cancers. One assumption could be that variations of their levels of expression might lead to switch between TSG and OG roles.
Interestingly, among the 50 best ranked TSGs, 4 genes are not mutated (CENPU, AEBP2, CDX1, ZNF652), and 22 genes are rarely mutated, less than 3 times. Such cases are not observed among the top ranked OGs. Indeed, in the case of TSG, decrease in expression or gene deletion leads to the same effect, i.e. loss of function, as deleterious mutations within the gene sequence. A TSG that loses its function based on default in expression cannot be retrieved by prediction methods based only on mutations, and LOTUS probably classified these rarely mutated genes probably according to their interactions in the PPI network. The case of CENPU is already discussed above. We could confirm that CDX1 is a known TSG [63]. Silencing of AEBP2 with long non coding RNA is associated to melanoma [64], while its deletion is observed in myeloid leukemia [65]. Similarly, silencing of ZNF652 by miRNA is involved in lymphoma [66], while some genetic variants of this gene lead to higher risks of prostate cancer [67]. Concerning the 22 rarely mutated TSGs, 9 of them are already discussed above, and most of the others are known to play a role in cancer. As an example, TNIP2 may be a TSG in pancreatic tumors, since it was shown that inhibition of MiR-1180, a short non coding mi-RNA over-expressed in pancreatic cancer targeting TNIP2, inhibited cell proliferation [68].
Taken together, these results show that, among the top TSG and OG ranked by LOTUS, many genes are indeed involved in cancer, and that LOTUS predictions correspond to relevant genes that are reliable candidates as cancer driver genes, at least for some tumor types.
In this section, we do not consider cancer as a single disease, but as a variety of diseases with different histological types that can affect various organs. Indeed, an important question in cancer research is to identify driver genes for each type of cancer. One way to solve this problem is to use a prediction method that is trained only with driver genes known for the considered cancer. Such single-task methods may however display poor performance because the number of known drivers per cancer is often too small to derive a reliable model. Indeed, scarce training data lead to a potential loss of statistical power as compared to the problem of identification of pan-cancer driver genes where data available for all cancers are used.
In this context, we investigate two multitask versions of LOTUS, where we predict driver genes for a given cancer based on the drivers known for this cancer but also on all driver genes known for other cancer types. For a given cancer type, this may improve driver genes prediction by limiting the loss of statistical power compared to the aforementioned single-task approach.
For that purpose, we derive a list of 30 cancer diseases from the 20/20 mutation dataset as explained in Methods. This complete list is available in S2 Table. As expected, many cancer types have only few known cancer genes (Fig 4).
Since we want to evaluate the performance of LOTUS in a cross-validation scheme, we only consider diseases with more than 4 known driver genes in order to be able to run a 2-fold cross-validation scheme. This leads us to keep 27 cancer types for TSG prediction and 27 for OG prediction. Note however that, for each cancer type, prediction are made while sharing all the driver genes known for the 30 diseases, according to their similarities with these cancer types.
The 2-fold cross-validated CE of LOTUS for the 27 considered cancer types is presented in Table 12 (for TSG) and Table 13 (for OG). We compare four variants of LOTUS, as explained in Methods: (1) single-task LOTUS treats each disease in turn independently from the others using only the mutation data related to the considered disease to calculate gene features, and only the driver genes known for this disease are used to train the algorithm; (2) Aggregation LOTUS is also a single-task version of LOTUS, but gene features are calculated using the complete mutation database of gene mutations in all cancers. In addition, for each disease, the train set consists of known drivers for all the other cancers and have of the drivers known for the considered disease. Then the prediction performance are calculated for the other half of known drivers for this disease, which constitute the test set in the 2-fold cross validation scheme. Therefore, Aggregation LOTUS is a single-task algorithm that uses much richer information than the basic Single-task LOTUS; (3 and 4) Two multitask versions of LOTUS use either a standard multitask strategy that does not take into account the relative similarities between diseases (Multitask LOTUS), or a more refined multitask strategy where driver gene information is shared between cancer types according to their similarity based on biological information (Multitask LOTUS2). Finally, we compare these performances with those of DiffMut, as a single-task method using only the mutation data related to the considered disease, as for single-task LOTUS.
For most diseases (25/27 for TSG, 27/27 for OG), single-task LOTUS and DiffMut lead to the worst CE, confirming the difficulty to treat each cancer type individually, due to the small number of known driver genes and to the smaller mutation database available for each cancer type type. Interestingly, Aggregation LOTUS often leads to a strong improvement in performance. This shows that different cancer types often share some common mechanisms and driver genes, and therefore, simply using all the available information as in a pan-cancer paradigm improves the performance of driver gene prediction for each disease. However, in many cases, the multitask LOTUS and LOTUS2 algorithms lead to an additional improvement over Aggregation LOTUS, LOTUS2 leading in general to the best results (in 17 types out of 27 for TSG prediction, and in 17 types out of 27 for OG prediction). On average, the decrease in CE between Aggregate LOTUS and LOTUS2 is of 20% for OG and 18% for TSG. The improvement in performance observed between Aggregate LOTUS and LOTUS2 shows that, besides some driver mechanisms common to many cancers, each cancer presents some specific driver mechanisms that can only be captured by prediction methods able to integrate some biological knowledge about the different diseases. The above results show that multitask algorithms allowing to share information between cancers according to their biological similarities such as LOTUS2, rather than on more naive rules, better capture these specific driver genes. They also show that the kernel Kdiseases = Kdescriptors built on disease descriptors contains some relevant biological information to compare diseases.
To measure how different the predictions of LOTUS2 are for each cancer type, we compared the first 50 predictions for each type. Aggregating all predictions for TSGs (respectively OGs) results in 210 genes (respectively 224 genes), which shows that various cancer types share some drivers, but that the prediction lists are different. Indeed, some drivers with high penetrance (such as TP53) are expected to be found in most cancer types, whereas other drivers are more specific to given organs or cell types, in particular since all genes are not expressed in all cell types.
In addition, multitask algorithms based on task descriptors (here, disease descriptors) appear to be promising in order to include prior knowledge about diseases and share information according to biological features characterizing the diseases.
Finally, note that we did not try to run TUSON, MutSigCV, MUFFINN or 20/20+ to search for cancer specific driver genes in the single-task setting (they cannot be run in the multi-task setting). Indeed, according to the results of pan-cancer studies in the single-task setting, they do not perform as well as single-task LOTUS. Considering that single-task LOTUS and DiffMut were far from reaching the performance of multi-task LOTUS for prediction of cancer specific driver genes, TUSON, MutSigCV, MUFFINN and 20/20+ are not expected to reach these performance either.
Our work demonstrates that LOTUS outperforms several state-of-the-art methods on all tested situations for driver gene prediction. This improvement results from various aspects of the LOTUS algorithm. First, LOTUS allows to include the PPI network information as independent prior biological knowledge. In the single-task setting, we proved that this information has significance for the prediction of cancer driver genes. Because LOTUS is based on kernel methods, it is well suited to integrate other data from multiple sources such as protein expression data, information from chip-seq, HiC or methylation data, or new features for mutation timing as designed in [69]. Further development could involve the definition of other gene kernels based on such type of data, and combine them with our current gene kernel, in order to evaluate their relevance in driver gene prediction.
We also showed how LOTUS can serve as a multitask method. It relies on a disease kernel that controls how driver gene information is shared between diseases. Interestingly, we showed that building a kernel based on independent biological prior knowledge about disease similarity leads on average to the best prediction performance with respect to single-task algorithms, and also with respect to a more generic and naive multitask learning strategy that does not incorporate knowledge about the cancer types. Again, the kernel approach leaves space for integration of other types and possibly more complex biological sources of information about diseases. Our multitask approach thus allows to make prediction for cancer types with very few known driver genes, which would be less reliable with the single-task methods. We considered here only diseases with at least 4 known driver genes, in order to perform cross-validation studies, which was necessary to evaluate the methods. However, it is important to note that in real-case studies, at the extreme, both versions of multitask LOTUS could make driver gene prediction for the 30 cancer types, including those for which no driver gene is known.
LOTUS is a machine learning algorithm based on one-class SVM. In fact, the most classical problem in machine learning is binary classification, where the task is to classify observations into two classes (positives and negatives), based on training sets P of known positives and N of known negatives. Driver gene detection can be seen as binary classification of TSGs vs. neutral genes, and of OGs vs. neutral genes. However, although the P set is composed of known driver genes, it is not straightforward to build the N set because we cannot claim that some genes cannot be drivers. Thus, driver gene detection should rather be seen as binary classification problem with only one training set P of known positives. This problem is classically called PU learning (for Positive-Unknown), as opposed to PN learning (for Positive-Negative).
The classical way to solve PU learning problems is to choose a set N of negatives among the unlabeled data and apply a PN learning method. For example, one can consider all unknown items as negatives (some of which may be reclassified afterwards as positives), or randomly choose bootstrapped sets of negatives among the unknown, like in [34]. Both methods assume that a minority of the unlabeled items are in fact positives, which is expected for driver genes.
The one-class SVM algorithm [70] can also be used as a PU learning method, in which a virtual item is chosen as the training set of negatives. We preferred this approach because in preliminary studies, we found that it had slightly better performances than PU learning methods and was also faster.
For LOTUS, as for all machine learning algorithm, the set of known driver genes is critical: if this set is poorly chosen (i.e., if some genes were wrongly reported as driver genes, or more likely, if the reported genes are not the best driver genes), the best algorithm might not minimize the CE. To circumvent this problem, we propose two new approaches for future developments: one could build a multi-step algorithm that iteratively removes some genes from the positive set and labels them as unknown, and relabel as positives some of the best ranked unknown genes. We believe that such an algorithm would make the set of positives converge to a more relevant list. Alternatively, one could assign (finite) scores to the known driver genes before performing classification and increment these scores at each step.
LOTUS is a new machine learning-based method to predict new cancer driver genes, given a list of know ones. In the simplest, pan-cancer setting, we consider a list of N known driver genes {g1, …, gN}, and the goal of LOTUS is to learn from them a scoring function f(g), for any other gene g, that predicts how likely it is that g is also a cancer gene. Since TSGs and OGs have different characteristics, we treat them separately and build two scoring functions fTSG and fOG that are trained from lists of know TSGs and OGs, respectively.
LOTUS learns the scoring function f(g) with a one-class support vector machine (OC-SVM) algorithm [70], a classical method for novelty detection and density level set estimation [71]. The scoring function f(g) learned by a OC-SVM given a training set {g1, …, gN} of known cancer genes takes the form:
f ( g ) = ∑ i = 1 N α i K ( g i , g ) , (1)
where α1, …, αN are weights optimized during the training of OC-SVM [70], and K(g, g′) is a so-called kernel function that quantifies the similarity between any two genes g and g′. In other words, the score of a new gene g is a weighted combination of its similarities with the known driver genes.
The kernel K encodes the similarity among genes. Mathematically, the only constraint that K must fulfill is that it should be a symmetric positive definite function [30]. This leaves a lot of freedom to create specific kernels encoding prior knowledge about relevant information to predict driver genes. In addition, one can easily combine heterogeneous information in a single kernel by, e.g., summing two kernels based on different sources of data. In this work, we restrict ourselves to the following basic kernels, and leave for future work a more exhaustive search of optimization of kernels for cancer gene prediction.
The pan-cancer LOTUS approach can also be used for cancer-specific predictions, by restricting the training set of known cancer driver genes to those genes known to be driver in a particular cancer type. However, for many cancer types, only few driver genes have been validated, creating a challenging situation for machine learning-based methods like LOTUS that rely on a training set of known driver genes to learn a scoring function. Since cancer driver genes of different cancer types are likely to have similar features, we propose instead to learn jointly cancer type-specific scoring functions by sharing information about known driver genes across cancer types, using the framework of multitask learning [33, 34]. Instead of starting from a list of known driver genes, we now start from a list of known (cancer gene, cancer type) pairs of the form {(g1, d1), …, (gN, dN)}, where a sample (gi, di) means that gene gi is a known cancer gene in disease di. Note that a given gene (and a given cancer type) may of course appear in several such pairs.
The extension of OC-SVM to the multitask setting is straightforwardly obtained by creating a kernel for (gene, disease) pairs of the form:
K p a i r ( ( g , d ) , ( g ′ , d ′ ) ) = K g e n e ( g , g ′ ) × K d i s e a s e ( d , d ′ ) ,
where Kgene is a kernel between genes such as that used in pan-cancer LOTUS and Kdisease is a kernel between cancer types described below. We then simply run the OC-SVM algorithm using Kpair as kernel and {(g1, d1), …, (gN, dN)} as training set, in order to learn a cancer type-specific scoring function of the form f(g, d) that estimates the probability that g is a cancer gene for cancer type d.
The choice of the disease kernel Kdisease influences how information is shared across cancer types. One extreme situation is to take the uniform kernel Kuniform(d, d′) = 1 for any d, d′. In that case, no distinction is made between diseases, and all known cancer driver genes are pooled together, recovering the pan-cancer setting (with the slight difference that genes may be counted several times in the training set if they appear in several diseases). Another extreme situation is to take the Dirac kernel KDirac(d, d′) = 1 if d = d′, 0 otherwise. In that case, no information is shared across cancer types, and the joint model over (gene, disease) pairs is equivalent to learning independently a model for each disease, as in the single-task approach.
In order to leverage the benefits of multitask learning and learn disease-specific models by sharing information across diseases, we consider instead the following two disease kernels:
When comparing LOTUS to TUSON and MUFFINN, we use a dataset of somatic mutations collected from COSMIC [14], TCGA (http://cancergenome.nih.gov/) and [18], that was used in [26]. This dataset contains a total of 1, 195, 223 mutations across 8, 207 patients affecting 18, 843 genes.
When comparing LOTUS to DiffMut and 20/20+, we use a dataset of somatic mutations borrowed from [28]. This dataset contains a total of 729, 205 mutations across 7, 916 patients affecting 19, 320 genes.
When comparing LOTUS to MutSigCV, we use an example dataset available on GenePattern. This dataset contains a total of 137, 343 mutations across 177 patients of lung squamous cell carcinoma affecting 16, 885 genes.
We obtained the PPI network from the HPRD database release 9 from April 13, 2010 [74]. It contains 39, 239 interactions among 7, 931 proteins. As for known pan-cancer driver genes, we consider three lists in our experiments: (i) the TUSON train set, proposed in [26], consists of two high confidence lists of 50 OGs and 50 TSGs extracted from CGC (release v71) based on several criteria, in particular excluding driver genes reported through translocations; (ii) the 20/20 train set, proposed in [28] to train the 20/20+ method, contains 53 OGs and 60 TSGs; finally, (iii) the CGCv86 train set consists of two broader lists that we extracted from CGC release v86 of the COSMIC database: we consider as OGs the genes annotated as “oncogene”, “oncogene, TSG”, “oncogene, fusion”, “oncogene, TSG, fusion”, and as TSGs the genes annotated as “TSG”, “oncogene, TSG”, “TSG, fusion”, “oncogene, TSG, fusion”. For cancer type-specific lists of driver genes, we only consider the CGCv86 train sets. We distinguished 30 diseases based on the available annotations describing patients in the mutation matrix, only merging “Kidney Chromophobe”, “Kidney Papillary Cell Carcinoma” and “Kidney Clear Cell Carcinoma” into “Kidney Cancer”, “DLBCL” and “Lymphoma B-Cell” into “Lymphoma B-Cell” and neglecting the unspecific “CARC”. The names of these diseases and their numbers of associated TSGs and OGs can be found in S2 Table. For each of the resulting diseases, 0 to 56 TSGs/OGs were known in CGCv86. We considered only diseases with at least 4 known TSGs or OGs available, in order to have enough learning data points to perform a two-fold cross-validation scheme, which led us to consider 27 diseases for TSG prediction and 27 for OG prediction.
Among LOTUS, TUSON, 20/20+, DiffMut, MUFFINN and MutSigCV, we can distinguish on the one side, the unsupervised methods MutSigCV, MUFFINN and DiffMut that score candidate genes independently of any training set of known drivers, and the supervised methods LOTUS, TUSON and 20/20+ that make predictions based on a training set of known driver genes.
In addition, all methods use gene descriptors that are calculated based on a mutation databases, and therefore, changing the mutation database will change the prediction performance.
In order to make fair comparison between LOTUS and the other five methods, we performed several experiments in which LOTUS is trained with the training set of the TUSON (respectively 20/20+) paper when compared to TUSON and MUFFINN (respectively 20/20+). In addition, in all experiments, the gene features calculated for LOTUS and MUFFINN were based on the same mutation databases as those used by the other methods in their respective papers.
Therefore, for a fair comparison between LOTUS, MUFFINN and TUSON, we use the mutation database available on the website of the authors along with their training sets of OGs and TSGs provided in [26]. We evaluate the performance of LOTUS on this dataset by 5-fold cross-validation repeated twice (see Methods). For TUSON, we use the prediction results available in [26] and evaluate the CE as the mean number of non-driver genes that are ranked before known driver genes of the TUSON train sets. Finally, we downloaded and re-ran MUFFINN.
For a fair comparison between LOTUS and 20/20+, we use the mutation database of 20/20+ and the training sets of OGs and TSGs provided by the authors on their website [28]. We evaluate the performance of LOTUS as above. However we note that the 20/20+ score itself is obtained by a bootstrap procedure similar to our cross-validation approach [28].
For a fair comparison between LOTUS and MutSigCV, we use the example mutation database available only for lung squamous cell tumours. Since MutSigCV does not use a train set of driver genes, we trained LOTUS with known OGs and TSGs available in CGCv86 for lung squamous cell tumours. MutSigCV provides a ranked list of genes that does not distinguish TSG and OG. Therefore, the CE is obtained by averaging the numbers of non-driver genes ranked before each driver genes in the train sets used for LOTUS.
Finally, for a fair comparison between LOTUS and DiffMut, we use the 20/20+ mutation dataset for both methods. LOTUS is trained with the 20/20+ training sets of OGs and TSGs. We run DiffMut using the latest version of the algorithm, and we compute the CE related to the 20/20+ training sets of OGs and TSGs.
To assess the performance of a driver gene prediction method on a given gold standard of known driver genes, we score all genes in the COSMIC database and measure how well the known driver genes are ranked. For that purpose, we plot the ROC curve, considering all known drivers as positive examples and all other genes in COSMIC as negative ones, and define the consistency error (CE) as
C E = # N × ( 1 - A U C ) ,
where # N is the number of negative genes, and AUC denotes the area under the ROC curve. In other words, CE measures the mean number of “non-driver” genes that the prediction method ranks higher than known driver genes. Hence, a perfect prediction method should have CE = 0, while a random predictor should have a CE near # N / 2.
To estimate the performance of a machine learning-based prediction method that estimates a scoring function from a training set of known driver genes, we use k-fold cross-validation for each given gold standard set of known driver genes. In k-fold cross-validation, the gold standard set is randomly split into k subsets of roughly equal sizes. Each subset is removed from the gold standard in turn, the prediction method is trained on the remaining k − 1 subsets, and its CE is estimated considering the subset left apart as positive examples, and all other genes of COSMIC not in the gold standard set as negative examples. A mean ROC curve and mean CE is then computed from the k resulting ROC curves. This computation is repeated several times to consider several possibly different partitions of the gold standard set.
Each version of LOTUS depends on a unique parameter, the regularization parameter C of the OC-SVM algorithm. Each time a LOTUS model is trained, its C parameter is optimized by 5-fold cross-validation on the training set, by picking the value in a grid of candidate values {2−5/2, 2−4/2, …, 25/2} that minimizes the mean CE over the folds.
We compare the performance of LOTUS to five other state-of-the-art methods: MutSigCV [21], which is a frequency-based method, TUSON [26] and 20/20+ [28] that combine frequency and functional information, MUFFINN [32] and DiffMut that analyses mutation profiles on genes.
MutSigCV searches driver genes among significantly mutated genes which adjusts for known covariates of mutation rates. The method estimates a background mutation rate for each gene and patient, based on the observed silent mutations in the gene and noncoding mutations in the surrounding regions. Incorporating mutational heterogeneity, MutSigCV eliminates implausible driver genes that are often predicted by simpler frequency-based models. For each gene, the mutational signal from the observed non-silent counts are compared to the mutational background. The output of the method is an ordered list of all considered genes as a function of a p-value that estimates how likely this gene is to be a driver gene.
TUSON uses gene features that encode frequency mutations and functional impact mutations. The underlying idea is that the proportion of mutation types observed in a given gene can be used to predict the likelihood of this gene to be a cancer driver. After having identified the most predicting parameters for OGs and TSGs based on a train set (called the TUSON train set in the present paper), TUSON uses a statistical model in which a p-value is derived for each gene that characterizes its potential as being an OG or a TSG, then scores all genes in the COSMIC database, to obtain two ranked lists of genes in increasing orders of p-values for OGs and TSGs.
The 20/20+ method encodes genes based on frequency and mutation types, and other biological information. It uses a train set of OGs and TSGs (called the 20/20 train set in the present paper) to train a random forest algorithm. Then, the random forest is used on the COSMIC database and the output of the method is again a list of genes ranked according to their predicted score to be a driver gene [28]. We did not implement this method, so we decided to evaluate its performance only on its original training set: the 20/20 dataset. Moreover, we applied the same method to compute the CE as for MutSigCV and TUSON, which should actually give an advantage to 20/20+, since it is harder to make predictions in a cross-validation loop using a smaller set of known driver genes.
DiffMut uses a dataset of somatic mutations and a dataset of healthy genomes, but no training sets of known driver genes. It compares the mutation profiles on a gene in the mutation dataset with the nucleotide variation profile in the healthy genomes, and computes for every gene a score that allows to rank all genes according to their potential as OG or TSG.
MUFFINN uses a dataset of somatic mutations and extracts the number of non-synonymous mutations per gene. Then, it computes a score (either DNmax or DNsum) and propagates these scores on a functional gene network (either HumanNet or STRING). The final scores are used to compute four different rankings for all genes. Among these four possibilities, we systematically used the version that yields the best result for MUFFINN,. Note however that, in practice, the user would not know which version should be preferred.
We implemented LOTUS and performed all experiments in R using in particular the kernlab package for OC-SVM [75]. The code and data to reproduce all experiments are available at https://github.com/LOTUSproject/LOTUS.
|
10.1371/journal.pgen.1008289 | Schnyder corneal dystrophy-associated UBIAD1 mutations cause corneal cholesterol accumulation by stabilizing HMG-CoA reductase | Schnyder corneal dystrophy (SCD) is a rare genetic eye disease characterized by corneal opacification resulted from deposition of excess free cholesterol. UbiA prenyltransferase domain-containing protein-1 (UBIAD1) is an enzyme catalyzing biosynthesis of coenzyme Q10 and vitamin K2. More than 20 UBIAD1 mutations have been found to associate with human SCD. How these mutants contribute to SCD development is not fully understood. Here, we identified HMGCR as a binding partner of UBIAD1 using mass spectrometry. In contrast to the Golgi localization of wild-type UBIAD1, SCD-associated mutants mainly resided in the endoplasmic reticulum (ER) and competed with Insig-1 for HMGCR binding, thereby preventing HMGCR from degradation and increasing cholesterol biosynthesis. The heterozygous Ubiad1 G184R knock-in (Ubiad1G184R/+) mice expressed elevated levels of HMGCR protein in various tissues. The aged Ubiad1G184R/+ mice exhibited corneal opacification and free cholesterol accumulation, phenocopying clinical manifestations of SCD patients. In summary, these results demonstrate that SCD-associated mutations of UBIAD1 impair its ER-to-Golgi transportation and enhance its interaction with HMGCR. The stabilization of HMGCR by UBIAD1 increases cholesterol biosynthesis and eventually causes cholesterol accumulation in the cornea.
| Schnyder corneal dystrophy (SCD) is a rare genetic eye disease caused by deposition of free cholesterol in the cornea. It is closely correlated with mutations in the UbiA prenyltransferase domain-containing protein-1 (UBIAD1) gene, which encodes an enzyme catalyzing biosynthesis of coenzyme Q10 and vitamin K2. The underlying mechanism by which UBIAD1 mutations result in SCD development is unclear. Here, we report that SCD-associated mutations trap UBIAD1 in the ER and block Insig-1 mediated HMGCR degradation. We also generated a heterozygous mouse model (Ubiad1G184R/+) that mimics human SCD. We conclude that SCD-associated UBIAD1 mutations decrease HMGCR degradation and subsequently increase cholesterol biosynthesis in the cornea.
| Schnyder corneal dystrophy (SCD) is a rare autosomal dominant genetic eye disease [1]. It is characterized by free cholesterol accumulation in the cornea that causes progressive corneal opacification with aging [1] [2]. Genetics studies have linked SCD to mutations in UbiA prenyltransferase domain-containing protein-1 (UBIAD1), also known as transitional epithelial response gene 1 (TERE1) [3–5]. Until now, 25 missense mutations altering 21 amino acids, including N102S and G186R (equivalent to N100S and G184R in mouse Ubiad1, respectively), have been identified in about 50 SCD families [6–8]. However, the causal relationship between UBIAD1 mutations and SCD development has not been proved until very recently [9].
The UBIAD1 protein belongs to the UbiA superfamily of intramembrane aromatic prenyltransferases, which catalyze the biosynthesis of a variety of lipophilic molecules such as ubiquinones, vitamin K, vitamin E, chlorophylls, hemes and archaeal tetraether lipids [10]. UBIAD1 has been identified as a vitamin K2 biosynthesis enzyme in humans and mice, and is essential for mouse embryonic development [11, 12]. In Drosophila, the UBIAD1 homolog was reported to generate vitamin K2 that functions as an electron carrier for sustaining mitochondrial function [13]. In zebrafish, UBIAD1 was proposed to synthesize ubiquinone coenzyme Q10 that protects against oxidative damages through regulating nitric oxide activity, thereby maintaining vascular endothelial cell survival [14]. However, how UBIAD1 mutations cause cholesterol accumulation in the cornea is not fully understood.
The endoplasmic reticulum (ER)-localized 3-hydroxy-3-methyglutaryl coenzyme A reductase (HMG-CoA reductase, HMGCR) is a rate-limiting enzyme of the cholesterol biosynthetic pathway catalyzing the synthesis of mevalonate [15]. As an important intermediate, mevalonate gives rise to not only cholesterol, but also nonsterol isoprenoids such as farnesyl pyrophosphate and geranylgeranyl pyrophosphate (GGPP) that further generates ubiquinone, dolichol and hemes [15]. Interestingly, these mevalonate-derived molecules are also the final products of UBIAD1 superfamily members, suggesting a potential link between HMGCR and UBIAD1. Theses sterols and nonsterol isoprenoids coordinate to accelerate degradation of HMGCR to prevent over-accumulation of cholesterol [16]. Accumulating sterols induce HMGCR binding to ER-anchored Insig-1 and Insig-2, which bring ubiquitin ligases gp78, TRC8 and RNF145 together with other cofactors to ubiquitinate HMGCR, resulting in its degradation in proteasome [17–22]. Elucidating the molecular pathway of HMGCR degradation is of potential clinic significance. Statins as competitive inhibitors of HMGCR can decrease the synthesis of sterols and nonsterol isoprenoids and dramatically increase HMGCR in the liver [23, 24]. The lanosterol analog HMG499 (also named Cmpd 81) can potentiate the cholesterol-lowering effect of statins through inducing HMGCR degradation [25].
Recently, Ubiad1 was found to be a HMGCR-binding protein through proximity-dependent biotinylation using HMGCR as the bait [26]. The homozygous Ubiad1 N100S knock-in (Ubiad1N100S/N100S) mice exhibit corneal opacification and HMGCR accumulation in tissues [9]. In this study, we identified HMGCR as a UBIAD1-associated protein through UBIAD1 immunoprecipitation coupled with mass spectrometry. We demonstrated that UBIAD1 competed with Insig-1 for binding HMGCR, preventing the latter from ubiquitination and degradation. All known SCD-associated UBIAD1 mutants localized in the ER and stabilized HMGCR protein. More importantly, we generated a different SCD-associated knock-in mouse line carrying Ubiad1 G184R mutation. The Ubiad1G184R/+ mice exhibited excess HMGCR protein in tissues and striking corneal opacifications with free cholesterol deposition. These phenotypes recapitulate clinical manifestations of human SCD patients, suggesting that Ubiad1G184R/+ mouse is an ideal model to study human SCD.
To explore the underlying connections between UBIAD1 and cholesterol metabolism, we performed a tandem affinity purification (TAP) coupled to mass spectrometry to identify UBIAD1-associated proteins, using HEK-293 cells stably expressing human wild-type (WT) or G186R mutant form of UBIAD1 fused with a TAP tag at the C terminus (S1A Fig). The G186R mutation is a mutation found in an early-onset SCD family [27]. Besides, according to the determined structure of archaeal UbiA, the G186 residue was proposed to locate in the surface-exposed loop, and the G186R mutation may affect the interactions with other proteins [10, 28]. We found three proteins among the top of the list: HMGCR, VCP/p97 and SEL1L (S1B and S1C Fig, S1 Table). HMGCR is the rate-limiting enzyme converting HMG-CoA to mevalonate in the cholesterol biosynthetic pathway [15]. VCP/p97 and SEL1L have been known to be involved in the degradation of HMGCR and other ER proteins [29–31].
Therefore, we focused on the effects of UBIAD1 on sterol-induced degradation of HMGCR protein. The cells stably expressing WT UBIAD1 or G186R mutant were treated with increasing concentrations of 25-hydroxycholesterol (25-HC) for 5 h. The endogenous HMGCR was degraded in a concentration-dependent manner in cells stably expressing UBIAD1 (WT). However, 25-HC failed to induce HMGCR degradation in the UBIAD1 (G186R)-expressing stable cells (Fig 1A). In addition, the UBIAD1 (G186R) mutation had no obvious effect on SREBP-2 processing (Fig 1A). We next validated these findings using co-transfection experiments. Co-expression of HMGCR with Insig-1 conferred sterol-regulated degradation of endogenous HMGCR, as Insig-1 is a rate-limiting co-factor for multiple E3s including gp78, TRC8 and RNF145 [18–22] (Fig 1B, lanes 1–2). The WT form of mouse Ubiad1 had little effect on HMGCR degradation, whereas Ubiad1 (G184R) almost completely blocked the degradation (Fig 1B, lanes 3–6). The K89 and K248 are two ubiquitination sites of HMGCR [16], and their mutations should abolish sterol-induced degradation of HMGCR. Neither Ubiad1 (WT) nor Ubiad1 (G184R) increased the amount of HMGCR (K89R, K248R) (Fig 1B, lanes 7–12). The amount of Ubiad1 (G184R) was about 2-fold that of Ubiad1 (WT) when equivalent amounts of plasmids were used.
We next analyzed the effect of Ubiad1 on HMGCR ubiquitination. 25-HC triggered pronounced ubiquitination of immunoprecipitated HMGCR in the presence of proteasome inhibitor MG-132, which, however, was markedly reduced by WT Ubiad1 (Fig 1C, second panel, compare lane 2 and 4). The Ubiad1 (G184R) further inhibited the ubiquitination of HMGCR (Fig 1C, second panel, compare lane 4 and 6). The total amount of HMGCR from input was more abundant in Ubiad1 (G184R)-expressing cells than those expressing Ubiad1 (WT) (Fig 1C, third panel, lanes 3–6), consistent with results in Fig 1A and 1B.
We then used co-immunoprecipitation experiments to address potential interaction between HMGCR, Insig-1 and Ubiad1. The results showed that HMGCR pulled down more amounts of Ubiad1 (G184R) than Ubiad1 (WT) (Fig 1D). Meanwhile, Insig-1 association with HMGCR was reduced (Fig 1D). Together, these results suggest that SCD-associated mutant Ubiad1 competes with Insigs to bind HMGCR, thereby blocking Insig-mediated ubiquitination and degradation of HMGCR.
We then analyzed whether SCD-associated UBIAD1 (G186R) had any effect on cellular cholesterol level. Using the colorimetric method, we measured the amount of total cholesterol in cells stably expressing WT or G186R form of UBIAD1. UBIAD1 (G186R)-expressing cells indeed had more cholesterol than control (Fig 2A). However, no difference in the amount of nonesterified fatty acid (NEFA) was detected between these two cell lines (Fig 2B). We next sought to determine whether the G186R mutation increased de novo synthesis of cholesterol by using [14C]-acetate to label newly synthesized cholesterol and fatty acid. [14C]-cholesterol was markedly increased in UBIAD1 (G186R)-expressing cells (Fig 2C), whereas [14C]-fatty acid remained comparable between WT- and UBIAD1 (G186R)-expressing cells (Fig 2D). These results together suggest that the UBIAD1 (G186R) mutation enhances synthesis of cholesterol.
Human UBIAD1 contains 338 amino acids with 8 predicated transmembrane helices, and 21 SCD-associated UBIAD1 nucleotide mutations in the coding sequence that altered amino acids at 19 positions are shown in Fig 3A. Protein sequence analysis showed that amino acids that are mutated in SCD are evolutionarily conserved from fly to human (S2 Fig). As UBIAD1 (G186R) dramatically stabilized HMGCR and increased cellular cholesterol level (Fig 1, Fig 2), we next examined the effect of other SCD-associated UBIAD1 mutations on sterol-induced degradation of HMGCR. Each of the 21 SCD-associated missense mutations of UBIAD1 were co-transfected with HMGCR and Insig-1 expression plasmids individually followed by 25-HC treatment for 5 h. HMGCR protein was reduced by 80% in cells expressing the WT form of human UBIAD1 after receiving 25-HC treatment. However, 25-HC-induced HMGCR degradation was only reduced by 10% to 50% in cells expressing any of the 21 UBIAD1 mutations (Fig 3B–3F, top panel), indicating that HMGCR protein can be stabilized by SCD mutants of UBIAD1. Interestingly, the protein levels of SCD-associated UBIAD1 mutants were all substantially increased without being affected by 25-HC, even though the cells were transfected with the same amount of plasmids (100 ng per dish) (Fig 3B–3F, second panel). The protein levels of UBIAD1 harboring SCD-associated mutations were 1.9 to 9.5-fold more than WT UBIAD1 (Fig 3B–3F, second panel), while the Insig-1 protein had no changes (Fig 3B–3F, second panel). These results suggest that the SCD mutations of UBIAD1 render HMGCR proteins more resistance to sterol-induced degradation.
We next examined the cellular localization of WT UBIAD1 and UBIAD1 harboring SCD mutations in CHO-K1 cells. Immunofluorescence experiments revealed that the ectopically expressed WT UBIAD1 preferentially co-localized with the Golgi marker GM130 (Fig 4). However, the SCD-associated UBIAD1 mutants had a diffused distribution and co-localized with the ER marker calnexin (Fig 4 and S3 Fig).
We further analyzed the distribution of WT and G186R UBIAD1 under different conditions. S4 Fig showed that WT UBIAD1 presented in Golgi in FCS condition and in ER in sterol-depletion condition. Nonsterol isoprenoid geranylgeraniol (GGOH), but not cholesterol or 25-HC, relocated WT UBIAD1 to Golgi (S4A and S4C Fig). However, the G186R mutant primarily located in the ER under different conditions, and seems to be trafficking-deficient (S4B and S4D Fig). We next analyzed the effects of 25-HC and GGOH on the association between HMGCR and UBIAD1. Results of S5A Fig showed that GGOH could further degrade HMGCR with 25-HC in WT Ubiad1 expressing cells, and had minimal effect in G184R mutant cells (S5A Fig). Co-immunoprecipitation results in S5B Fig showed that 25-HC stimulated HMGCR to bind more WT Ubiad1, and GGOH reduced the interaction between HMGCR and WT Ubiad1. However, the association of HMGCR and G184R mutated Ubiad1 was stronger than WT Ubiad1, and did not response to 25-HC and GGOH treatments (S5B Fig). Therefore, the SCD-associated mutations may impair ER-to-Golgi transportation of UBIAD1 and enhance UBIAD1-HMGCR interaction to prevent Insig-mediated degradation of HMGCR.
Although many missense mutations of UBIAD1 have been found in SCD patients, the causal relationship between these two remains to be proved. To further investigate the consequence of SCD-associated mutation of UBIAD1 in vivo, we generated Ubiad1 G184R (corresponding to G186R in human) knock-in mice using a knockout-first conditional ready strategy as shown in Fig 5A. The targeted allele was knocked out and conditional ready, and the mice were first crossed with Flp recombinase-expressing strain to generate the floxed allele that expressed WT Ubiad1. The mice were then crossed with EIIA-Cre recombinase transgenic mice to get the whole-body knock-in mice (Fig 5A). The WT mice had a single band of 400 bp, while the heterozygous knock-in mice (Ubiad1G184R/+) exhibited bands at both 400 bp and 500 bp (Fig 5B). Further sequencing of these two bands confirmed that the mutated band indeed carried the GGA-to-AGA mutation (Fig 5C).
No homozygous knock-in (Ubiad1G184R/G184R) mice were generated when the heterozygotes were intercrossed (Fig 5D). These results are not surprising as Ubiad1 knockout mice defective in vitamin K2 synthesis are embryonic lethal as well [12]. The WT and heterozygous knock-in mice (Ubiad1G184R/+) were born at an expected Mendelian ratio (34% vs 66%) (Fig 5D). Ubiad1G184R/+ knock-in mice appeared indistinguishable from WT littermates, and both had similar body weights even at the average age of 105-week-old (Fig 5E). Given that systemic hypercholesterolemia has been reported in some but not all SCD patients [1], we next sought to evaluate cholesterol levels of WT and Ubiad1G184R/+ mice. The serum levels of total cholesterol (TC), high-density-lipoprotein (HDL) and low-density-lipoprotein (LDL) cholesterol in Ubiad1G184R/+ knock-in mice were similar to those of WT mice (Fig 5F, 5G and 5H). The serum triglyceride (TG) level of Ubiad1G184R/+ mice was slightly, but not significantly, decreased relative to WT mice (Fig 5I). Collectively, these results suggest that the G184R mutation of Ubiad1 does not affect systemic cholesterol and triglyceride levels.
Since all SCD-associated point mutations including G186R impaired sterol-induced degradation of HMGCR and caused accumulation of HMGCR (Fig 1, Fig 3), we prepared mouse embryonic fibroblast (MEF) cells from Ubiad1G184R/+ mice and their WT littermates and analyzed HMGCR degradation. Compared with WT MEF cells, endogenous HMGCR in Ubiad1G184R/+ MEFs was partially resistant to ubiquitination and degradation induced by 25-HC (Fig 6A, S6 Fig), similar to the overexpression results (Fig 1A–1C). We next measured whether the G184R mutation would alter HMGCR protein levels in different mouse tissues. Strikingly, the amount of HMGCR was dramatically higher in the liver (2.5-fold), pancreas (12.1-fold), lung (4.5-fold), spleen (4.2-fold) and cornea (5.7-fold) of Ubiad1G184R/+ mice than WT mice (Fig 6B–6F). Together, these results indicate that Ubiad1 (G184R) protects HMGCR from degradation and increases HMGCR in various tissues.
Next, we examined mouse eyes using stereomicroscope. Prominent signs of corneal opacifications were found in 64% (21/33) of both male (11/17) and female (10/16) Ubiad1G184R/+ mice at 102-to-108 weeks of age (about 2 years, equivalent to the human age of 70) (Fig 7A, bottom panel). There was no difference between male and female mice. These corneal opacifications were haze-like and quite similar to human SCD (Fig 7A, bottom panel). As controls, none of the 17 aged WT littermates showed corneal opacification (Fig 7A, top panel). To further characterize this corneal opacification, the corneal section from these aged mice were stained with Filipin, a specific antibiotic binding to free cholesterol [32]. Free cholesterol in the WT cornea mainly localized to the epithelial cells and sporadically to the stromal cells as well (Fig 7B, top panel). However, in Ubiad1G184R/+ cornea the Filipin signals infiltrated throughout the anterior stroma underneath the epithelium in puncta or patches (Fig 7B, bottom panel), which correspond to the cholesterol crystals observed in the corneas of human SCD patients. Biochemical analysis showed that Ubiad1G184R/+ mice had higher levels of total and free cholesterol in the cornea than WT littermates, although both showed similar corneal TG levels (Fig 7C–7E). In contrast, the levels of total cholesterol, free cholesterol and TG levels in the liver, pancreas, lung and spleen were similar between WT and Ubiad1G184R/+ mice (S7A–S7L Fig). Collectively, these results demonstrate that the G184R mutation of Ubiad1 specifically cause free cholesterol accumulation in the anterior stroma of cornea, phenocopying human SCD.
Based on the above findings, we propose a working model depicting how UBIAD1 mutations cause cholesterol accumulation in the cornea (Fig 8). Under normal conditions in which the WT form of UBIAD1 mainly localizes in the Golgi, an elevation in sterols triggers HMGCR binding to Insig-1 or Insig-2, which recruits E3 ubiquitin ligases including gp78, TRC8 and RNF145 for ubiquitination and proteasomal degradation of HMGCR [18–22]. SCD-associated mutations of UBIAD1 impair its transportation from the ER to Golgi, resulting in a more stable protein in the ER that competes with Insig-1 for binding to HMGCR. As a consequence, sterol-induced ubiquitination and degradation of HMGCR is blocked. The increased cholesterol biosynthesis eventually causes cholesterol accumulation in the cornea in aged mice and humans.
The association of UBIAD1 with HMGCR was recently reported by Debose-Boyd and coworkers using HMGCR as a bait [26]. They showed that the SCD-associated mutants of UBIAD1 were sequestered in the ER and protected HMGCR from degradation, leading to the accumulation of HMGCR and cholesterol [26, 33]. They also found cholesterol accumulation in the cornea of aged Ubiad1N100S/N100S mice, which is another SCD model [9]. Our results and theirs are largely consistent and both support a role of HMGCR-UBIAD1 interaction in HMGCR stabilization. However, we employed an unbiased mass spectrometry analysis of immunoprecipitated UBIAD1 and constructed a different knock-in mouse line (Ubiad1G184R/+). Considering that all SCD patients are heterozygotes of UBIAD1 mutation, our Ubiad1G184R/+ mouse model may mimic human situation more closely. Indeed, Ubiad1G184R/+ mice displayed HMGCR accumulation in multiple tissues and elderly ones had free cholesterol deposition in the cornea.
Notably, corneal opacification and free cholesterol accumulation, albeit prominent in two-year-old Ubiad1G184R/+ heterozygote mice, were barely detected in the 3-month-old mice (data not shown). Consistently, SCD patients carrying heterozygous UBIAD1 mutations display slow progression of corneal opacification with aging. In addition, the homozygous Ubiad1G184R/G184R knock-in mice are embryonic lethal, while Ubiad1N100S/N100S mice are grossly normal. Because Ubiad1 knockout mice fail to survive through development owing to the defects in vitamin K2 synthesis [12], it is reasonable to speculate that the Ubiad1 G184R mutation may affect vitamin K2 synthesis more profoundly than Ubiad1 N100S.
Another intriguing phenomenon is that no obvious differences were detected in the serum and tissue levels of cholesterol and triglyceride between the WT and Ubiad1G184R/+ mice (Fig 5, Fig 7, S7 Fig), though HMGCR protein levels in all examined tissues were increased (Fig 6). Cholesterol accumulation was only detected in the cornea (Fig 7). Such phenomena are likely attributed to the unique anatomic structure of cornea, which lacks blood-vascular system and is separated from the systemic circulation [34]. Other tissues such as the liver, pancreas, lung and spleen, can intensively exchange lipids with the systemic circulation. In addition, compensatory mechanisms, such as the cholesterol esterification and efflux, may balance the cholesterol level in these tissues. According to the BioGPS expression database, the expression of acyl-CoA: cholesterol acyltransferases (ACAT)-1 and ACAT-2, which convert free cholesterol to cholesteryl ester for storage or secretion as lipoproteins [35], and ABCA1, which is an essential transporter for cholesterol efflux to HDL [36], is very low in the cornea [37]. Thus, the cornea cannot efficiently remove excess cholesterol by converting to cholesteryl ester or pumping out of the cell, leading to the accumulation of free cholesterol once HMGCR is stabilized by UBIAD1 mutations. Interestingly, familial lecithin-cholesterol acyltransferase (LCAT) deficiency and fish eye disease caused by partial loss-of-function of LCAT also exhibit evident corneal opacification due to free cholesterol accumulation [38]. These diseases highlight the importance of cholesterol homeostasis in maintaining corneal function.
Currently, corneal transplant surgery is the only way to restore vision in SCD patients [1]. It is urgent to develop other treatments in the future. As corneal overproduction of cholesterol is caused by HMGCR accumulation, it would be interesting to test whether local application of statins, the well-known inhibitors of HMGCR and are widely used for lowering cholesterol [39], in the form of eyedrops is effective for reducing corneal opacification. In addition, 2-hydroxypropyl-β-cyclodextrin (HPβCD) has an excellent ability to solubilize cholesterol and has been used to treat Niemann-Pick type C (NPC) disease caused by lysosomal cholesterol accumulation [40]. The local application of HPβCD would be another promising strategy to reduce corneal cholesterol. Our recently identified HMG499 (also named Cmpd 81) might also be used to treat SCD since HMG499 is a potent HMGCR degrader [25] and SCD is caused by HMGCR stabilization.
Collectively, we have found that the SCD-associated mutants of UBIAD1 bind and stabilize HMGCR, thereby increasing cellular cholesterol level. We have also generated and characterized a mouse model (Ubiad1G184R/+) for SCD disease that will be valuable for studying the underlying mechanism of SCD and developing therapeutic strategies as well.
All procedures and care of animals were carried out in accordance with the guidelines and protocols approved by the Institutional Animal Care and Use Committee at the Wuhan University under protocol number WDSKY0201408.
We obtained lovastatin (PHR1285), mevalonate (41288), 25-hydroxycholesterol (H1015), MG-132 (M8699), Filipin (F9765), geranylgeraniol (G3278), protease inhibitor cocktail (P8340), N-Ethylmaleimide (E3876), and paraformaldehyde (PFA) (P6148) from Sigma; [14C]-acetic acid sodium salt (NEC084H001) from Perkin Elmer; and FuGENE HD transfection reagent (E2312) from Promega; G418 (345810), digitonin (300410), henylmethylsulfonyl fluoride (PMSF) (52332), leupeptin (108975), pepstatin A (516481) and ALLN (208719) from Merck; Hoechst 33342 (H1399) from Invitrogen. Lipoprotein-deficient serum (LPDS) [41] and delipidated-fetal calf serum (FCS) [42] was prepared from FCS (S1580, Biowest) by ultracentrifugation in our laboratory [25].
Primary antibodies used for immunoblotting were as follows: mouse monoclonal anti-T7 (69522, Merck, 1 μg/ml), mouse monoclonal anti-Flag (F3165, Sigma-Aldrich, 1:1000), mouse monoclonal anti-Myc IgG-9E10 (CRL-1729, ATCC, 1 μg/ml), mouse monoclonal anti-ubiquitin IgG-P4D1(sc-8017, Santa Cruz Biotechnology, 1:1000), mouse monoclonal anti-hamster HMGCR IgG-A9 (CRL-1811, ATCC, 2 μg/ml), mouse monoclonal anti-clathrin heavy chain (610500, BD Biosciences, 1:1000), mouse monoclonal anti-Actin (A3853, Sigma, 1:5000), polyclonal anti-HMGCR antibody (H2) was raised against a C-terminal sequence (aa410-aa888) of human HMGCR in our laboratory [43]. Horseradish peroxidase-conjugated donkey anti-mouse (715-005-151, 1:5000) and anti-rabbit (711-005-152, 1:5000) secondary antibodies were obtained from Jackson ImmunoResearch Laboratories.
Primary antibodies used for immunofluorescence staining were as follows: rabbit polyclonal anti-GM130 (G7295, Sigma, 1:300), rabbit polyclonal to calnexin (ab22595, Abcam, 1:300). Alexa Fluor 488-labeled donkey anti-mouse IgG (A-21202, 1:500) and Alexa Fluor 555-labeled donkey anti-rabbit IgG (A-31572, 1:500) secondary antibodies were obtained from Invitrogen.
The following plasmids pCMV-Insig-1-Myc, pCMV-HMGCR-T7, pCMV-HMGCR-T7-K89R/K248R, pEF1a-HA-Ubiquitin were constructed in our laboratory [25]. The coding regions of UBIAD1 was amplified from human cell cDNA or mouse liver cDNA using standard PCR and cloned into cloned into pcDNA3-C-5xMyc vector. The plasmids encoding the variants of UBIAD1 were generated by site-directed mutagenesis based on full-length human or mouse UBIAD1. All the constructs were verified by DNA sequencing. The primer sequences are listed in S2 Table.
HEK-293 cells were grown in monolayer at 37°C in 5% CO2 in medium containing 10% FCS, Dulbecco’s modified Eagle’s medium (DMEM), 100 units/ml penicillin and 100 μg/ml streptomycin sulfate. CHO-K1 cells were grown in monolayer at 37°C in 5% CO2 in medium containing 5% FCS, Ham’s F-12 and DMEM (1:1), 100 units/ml penicillin and 100 μg/ml streptomycin sulfate.
On day 0, CHO-K1 cells were set up for experiments at 4×105 cells per 60-mm dish. On day 2, the cells were transfected with the indicated plasmids by using FuGENE HD reagent. The total amount of DNA in each transfection was adjusted to 2 μg per dish by addition of pcDNA3 mock vector.
For whole cell lysate, the cells were harvested and suspended in 120 μl of RIPA buffer (50 mM Tris-HCl, pH 8.0, 150 mM NaCl, 0.1% SDS, 1.5% NP40, 0.5% deoxycholate, 2 mM MgCl2) containing protease inhibitors (protease inhibitor cocktail 1:500, 10 μg/ml leupeptin, 5 μg/ml pepstatin A, 25 μg/ml ALLN, 1 mM PMSF and 10 μM MG-132). Protein concentrations of the extracts were determined with Pierce BCA protein assay kit (ThermoFisher, 23225), then mixed with 4×SDS loading buffer (150 mM Tris-HCl, pH 6.8, 12% SDS, 30% glycerol, 0.02% bromophenol blue, 6% β-mercaptoethanol). Proteins were separated with SDS-PAGE, transferred to nitrocellulose filters (GE Healthcare). Immunoblots were blocked by 5% non-fat milk/TBST and probed with indicated primary and HRP-conjugated secondary antibodies. Images were captured with Amersham Imager 680 (GE Healthcare), and intensities of each band were quantified by Image-Pro Plus 6 software (Media Cybernetics).
About 50 mg tissues excised from mouse were suspended in 500 μl of buffer A (10 mM HEPES/KOH, pH 7.6, 1.5 mM MgCl2, 10 mM KCl, 5 mM EDTA,5 mM EGTA, 250 mM sucrose) containing protein protease inhibitors, and homogenized by Precellys 24 (Bertin). The homogenized suspensions were then passed through a #7 needle for 30 times and centrifuged at 1000 g at 4°C for 7 min. The supernatant from the 1,000 g spin was centrifuged at 1 x 105 g for 30 min at 4°C. The pellets were the membrane fractions and were dissolved in 0.1 ml of SDS lysis buffer (10 mM Tris-HCl, pH6.8, 100 mM NaCl, 1% SDS, 1 mM EDTA, 1 mM EGTA).
HEK-293 cells were transfected with pCMV-Hs-UBIAD1-WT-TAP or pCMV-Hs-UBIAD1-G186R-TAP. One day later, cells were changed to medium containing 800 μg/ml G418. Fresh medium was exchanged every 2–3 days until colonies formed after about 2 weeks. Individual colonies were isolated with cloning cylinders, and the expression levels of UBIAD1 protein were verified by immunoblot.
HEK-293/Hs-UBIAD1-WT-TAP and HEK-293/Hs-UBIAD1-G186R-TAP stable cells were set up in 100-mm dish. Cells were harvested and lysed in immunoprecipitation (IP) buffer (1% digitonin, PBS, 5 mM EDTA, 5 mM EGTA) with protein protease inhibitors. Then cells were needled 15 times and centrifuged at 13,200 rpm for 20 min. The supernatant was pre-cleared with protein A/G agarose (sc-2003, Santa Cruz Biotechnology), then immunoprecipitated with rabbit IgG coupled agarose (A2909, Sigma) at 4°C for 2 hr. Agarose-captured proteins were released with TEV Protease (P8112S, New England Biolabs) in 500 μl IP buffer at RT for 1 hr. The released proteins were immunoprecipitated with anti-Flag M2 agarose beads (A2220, Sigma) at 4°C for 2 hr, and eluted with 3x Flag peptide at 4°C for 30 min. The eluted fractions were resolved with SDS-PAGE and stained with Coomassie blue. The identifies of proteins from eluted fractions were determined by tandem mass spectrometry.
The cells were harvested and suspended in 600 μl of IP buffer (1% digitonin, PBS, 5 mM EDTA, 5 mM EGTA, and 10 mM N-Ethylmaleimide) containing protein protease inhibitors, and passed through #7 needle 15 times. The cell lysates were centrifuged at 13,200 rpm at 4°C for 10 min, immunoprecipitated with anti-T7 antibody coupled agaroses (69026, Merck), and eluted with 2x SDS loading buffer. Whole cell lysates and pellets were subjected to SDS-PAGE and immunoblotted with indicated antibodies.
Cells were fixed with 4% PFA in PBS, permeabilized with 0.2% Triton X-100 (T8787, Sigma) for 8 minutes, blocked with 2% BSA for 1 hr and incubated with indicated primary antibodies for 1 hr at room temperature. Fluorescence-labeled secondary antibodies were used at concentration of 4 μg/ml in PBS containing 0.2% BSA for 45 minutes. Nuclei was stained with Hoechst 33342. Immunofluorescence images were captured by Leica TCS SP2 confocal microscope. Images were acquired at identical laser output, gain, and offset [44].
The protein sequences of UBIAD1 from difference species were aligned using ClustalW algorithm in MEGA X software [45]. The accession numbers of these protein sequences were used as followings: human (NP_037451.1), chimpanzee (JAA33188.1), rhesus (NP_001247708.1), tree shrew (XP_006145395.1), sheep (XP_004013772.1), horse (XP_008520311.1), elephant (XP_003413515.1), dog (XP_544571.1), rabbit (ETE69346.1), mouse (NP_082149.1), chicken (NP_001026050.1), snake (ETE69346.1), frog (NP_001016538.1), zebrafish (NP_001186655.1), sea urchin (XP_011664743.1) and fly (NP_523581.1).
Corneas were fixed by 4% PFA and sectioned with frozen cryostat (Leica CM3050 S) at 7 μm. Frozen sectioned slides were washed twice with PBS, stained with 50 μg/ml filipin in 10% FBS/PBS for 1 hr at room temperature, washed three times with PBS and mounted. Images were captured by fluorescence microscopy (Zeiss Axio Imager Z2) using a UV filter set, and the intensity of mercury lamp was turned to 10% of the maximal strength. Images were acquired at identical output, gain, and offset [46].
Cells were incubated in medium containing 10% delipidated-FCS, 1 μM lovastatin and 50 μM mevalonate for 16 hr. After depletion, cells were washed to remove lovastatin, and change to medium with 10% delipidated -FCS for 3 hr. Then [14C]-acetate (36 μCi/100-mm dish) were added and cells were treated for additional 2 hr. Cells were washed twice with PBS, dissolved by 0.5 ml 0.1 N NaOH, and saponified with ethanol and 75% potassium hydroxide for 2 hr. Then the nonpolar lipids (cholesterol) were extracted in petroleum ether and evaporated to dryness with N2. Following addition of concentrated HCl, polar lipids (fatty acids) were extracted in petroleum ether and evaporated to dryness with N2. The lipids were resolved by thin-layer chromatography (1.05582.0001, Merck). Radioactive signals were visualized with phosphoimager, and images were captured with Typhoon FLA 9000 (GE Healthcare).
The knockout first conditional ready targeting vector, containing a gene-trap cassette, was electroporated into ES cell line from C57BL/6J mice. Positive ES clones were verified by PCR and Southern blot, and three positive ES clones were injected to BALB/c blastocysts to generate chimeric mice. Male mice with high percentage chimeras were bred with female C57BL/6J mice (Shanghai Laboratory Animal Company, China) to get the heterozygous Ubiad1 knockout mice. To get a conditional knockin allele, these heterozygous knockout mice were bred with Flp recombinase deleter transgene (Jackson Laboratory, USA) to remove the Frt-flanked cassette. Subsequently, these mice were crossed with EIIA-Cre recombinase transgenic mice (Jackson Laboratory, USA) to remove Loxp-flanked cassette, thus getting the whole tissue Ubiad1 G184R knockin mice. The genotypes were identified by PCR using primers P1 (GCAAGCTGTATTTTGCCTTG), and P2 (CGAAAGTGATGAGGATGACGAGGT). The male and female Ubiad1G184R/+ were intercrossed to get Ubiad1+/+ and Ubiad1G184R/+ mice littermates for experiments. All mice were maintained on a 12-h light/dark schedule, and fed ad libitum access to water and standard chow diet (Shanghai Laboratory Animal Company, China).
Blood were collected from anesthetized mice, and serum was prepared from blood by centrifuging at 1500 × g for 10 min. Tissues were excised and weighted, then homogenized in chloroform/methanol (2:1) with Precellys 24 (Bertin). The lipids were extracted and dried under N2, and dissolved in ethanol. Total cholesterol and free cholesterol levels in serum and tissues were determined with Amplex Red Cholesterol Assay Kit (A12216, Invitrogen). HDL cholesterol, LDL cholesterol and triglyceride levels in the serum and liver were measured with HDL cholesterol, LDL cholesterol and triglyceride Assay Kit (Shanghai Kehua Bio-engineering, China), respectively.
Female Ubiad1G184R/+ mice were mated with male Ubiad1G184R/+ mice. Pregnant mice on the 13-day of conception were sacrificed by brief exposure to CO2. The uterine horns were excised into Petri dish and all embryos were carefully dissected. Each embryo was transferred into a new dish; the heads were used for genotyping. The rest of embryo was minced thoroughly using sharp Iris scissors, and digested with trypsin/EDTA for 20 minutes. Trypsin was neutralized, and the MEFs were collected by centrifuging. MEFs were cultured in medium with 10% FCS/DMEM, 100 units/ml penicillin and 100 μg/ml streptomycin sulfate.
Mice were first anaesthetized with 1% pentobarbital sodium in PBS (10 μl/g), then corneal opacifications were observed and images were captured with stereoscopic microscope (Olympus SZX16). The corneas were excised from eyes and carefully removed other tissues under stereomicroscope, then used for protein and lipid analyses.
The statistical analyses were carried out using GraphPad Prism 6 software. Data were presented as means ± SD and analyzed by unpaired two-tailed Student’s t-test. Statistical significance was set at p < 0.05.
|
10.1371/journal.pgen.1004066 | ComEA Is Essential for the Transfer of External DNA into the Periplasm in Naturally Transformable Vibrio cholerae Cells | The DNA uptake of naturally competent bacteria has been attributed to the action of DNA uptake machineries resembling type IV pilus complexes. However, the protein(s) for pulling the DNA across the outer membrane of Gram-negative bacteria remain speculative. Here we show that the competence protein ComEA binds incoming DNA in the periplasm of naturally competent Vibrio cholerae cells thereby promoting DNA uptake, possibly through ratcheting and entropic forces associated with ComEA binding. Using comparative modeling and molecular simulations, we projected the 3D structure and DNA-binding site of ComEA. These in silico predictions, combined with in vivo and in vitro validations of wild-type and site-directed modified variants of ComEA, suggested that ComEA is not solely a DNA receptor protein but plays a direct role in the DNA uptake process. Furthermore, we uncovered that ComEA homologs of other bacteria (both Gram-positive and Gram-negative) efficiently compensated for the absence of ComEA in V. cholerae, suggesting that the contribution of ComEA in the DNA uptake process might be conserved among naturally competent bacteria.
| Horizontal gene transfer (HGT) plays a key role in transferring genetic information from one organism to another. Natural competence for transformation is one of three modes of HGT used by bacteria to promote the uptake of free DNA from the surrounding. The human pathogen Vibrio cholerae enters such a competence state upon growth on chitinous surfaces, which represent its natural niche in the aquatic environment. Whereas we have gained a reasonable understanding on how the competence phenotype is regulated in V. cholerae we are only at the beginning of deciphering the mechanistic aspects of the DNA uptake process. In this study, we characterize the competence protein ComEA. We show that ComEA is transported into the periplasm of V. cholerae and that it is required for the uptake of DNA across the outer membrane. We demonstrate that ComEA aggregates around incoming DNA in vivo and that the binding of DNA is dependent on specific residues within a conserved helix-hairpin-helix motif. We propose a model indicating that the DNA uptake process across the outer membrane might be driven through ratcheting and entropic forces associated with ComEA binding.
| Recombination between the bacterial chromosome and DNA fragments that enter the cell through horizontal gene transfer (HGT) either replace damaged or mutated alleles with the original alleles, thereby repairing the gene, or transfer mutated alleles or new genes to naïve strains. Thus, HGT plays a key role in transferring genetic information from one bacterium to another and maintaining the balance between genome maintenance and evolution. Natural competence for transformation is one of three modes of HGT in bacteria and promotes the uptake of free DNA from the environment (for recent reviews see [1]–[6]).
Many naturally transformable bacteria have been described [7], including the pathogenic bacterium Vibrio cholerae [6], [8]. The physiological state of natural competence of this Gram-negative bacterium is associated with its primary niche, the aquatic environment. Within this habitat, V. cholerae attaches to the exoskeleton of zooplankton or zooplankton molts [9]. Those exoskeletons comprise the polymer chitin, which is the natural inducer of competence in V. cholerae [6], [8], [10]. Whereas the regulatory network driving competence has been well investigated (reviewed by Seitz and Blokesch [6]), so far very little is known about the DNA uptake complex of V. cholerae [11]. With respect to the DNA uptake machinery of naturally transformable bacteria it has been suggested that a (pseudo-)pilus [1], [2], similar to type IV pili (Tfp) [12], represents a core element of the DNA import machinery. However, it is still unclear how the proteins interact to pull the transforming DNA through the cell envelope [3]. A proposed mechanism for DNA uptake involves repeating cycles of pilus extension and retraction [1], [2], [4], [13] although recent review articles suggested that other competence proteins, such as ComEA, might be involved in pulling the DNA into the cell [4], [14] (though without experimental evidence). The present study reinforces those ideas and shows that ComEA is a prerequisite for DNA uptake in naturally competent V. cholerae. Furthermore, based on an earlier study on DNA ejection from bacteriophages [15] we propose a model suggesting that the DNA translocation across the outer membrane is possibly accomplished by ratcheting and entropic forces associated with the binding of ComEA to the incoming DNA.
Currently, the majority of studies on the cellular localization of competence proteins were performed on Gram-positive bacteria [16]–[19], whereas far less is known about competence protein localization in Gram-negative bacteria. We recently identified the minimal competence gene set of V. cholerae and provided first insight into the DNA uptake machinery of this organism [11]. Notably, through the analysis of knockout strains lacking specific components of the DNA uptake complex we demonstrated that natural transformation still occurred in the absence of the proteins involved in the Tfp structure and biosynthesis though at very low frequencies. Such rare transformants were never detectable for comEA− strains [11], suggesting that ComEA plays an important role in the DNA uptake process, the focus of this work.
In studies on B. subtilis and S. pneumoniae it was reported that binding of transforming DNA to those Gram-positive cells is at least partially mediated by ComEA and that ComEA is “absolutely required” for DNA uptake and transformation [20]–[22]. Likewise, ComE (ComEA homolog)-negative strains of Neisseria gonorrhoeae [23] and V. cholerae [8], [24] were severely or completely impaired for natural transformability, indicating that ComEA might also play an important role in Gram-negative bacteria. A recent study by Lo Scrudato and Blokesch indicated that comEA and the gene encoding the inner membrane transporter comEC were differentially regulated from the Tfp-like components of the DNA uptake machinery [25], [26], which, together with our study on the DNA uptake machinery, suggest that DNA transport might be a multi-step process in V. cholerae (as previously proposed for Helicobacter pylori [14], [27], which does not contain a bona fide Tfp-based DNA uptake machinery). Here, we show that the Tfp-like elements of the DNA uptake machinery of V. cholerae are not sufficient to translocate DNA across the outer membrane and that the competence protein ComEA plays an essential role in this process.
In a previous study by Chen and Gotschlich the authors predicted a 19-residues signal sequence for sec-dependent transport of the ComEA-homolog of Neisseria gonorrhoeae (ComE) into the periplasm [23]. Such a predictable signal sequence (amino acid residues 1–25) is also present in ComEA of V. cholerae. To experimentally address the localization of the ComEA protein we aimed at visualizing it in vivo by constructing a functional translational fusion between ComEA and mCherry. Using this construct we observed a uniform localization pattern of ComEA (Fig. 1A), which is consistent with the presence of such an N-terminal signal sequence and the transport of ComEA to the periplasm. To validate this microscopical observation, we generated a translational fusion between comEA and the gene encoding beta-lactamase (bla; without the region encoding the signal sequence), which replaced the wild-type comEA allele on the V. cholerae large chromosome. The resulting strain retained natural transformability at a frequency of 2.5×10−5±3.0×10−5 compared with 7.9×10−5±2.5×10−5 for the parental wild-type strain (average of four biological replicates ± SD) indicating the functionality of the fusion construct. Most importantly, the construct conferred full resistance to ampicillin, which provides further evidence for the periplasmic localization of ComEA-bla as beta-lactamase can only exert activity against beta-lactam antibiotics in the periplasm of Gram-negative bacteria (Fig. S1).
Next, we aimed to investigate whether the ComEA protein is motile within the periplasm. To this extent we used a fluorescence loss in photobleaching (FLIP; Fig. 1B) approach because photobleaching can reveal protein dynamics in live cells [28]. In contrast to fluorescence recovery after photobleaching (FRAP), where fluorescent proteins within a small area of the cell are bleached and the back-diffusion of the surrounding non-bleached proteins into this region is recorded, FLIP consist of repetitive bleaching of the same region (e.g. region of interest 1 in Fig. 1B), thereby preventing fluorescence recovery in that region. Moreover, any mobile protein from elsewhere in the same compartment (e.g. region of interest 2 in Fig. 1B) will also enter this continuously photo-bleached area, eventually resulting in a complete loss of fluorescence in the compartment. In contrast, any not connected compartment will be spared from bleaching (e.g. region of interest 3 in Fig. 1B). Therefore, FLIP is often used to reveal the mobility of proteins within certain compartments of the cell [29], which is what we were aiming for. Indeed, our FLIP experiments indicated that ComEA was highly motile within the periplasm (Fig. 1B). Likewise, a translational fusion between the signal sequence of ComEA (amino acid residues 1–25; ss[ComEA]) alone and mCherry resulted in a similar localization (Fig. 1A) and mobility pattern (Fig. S2). This uniform localization pattern differed from that obtained from previous studies on B. subtilis, where Hahn et al. used immunofluorescence microscopy to show that ComEA localizes in a non-uniform punctate manner [16]. Kaufenstein et al. confirmed those data and concluded that the distinct assemblies of ComEA were mobile [19].
Studies using purified tagged ComEA/ComE homologs demonstrated that the protein binds DNA in vitro; thus, ComEA was considered as a DNA receptor protein [21], [23], [30], [31]. DNA binding could be attributed to a conserved helix-hairpin-helix (HhH) motif [32]. Notably and in contrast to helix-turn-helix or helix-loop-helix motifs, which are widespread in proteins that interact with DNA in a sequence-dependent manner, HhH motifs bind DNA in a non-sequence-specific manner. Such binding is based on hydrogen bonding between the protein and the DNA phosphate groups [32] and HhH motifs have been described in various protein classes, including DNA polymerases, DNA ligases or DNA glycosylases [32], [33]. However, the in vivo binding of DNA through ComEA has never been demonstrated. We genetically engineered a fusion protein between ComEA and GFP, which was transported across the inner membrane via the Tat-transport machinery in a folded state (as GFP is improperly folded when translocated to the periplasm in a sec pathway dependent manner [34]). Interestingly, the protein failed to translocate in Escherichia coli; instead, ComEA was tightly bound to the bacterial chromosome, which appeared as a highly compacted structure (Fig. 2A). The increased protein expression levels resulted in cell death, indicating that the strong binding of ComEA to the DNA in vivo interfered with cellular processes. Due to this lack of translocation of ComEA-GFP into the periplasm and the in vivo binding to the chromosome we conducted further experiments using the ComEA-mCherry fusion despite the lower signal intensity of mCherry compared with GFP.
To investigate the function of ComEA in vivo, we excluded artifacts caused by artificial (over-)expression as those have been recognized as having detrimental effects on subcellular localization [35]. Thus, all V. cholerae strains used in these experiments were generated through the substitution of chromosomal comEA with diverse comEA-mCherry alleles. In these strains, the expression of comEA-mCherry was driven through its native promoter and consequently co-regulated with other competence genes. The functionality of the chromosomally encoded ComEA fusion protein was confirmed using a transformation assay, and the chromosomally-encoded fusion protein was uniformly localized within the periplasm (Fig. 2B and Fig. S3). Importantly, the addition of external transforming DNA (tDNA) led to the formation of distinctive ComEA-mCherry foci (Fig. 2B). The size and numbers of these protein aggregates was dependent on the length of the supplemented tDNA. Periplasmic mCherry alone did not aggregate (ss[ComEA]-mCherry; Fig. 2B). A similar relocalization pattern after the addition of external DNA was also observed when the cells were grown on chitin surfaces mimicking the natural reservoir V. cholerae (Fig. S4). This observation suggested that ComEA binds transforming DNA in the periplasm thereby potentially contributing to DNA translocation across the outer membrane.
To test this hypothesis, we repeated the experiments using YoYo-1-labeled DNA. Indeed, a perfect colocalization pattern was observed when the fluorescent signals of ComEA-mCherry and DNA were compared (Fig. 2C). Foci formation through ComEA and colocalization with YoYo-1-labeled DNA were absent in a strain lacking the outer membrane pore PilQ [2], [8], [11], whereas the absence of the inner membrane transporter ComEC did not interfere with ComEA-DNA colocalization (Fig. 2C). Similar foci formation of YoYo-1-labeled DNA was also observed in a strain carrying wild-type ComEA, excluding a translational artifact resulting from the mCherry-fusion (Fig. 2C). Notably, YoYo-1 foci were absent in a comEA-negative strain, which was also the case for a strain lacking the major Tfp subunit PilA (Fig. S5A).
Using a whole-cell duplex PCR-based DNA uptake assay [24], [11] that aims at detecting DNA strands, which have either entered the periplasm or have already reached the cytoplasm of the competent bacteria (thereby becoming resistant against externally applied DNase), we confirmed that tDNA (both unlabeled or YoYo-1-labeled) was undetectable in comEA-negative strains even though it was readily detectable in the wild-type strain and in comEC negative derivatives (Fig. 2D and Fig. S5).
Whereas the absence of YoYo-1 labeled DNA foci and PCR-amplifiable DNA in comEA negative strains is indicative of a failure to transport tDNA across the outer membrane, such results would also be consistent with ComEA's main function being to protect and stabilize incoming tDNA against potential nucleases. Indeed, two nucleases have been described for V. cholerae, Dns and Xds, which are solely responsible for extracellular nuclease activity in this organism [36]. Interestingly, Focareta and Manning demonstrated that even though Dns can be recovered from culture supernatants, it was also detectable in the periplasmic space of V. cholerae [37]. We recently confirmed the extracellular localization of Dns [38] but also its at least partial association with the bacterial cells (through western blot analysis; [25]). Moreover, Blokesch and Schoolnik showed that expression of dns has to be silenced in V. cholerae to allow natural transformation to occur at high cell density [26], [38]. Thus, to rule out the possibility that ComEA might protect incoming tDNA against either of those two nucleases we tested dns, xds, and comEA single, double, and triple mutants for natural transformation and the recovery of DNase resistant tDNA in whole cells (Fig. S6). Notably, the absence of dns resulted in higher transformability (Fig. S6A), consistent with an early study [38], and in the detection of increased amounts of DNase-resistant tDNA within the bacteria (Fig. S6B). However, no transformants or translocated tDNA were detectable if comEA was concomitantly absent (Fig. S6). We therefore conclude that ComEA's main role is not to protect incoming tDNA against degradation by the nucleases Xds or Dns, though we cannot exclude the presence of any other hitherto unidentified nuclease in the periplasm of V. cholerae. Instead, we suggest that translocation of tDNA across the outer membrane is not solely driven through Tfp-like elements of the DNA uptake machinery but also requires ComEA.
To gain insights into the molecular mechanism through which ComEA binds dsDNA, we predicted the structure of ComEA and characterized the interactions of this protein with the transforming DNA. First, we used comparative modeling to create a 3D structure of ComEA using the X-ray structure of the ComEA-related protein HB8 from Thermus thermophilus (PDB ID: 2DUY, unpublished) as a template (Fig. 3, movie S1). Based on structural similarity with structures from the HhH family [39], we identified K62 and K63 as candidate residues for DNA binding interactions and could model the putative ComEA-DNA adduct (Fig. 3B). The electrostatic potential of the ComEA model is consistent with the identified DNA-binding region, showing positively charged regions corresponding to the lysine pair (Fig. 3C).
To validate this model, we used site-directed mutagenesis to create ComEA variants with single or double amino acid substitutions. All comEA-mCherry alleles were inserted into the chromosome, thereby replacing the wild-type comEA copy. The ComEA-mCherry variants were tested for expression and periplasmic localization, foci formation upon provision of tDNA, for their ability to induce DNA translocation into a DNase resistant state (using the DNA uptake assay) and to restore natural transformation (Fig. 4 and Fig. S7). Consistent with the in silico predictions, K63 was of major importance. ComEAK63A was severely impaired for natural transformation (∼250-fold reduction; Fig. 4C), resulting in DNA uptake levels below the limit of detection (Fig. 4B). The substitution of K63 with a negatively charged residue (ComEAK63E) or the concomitant exchange of K62 (ComEAK62/63A) completely abolished natural transformation (Fig. 4C). The ComEA-DNA model also explains why K63 has the major role in DNA binding: while K62 is engaged with a single backbone phosphate moiety, K63 is inserted into the DNA minor groove, chelating the backbone of both strands (Fig. 3B, inset). Moreover, a substitution of the nearby glycine residue at position 60 by alanine had no effect on DNA binding and transformation, whereas strains producing ComEAG60V and ComEAG60E were impaired in DNA uptake and were non-transformable (Fig. 4). We suggest that the combined effect of impairing the interactions of K62 and K63 with the dsDNA (as in the case of ComEAG60E) and perturbing the HhH1 GIG hairpin motif (Fig. 3A) has a major impact on the ability of ComEA to bind DNA.
To further investigate whether the lysine pair is indeed involved in DNA binding we heterologously expressed those variants as tat-ComEA-GFP fusions in E. coli (Fig. S8; as for wild-type ComEA in Fig. 2A). Using this approach we showed that the ComEAK63E and ComEAK62/63A variants behaved differently from WT ComEA in that they localized evenly within the cytoplasm. In addition, most of the E. coli cells did not show any compaction of the chromosome (and if so the variant did not co-localize with the compacted chromosome). The same phenotype was observable for variants that lacked either of the two HhH motifs (Fig. S8), suggesting that those variants had lost their ability to bind DNA. In contrast, a K63A variant showed an intermediate phenotype (Fig. S8) consistent with the ∼250-fold decreased transformation frequency observed for the ComEAK63A-mCherry variant in V. cholerae (Fig. 4).
Apart from this patch at HhH1, the only other amino acid important for the in vivo functionality of ComEA was the conserved arginine residue at position 71 (Fig. 3A). The DNA uptake ability of ComEAR71A was slightly reduced, and less DNA-protein foci were observed for this variant (Fig. S7). However, the strain containing ComEAR71A remained naturally transformable, a feature that was completely abolished for the ComEAR71D variant. The latter mutant protein was also unable to bind DNA within the periplasmic space and did not foster the uptake of transforming DNA (Fig. S7). Based on our ComEA model structure, R71 is located in a position not particularly favorable for DNA binding (Fig. 3B); therefore, it is likely that R71 might be important for the structural stability of ComEA.
To unambiguously show that the lysine residues are required for DNA binding we purified a tagged (Strep-tag II) version of ComEA, ComEAK62/63A, ComEA-mCherry, and ComEAK62/63A-mCherry (Fig. S9). The purified ComEA protein showed an unexpected UV-Vis spectrum, which was consistent with bound DNA (due to an absorption peak around 260 nm; Fig. S9A). Interestingly, if we compared purified ComEA-mCherry with the ComEAK62/63A-mCherry, we observed that the peak at 260 nm was absent in this variant, indicating that the protein was indeed no longer able to bind DNA.
To remove any pre-bound DNA from the ComEA protein we included a DNase treatment step prior to the elution of the protein from the affinity column (see Material and Methods; Fig. S9C and D). All four proteins were tested for in vitro binding to DNA using an electrophoretic mobility shift assay (EMSA). Notably, ComEA-mCherry and ComEA bound to DNA in a concentration dependent manner as visualized by the retarded migration of the DNA probe (Fig. 5A and Fig. S10A) and the likewise changed migration of the protein (visualized by the fluorescence of mCherry; Fig. 5A). Notably, the K62/63A variants of ComEA did not change the migration behavior of the DNA probe (Fig. 5B and Fig. S10B), again confirming that the protein had lost the ability for DNA binding.
It should be noted that the shifted DNA signal was detectable at DNA to protein ratios as low as 1∶10 and the probe seemed completely shifted at a ratio of 1∶25–30 (Fig. 5A and Fig. S10A), which was significantly lower than what has been described for the B. subtilis ComEA homolog (98% of the DNA probe was shifted when 5.5×10−11 M of DNA was incubated with 1.6 µM of purified protein; [21]) or for the neisserial ComE ortholog [23]. A possible explanation for this difference could be that the ComEA/ComE proteins investigated in those earlier studies were pre-occupied by DNA as we observed for ComEA of V. cholerae in the absence of DNase treatment.
Provvedi and Dubnau suggested that the in vitro DNA binding behavior of the ComEA protein of B. subtilis was indicative of cooperative binding [21]. To test whether any cooperative binding was observable for ComEA of V. cholerae we used Atomic Force Microscopy (AFM). AFM allows investigating the extent of ComEA-mCherry binding to a DNA fragment and to also determine where on the DNA the protein is bound (e.g. fractional occupancies at any specific site, binding to the ends, or to nonspecific sites). To minimize overestimation of the binding affinity that can occur in the case when coverage of protein on the surface is too high, such that the protein coincidently lands on DNA, we kept the DNA-protein molecular ratio low by not exceeding a ratio of 1∶10 (DNA to protein). Prior to AFM imaging, we pre-incubated the ComEA-mCherry protein with a random PCR fragment (809 bp) at a molecular ratio of 1∶2.5 or 1∶10. As illustrated in Fig. 5 we observed a mixture of bare DNA molecules, free protein molecules, and protein/DNA complexes. To identify the ComEA-mCherry protein in topographic AFM images we used height and width criteria (height >2 nm, width from 10 to 20 nm). Using an approach reported by Yang et al. [40] we found that the probability of protein molecules located on DNA was 5 times higher than it would be for stochastically binding of the protein to the mica surface. Moreover, in the case of a DNA to protein ratio of 1∶10 we observed 2.5-fold higher affinity of the protein to the free ends of DNA than to random sites on the DNA strand. These AFM data indicate that, at least at the measured concentrations, no cooperative binding of the ComEA protein to DNA occurred and again contradicts the hypothesis that binding of ComEA might primarily protect the tDNA from degradation. Such protective effect has been demonstrated for the competence protein DprA of Streptococcus pneumoniae [41], which binds the single-stranded tDNA after its translocation into the cytoplasm. Indeed, Mortier-Barrière et al., described in their study that DprA binding to DNA appeared to be cooperative since fully covered protein-DNA complexes were observed next to free ssDNA molecules at a protein to nucleotide ratio of 1∶20. We never observed such scenario for ComEA's binding to dsDNA using AFM (though we used a ∼4-fold lower protein to nucleotide ratio).
Interestingly, a passive DNA uptake mechanism has recently been proposed for single-stranded T-DNA translocation into plant cells involving the VirE2 protein of Agrobacterium tumefaciens [42]. We reasoned that if a similar mechanism is responsible for DNA uptake in competent V. cholerae cells, although dsDNA is involved and ComEA shows no similarity to VirE2, then the aggregation of ComEA should occur at one distinct DNA entry point (most likely next to the PilQ secretin). To test this hypothesis, we performed time-lapse microscopy experiments using ComEA-mCherry-expressing V. cholerae strains in the presence of external DNA (Fig. 6). We consistently observed the accumulation of ComEA as one large focus before smaller subclusters separated from the main ComEA focus and spread throughout the periplasm until the uniform localization of ComEA was restored (Fig. 6, movies S2, S3, S4).
Based on the data presented above we hypothesize that ComEA might play a direct role in the translocation of DNA across the outer membrane solely based on its ability to bind to DNA. If this were the case then ComEA homologs of other naturally competent bacteria should be able to replace ComEA of V. cholerae. And indeed, ComEA of B. subtilis was able to efficiently compensate for the absence of ComEA of V. cholerae (Fig. 7). Moreover, even the C-terminal (HhH)2 motif of ComEA of B. subtilis alone, which was shown to bind DNA in vitro [21], was sufficient to restore natural transformation of a comEA negative V. cholerae strain as were the ComEA homologs from N. gonorrhoeae, Haemophilus influenzae, and Pasteurella multocida (Fig. 7). It should be noted that Sinha et al. suggested that H. influenzae might contain an additional but so far unidentified paralog of comE1 due to the modest effect observed for a comE1 minus strain [43].
It is tempting to speculate that ComEA might fulfill a similar role in Gram-positive bacteria. Indeed, the localization of ComEA has been previously described for B. subtilis [16], [17], [19] but those studies were either based on immunofluorescence microscopy [16], which does not allow following protein localization over time, or were done in the absence of tDNA [17], [19]. Therefore, it was concluded by Kaufenstein et al. that ComEA localizes to many sites of the cell membrane and only occasionally co-localizes with the polar DNA uptake machinery, which was mainly achieved by changing the artifical inducer concentration [19]. However if the cell wall would be considered as a similar barrier in Gram-positive bacteria as the outer membrane is in Gram-negatives, creating a kind of periplasmic space between the cell wall and the (inner) membrane as suggested by Matias and Beveridge [44], then the binding of ComEA could also participate in the transport of DNA across the cell wall layer. However, in contrast to ComEA of Gram-negative bacteria, ComEA of Gram-positives is anchored to the membrane and therefore accumulation of ComEA can only occur in two dimensions, which might still be sufficient to prevent backward diffusion of the tDNA and contribute to DNA translocation across the cell wall. Notably, while this article was under revision Bergé et al. published a study on the nuclease EndA of naturally competent Streptococcus pneumoniae [45]. The authors demonstrated that EndA aggregates at midcell in this Gram-positive bacterium and that this recruitment is dependent on “the dsDNA receptor” ComEA [45]. Interestingly, ComEA also localized to the midcell and the authors speculated “a direct interaction of EndA and ComEA, an hypothesis which received indirect support” [45].
Our findings suggest that the ability of ComEA proteins to bind to dsDNA emerging from the PilQ pore can potentially prevent the retrograde movement of the substrate, and ComEA binding might contribute to pull DNA into the periplasm (Fig. 8). It has been suggested that ratcheting produced through binding proteins can significantly accelerate translocation events [46], [47], as for the case of phage DNA injection into bacterial cells [15]. Based on our data, a similar mechanism can be envisioned for the ComEA-mediated transfer of DNA into the periplasm, with the rate of uptake depending on the specific binding kinetics and concentration of ComEA [15]. We hypothesize that ComEA-mediated DNA internalization might start occurring once short stretches of tDNA would enter the periplasm (most likely through the outer membrane secretin PilQ and potentially after a single Tfp retraction event). The ratio between the periplasmic ComEA protein and the incoming tDNA should be high at that stage thereby leading to an increased ComEA effective binding density, which, potentially together with the higher affinity of ComEA for DNA ends as observed by AFM (Fig. 5), would promote efficient DNA internalization. The absence of cooperative ComEA-DNA binding revealed by our AFM data (Fig. 5) is not an obstacle to a ComEA-mediated ratchet mechanism of internalization, as cooperativity would only contribute to increase the relative speed of the process [15], [46], [47]. The binding of proteins has undeniably been recognized as a driving force, both in the translocation of proteins as well as of DNA [48]. To this extent, Salman et al. investigated the translocation of double-stranded (ds) DNA through the nuclear pore complex using a combination of epifluorescence microscopy and single-molecule manipulation techniques [49]. They presented evidence that the DNA uptake process in their reconstituted system was based on a passive ratchet, directed by the retention of the already translocated segment of the DNA [49]. We suggest that ComEA might play a similar role in the DNA uptake process in naturally competent V. cholerae cells.
In summary, we used a cell biological approach to better understand DNA uptake in naturally competent V. cholerae cells. We visualized the competence protein ComEA and observed the in vivo binding of this protein to dsDNA in real time. Structural modeling and AFM experiments suggested that the binding of ComEA to DNA is primarily responsible for DNA translocation across the outer membrane. Consistent with this suggestion, ComEA variants unable to bind to DNA in vivo were also defective in promoting DNA uptake and natural transformation. We hypothesize that ComEA encounters incoming DNA immediately after short stretches of DNA have crossed the outer membrane (through the PilQ secretin or in exceptional cases also in a Tfp-independent manner [11]) and that ComEA subsequently promotes DNA translocation across the outer membrane without the need for any external energy source (Fig. 8). ComEA might therefore be more than a DNA receptor protein, but rather a crucial player for mediating DNA uptake in V. cholerae and potentially also other naturally competent bacteria.
Vibrio cholerae strains and plasmids used in this study are listed in Table S1. Escherichia coli strain DH5α [50] was used as host for cloning purposes and for heterologous expression of ComEA and its variants for protein purification. Genomic DNA (gDNA) extracted from E. coli BL21 (DE3) [51] was utilized to test DNA uptake by PCR as described [24]. E. coli S17-1λpir [52] served as donor strain for bacterial mating with V. cholerae.
All V. cholerae and E. coli strains were grown aerobically in Luria-Bertani (LB) medium at 30°C and 37°C, respectively. Solid LB plates contained 1.5% agar. For tfoX expression and induction of other constructs under control of the PBAD promoter the LB medium was supplemented with 0.02% L-arabinose (L-ara). For expression of tat-gfp, tat-comEA-gfp, and its derivatives in E. coli DH5α (Fig. 2A and Fig. S8) L-ara concentrations were lowered to 0.002%. Thiosulfate Citrate Bile Salts Sucrose (TCBS) agar plates were prepared following the manufacturer's instructions (Fluka) and used to counterselect E. coli after bacterial mating. For sucrose-based counterselection, NaCl-free LB medium containing 6% sucrose was used. LB medium and LB agar plates were supplemented with antibiotics when required. Final concentrations of antibiotics were 50 µg/ml, 75 µg/ml and 100 µg/ml for gentamicin, kanamycin, and ampicillin, respectively. The ampicillin concentration was lowered to 50 µg/ml for V. cholerae strains induced for competence.
Standard molecular biology-based methods were used for DNA manipulations. Restriction enzymes and DNA modifying enzymes were obtained from New England Biolabs, Taq DNA polymerase (GoTaq) was obtained from Promega and used for colony PCR, and Pwo DNA Polymerase (Roche) was used for high-fidelity PCR amplifications. Modified DNA sequences were verified using Sanger sequencing (Microsynth, CH).
All plasmid constructs were based on pBAD/Myc-HisA (Invitrogen), which contains the araBAD (PBAD) promoter followed by a multiple cloning site (MCS) for dose-dependent protein expression. A derivative of pBAD/Myc-HisA, pBAD(kan), was created through substitution of the ampicillin resistance cassette (bla) with a kanamycin resistance cassette (aph). The genes and translational fusion constructs were PCR amplified and cloned into the MCS of pBAD/Myc-HisA or pBAD(kan). For the amplification of V. cholerae genes, the gDNA of strain A1552 [53] served as a template. The accuracy of the plasmids was verified through sequencing.
Genes were deleted from the parental strain A1552, using either a gene disruption method based on the counter-selectable plasmid pGP704-Sac28 [54], or natural transformation and FLP recombination, as recently described (TransFLP method [55]–[57]).
Strains containing comEA-mCherry or site-directed variants thereof were constructed using the TransFLP method [55]–[57]. For the construction of ComEA site-directed variants, a silent ‘watermark’ restriction site was inserted close to or including the changed nucleotide sequence. This watermark simplified screening purposes after homologous recombination.
The comEAB.s. gene (or parts thereof) was amplified from gDNA derived from B. subtilis strain 168. The DNA fragment containing comE1 from Neisseria gonorrhoeae (N.g.; Neisseria gonorrhoeae strain FA 1090, NCBI Reference Sequence: NC_002946.2; locus YP_208252), Haemophilus influenzae (H.i.; Haemophilus influenzae strain R2846, NCBI Reference Sequence: NC_017452.1; locus YP_005829750), and Pasteurella multocida (P.m.; Pasteurella multocida subsp. multocida str. Pm70, NCBI Reference Sequence: NC_002663.1; locus NP_246604, hypothetical protein PM1665) was synthesized using the GeneArt® Strings™ technology (Life technologies/Invitrogen) and served as PCR template for the TransFLP strain construction method [55]–[57]. The beta-lactamase gene (bla) was amplified from plasmid pBR-flp [55]–[57]. All strains were verified through colony PCR (in part followed by restriction enzyme digestion according to inserted watermarks) and confirmed through PCR amplification and sequencing.
Microscopy images were obtained using a Zeiss Axio Imager M2 epifluorescence microscope. Details about the instrumentation and configurations are provided elsewhere [25]. All bacterial samples were mounted on 2% agarose/PBS pads. Image processing and annotation was done using ImageJ and Adobe Illustrator.
Strains carrying fluorescent fusion constructs were grown aerobically for ∼5 h in LB supplemented with the respective antibiotics and 0.02% L-arabinose (0.002% L-ara for E. coli experiments; Fig. 2 and Fig. S8). The strains carrying chromosomally encoded fluorescent fusion proteins were grown aerobically and at 30°C in LB supplemented with 0.02% L-ara for ∼7 h (OD600 2.5; [11]). The samples were washed once in PBS and immediately imaged.
The staining of chromosomal DNA was performed through the addition of 4′,6-diamidino-2-phenylindole (DAPI; final concentration 5 µg/ml) to the bacterial cultures for at least 5 min.
To characterize the ComEA-mCherry localization dynamics during DNA uptake, comEA-mCherry-expressing strains were grown as described above. A total of 50 µl of washed culture was mixed with 1 µg of either gDNA derived from V. cholerae strain A1552-lacZ-Kan [58], commercially available phage lambda DNA (Roche) or a 10.3 kb fragment amplified through PCR. After 5 min of incubation with the DNA the bacteria were mounted on agarose pads and imaged. To visualize the DNA during the relocalization of ComEA-mCherry, phage lambda DNA (Roche) was pre-stained with 10 µM YoYo-1 (Molecular Probes/Invitrogen) at 4°C corresponding to a base pair to dye ratio of 15∶1. The bacterial culture was mixed with the pre-stained DNA and incubated for 20 min. The cells were washed in PBS, mounted on agarose pads and imaged.
For time-lapse microscopy, the samples were prepared as described above, but immediately imaged after the addition of DNA. The images were taken every 3 or 120 sec as indicated in the figure and movie legends. For time-lapse imaging, the agarose pads were sealed using a mixture of Vaseline, lanolin and paraffin (VALAP).
Fluorescence loss in photobleaching (FLIP) experiments were performed on a Zeiss LSM710 microscope equipped with a 561 nm solid-state laser (20 mW). A Plan-Apochromat 63×/1.40 Oil objective was used. The microscope was controlled with the Zen 2009 software suite (Zeiss). Time intervals ranged from 104 to 120 ms/frame for live cells to max. 160 ms/frame for fixed cells. The maximum (100%) laser power was used for bleaching.
V. cholerae strains ΔcomEA-TntfoX harboring pBAD(kan)-comEA-mCherry or pBAD(kan)-ss[ComEA]-mCherry were grown aerobically for 5 h in LB supplemented with 0.02% arabinose and 75 µg/ml of kanamycin. After the cells were mounted, the slides were sealed and the bacteria were immediately imaged (live samples; Fig. 1 and Fig. S2A). Alternatively, the cells were fixed for 30 min (4% paraformaldehyde/150 mM phosphate buffer) before imaging (fixed samples; Fig. S2B).
For FLIP data acquisition a circular bleaching region of ∼440 nm width was defined at one cell pole (region-of-interest (ROI) 1; labeled as 1 in Fig. 1). A circular ROI of the same size was defined at the opposite cell pole of the same bacterium (labeled as 2 in Fig. 1) and in an adjacent cell (labeled as 3 in Fig. 1). The average fluorescence intensity of all regions was recorded. Bleaching of ROI 1 was initiated after a lag of 20 frames and repeated after each frame. The acquired data were exported and processed in ‘R’ [59]. The recorded fluorescence intensities were normalized to the average fluorescence intensity of the first 10 frames. Moving averages were calculated using the SMA(x, n = 5) function from the ‘TTR’ package [59].
Transformation assays were performed as previously described [25] with gDNA of strain A1552-lacZ-Kan [58] as transforming material. Transformation frequencies were calculated as the number of transformants divided by the total number of colony forming units (CFU). Differences in transformation frequencies were considered significant for P-values below 0.05 (*) or 0.01 (**) as determined by Student's t-test on log-transformed data.
DNA uptake was verified using a whole-cell duplex PCR assay as described [24] with slight modifications. Briefly, competence-induced bacteria were grown aerobically until an OD600 of 1.0–1.5 before genomic DNA (gDNA) (2 µg/ml) of E. coli strain BL21 (DE3) was added for 2 h. For the uptake of YoYo-1-labeled DNA gDNA of E. coli strain BL21 (DE3) was pre-labeled as described for the microscopy experiments and YoYo-1 was maintained in the solution throughout the 2 h incubation period. Next, cells were harvested and treated with DNase I (Roche) for 15 min at 37°C. Excess nuclease was removed by washing and cells were resuspended in 100 µl PBS. ∼3×106 bacteria were used as template in a whole-cell duplex PCR. Primer pairs were specific for the donor DNA derived from E. coli BL21 (DE3) and for gDNA of the V. cholerae acceptor strain (at a 10-fold lower concentration). The latter reaction served as control for the total number of acceptor bacteria [24].
A 3D model structure was produced for ComEA (truncating the first 37 residues including the 25 residue-containing signal peptide) using comparative modeling (MODELLER package [60]) on the Thermus thermophilus HB8 (PDB ID: 2DUY) template (with 43% sequence identity) (Fig. 3). The ComEA-DNA complex was modeled, to identify structurally similar DNA-binding proteins using the DALI server [39]. The DNA polymerase, PolC, from Geobacillus kaustophilus (PDB ID: 3F2D) [61] was selected as the best match, with 24% sequence identity and a root mean square deviation (RMSD) of 2.4 Å compared with the modeled ComEA of V. cholerae. The PolC X-ray structure complexed with DNA was used to identify potential DNA poses on the V. cholerae ComEA model using the Chimera MatchMaker tool [62]. This assessment led to the production of a DNA-ComEA model (Fig. 3B, movie S1), which was further refined and equilibrated using the minimization and molecular simulations detailed below. The estimated binding energy for the ComEA-DNA association is in the order of 29±8 kcal/mol, based on MM/PBSA calculation on the MD trajectory.
Molecular dynamics simulation was used to relax and study the dynamics and energetics of ComEA and the ComEA-DNA complex for 55 and 50 ns, respectively. The MD simulations were run using the NAMD simulation package [63] with Amber force field (with Barcelona modification for nucleic acids [64] and the TIP3P water model [65]. The systems were first energy minimized using constrained C-alpha atoms, followed by analysis without any constraint for 2000 steps. To equilibrate the system, the temperature was gradually increased up to 300 K in the NVT ensemble and maintained at 300 K for 100 ps with a 1 fs time step. Finally, an NPT simulation was run at 300 K for 500 ps with a 2 fs time step to complete the equilibration procedure. The equilibrated structure was used as starting point for production simulations. All production MD simulations were run at 1 bar with a time step of 2 fs, using SHAKE algorithm [66] on all bonds and PME [67] for treating electrostatic interactions. To control the temperature and the pressure, Langevin dynamics and the Nose-Hoover Langevin piston, respectively, were used [68], [69]. The trajectories were saved every 500 steps in the production simulations. To characterize the binding affinity of different systems, the free binding energies were calculated using the MMPBSA.py package [70]. 100 frames were sampled from the trajectories for analysis using MMPBSA.py. The entropy portion of the free energy was not considered in the calculation. In addition, the PME module in VMD was used to estimate the electrostatics potential of the modeled ComEA monomers (Fig. 3C).
ComEA, ComEAK62/63A, ComEA-mCherry, and ComEAK62/63A-mCherry (all containing the eight amino acid Strep-tag II sequence at the C-terminus) were purified as previously described [71] with minor modification. Briefly, E. coli cells containing the respective plasmids (Table S1) were grown aerobically at 37°C until an OD600 of 1.0. At that time expression was induced by the addition of 0.2% arabinose to the culture medium and the cells were further incubated for 2 hours before their harvest at 4°C and storage of the cell pellet at −80°C. The cells were lysed by sonication (Vibra-cell; 10 min. in total with 30 sec on and 30 sec off intervals and an amplitude of 80%) and the lysate was further processed as described [71]. Notably, after realization that the protein was pre-occupied by DNA (see results section), we included a on-column DNase treatment step (10 µg/ml of DNase I (Roche) in 100 mM Tris/HCl pH 8.0 buffer containing 20 mM MgCl2 and 0.2 mM CaCl2; 30 min. at 30°C) after the soluble protein fraction was loaded onto the streptactin resin and washed with 5 column volumes of washing buffer. The DNase I treatment step was followed by extensive washing of the column (10 to 30 volumes) before the respective protein was eluted as described [71]. The eluted proteins were concentrated using Amicon Ultra spin columns (with a MWCO of 3 kDa or 10 kDa; Millipore). For the AFM experiment, the protein was dialyzed against AFM buffer (5 mM Tris/HCl pH = 8.0 and 10 mM MgCl2). The protein concentration was determined according to Bradford [72].
Electrophoretic Mobility Shift Assays were basically performed as previously explained [71]. However, as preliminary experiments indicated that neither the absence of DTT nor the storage of the protein in the absence of glycerol and at 4°C did change the results of the experiments, the protocol was changed accordingly. The 200 bp DNA fragment was PCR-amplified using gDNA of strain A1552 as template and represented the upstream region of the comEA gene. Other DNA fragments (e.g. the aphA promoter region as previously tested [71]) were similarly shifted (data not shown). The protein/DNA mixture was incubated for 5 min at room temperature before electrophoretic separation on an 8% polyacrylamide gel. DNA was visualized by ethidium bromide staining [71] whereas the fusion proteins (ComEA-mCherry and ComEAK62/63A-mCherry were detected using a Typhoon scanner (GE Healthcare; excitation at 532 nm (green) and emission detected with a 610 BP30 (red) filter).
To prepare the protein/DNA complex we mixed 0.85 ng/µl of a PCR-amplified DNA fragment (809 bp) with the protein in the molecular ratios of 1∶2.5 and 1∶10 (DNA∶protein) in buffer containing 5 mM Tris/HCl pH 8.0 and 10 mM MgCl2. After incubation for 10 min at 37°C, 15 µl of the mixture was deposited on freshly cleaved mica and rinsed thoroughly with ddH2O for two minutes. Preparation of the sample with bare DNA was done under the same conditions but in the absence of the protein. The AFM images were acquired in air and in tapping mode using an Asylum Research Cypher microscope. We used Olympus silicon cantilevers (Olympus OMCL-AC240TS-R3) with a spring constant of 1.7 N/m and a resonant frequency of 70 kHz. The typical scan rate was 2.0 Hz.
|
10.1371/journal.pgen.1005596 | The Rise and Fall of an Evolutionary Innovation: Contrasting Strategies of Venom Evolution in Ancient and Young Animals | Animal venoms are theorized to evolve under the significant influence of positive Darwinian selection in a chemical arms race scenario, where the evolution of venom resistance in prey and the invention of potent venom in the secreting animal exert reciprocal selection pressures. Venom research to date has mainly focused on evolutionarily younger lineages, such as snakes and cone snails, while mostly neglecting ancient clades (e.g., cnidarians, coleoids, spiders and centipedes). By examining genome, venom-gland transcriptome and sequences from the public repositories, we report the molecular evolutionary regimes of several centipede and spider toxin families, which surprisingly accumulated low-levels of sequence variations, despite their long evolutionary histories. Molecular evolutionary assessment of over 3500 nucleotide sequences from 85 toxin families spanning the breadth of the animal kingdom has unraveled a contrasting evolutionary strategy employed by ancient and evolutionarily young clades. We show that the venoms of ancient lineages remarkably evolve under the heavy constraints of negative selection, while toxin families in lineages that originated relatively recently rapidly diversify under the influence of positive selection. We propose that animal venoms mostly employ a ‘two-speed’ mode of evolution, where the major influence of diversifying selection accompanies the earlier stages of ecological specialization (e.g., diet and range expansion) in the evolutionary history of the species–the period of expansion, resulting in the rapid diversification of the venom arsenal, followed by longer periods of purifying selection that preserve the potent toxin pharmacopeia–the period of purification and fixation. However, species in the period of purification may re-enter the period of expansion upon experiencing a major shift in ecology or environment. Thus, we highlight for the first time the significant roles of purifying and episodic selections in shaping animal venoms.
| While the influence of positive selection in diversifying animal venoms is widely recognized, the role of purifying selection that conserves the amino acid sequence of venom components such as peptide toxins has never been considered. In addition to unraveling the unique strategies of evolution of toxin gene families in centipedes and spiders, which are amongst the first terrestrial venomous lineages, we highlight the significant role of purifying selection in shaping the composition of animal venoms. Analysis of numerous toxin families, spanning the breadth of the animal kingdom, has revealed a striking contrast between the evolution of venom in ancient and evolutionarily young animal groups. Our findings enable the postulation of a new theory of venom evolution. The proposed ‘two-speed’ mode of evolution of venom captures the fascinating evolutionary history and the dynamics of this complex biochemical cocktail.
| Venom is an intriguing evolutionary innovation that is utilized by various animals for predation and/or defense. This complex biochemical cocktail is characterized by a myriad of organic and inorganic molecules, such as proteins, peptides, polyamines and salts that disrupt the normal physiology of the envenomed animal. Evolution of venom has been intensively investigated in more recently diverged lineages (for simplicity, we refer to them as ‘evolutionarily younger’ lineages), such as advanced snakes and cone snails, which originated ~54 [1] and ~33–50 [2, 3] million years ago (MA), respectively. Several venom-encoding genes in these animals have undergone extensive duplications [4, 5] and evolve rapidly under the influence of positive selection [6–10]. In contrast, the evolution of venom in most of the ancient lineages, such as cnidarians (corals, sea anemones, hydroids and jellyfish), coleoids (octopus, squids and cuttlefish), spiders and centipedes, remains understudied, if not completely overlooked. Perhaps the only exhaustively investigated ancient venomous clade are the scorpions, which originated in the Silurian about 430 MA [11, 12]. Moreover, certain potent toxins in species separated by considerable geographic and genetic distance can exhibit remarkable sequence conservation (Fig 1). Yet, research to date has solely focused on how positive selection has expanded the venom arsenal, while completely ignoring the role of negative (purifying) selection.
Phylum Cnidaria consists of animals such as sea anemones, jellyfish, corals and hyrdroids that originated in the Ediacaran Period, approximately 600 MA [13–15]. They are characterized by unique stinging organelles called nematocysts with which they inject venom. Cnidaria represents the oldest venomous lineage known and includes some of the most notorious animals, such as the sea wasp (Chironex fleckeri), a species of box jellyfish. Coleoids, which first appeared in the Early Devonian 380–390 MA [16], represent yet another neglected lineage of ancient venomous animals. Although the venomous nature of coleoids was established as early as 1888 [17], their venoms have received scant attention from toxinological research [17–20].
Centipedes are amongst the oldest living terrestrial venomous animals, with the fossil record extending back to ~420 MA [21]. All ~3,300 species of centipedes (class Chilopoda) described to date belong to five extant orders: Craterostigmomorpha, Geophilomorpha, Lithobiomorpha, Scolopendromorpha and Scutigeromorpha. They inject venom into victims via modified first pair of trunk limbs (forcipules) and use venom for predation and defense. Venoms of certain centipedes can cause excruciating pain, paresthesia, edema, necrosis [22, 23] and can be fatal to mammals as large as dogs [24]. Yet, only a handful of centipede toxins have been pharmacologically characterized to date. Similarly, despite their remarkable ability to target a diversity of ion channels, only toxins from certain medically significant species of spiders have been investigated to date [25]. Thus, the evolutionary history and phyletic distribution of venom from these aforementioned ancient lineages, which represent the first venomous animal groups, remain understudied [18, 26–28]. It should be noted that the divergence times of these lineages can be safely assumed to be equivalent to the time of origin of venom in those respective lineages, as all of the examined lineages are (i) venomous, (ii) do not share between them a common venomous ancestor, and/or (iii) for most of them the fossil data clearly indicates the presence of a venom delivery apparatus [29–36].
By examining a large number of nucleotide sequences from a diversity of species, we report for the first time the molecular evolutionary histories of a number of venom protein families in centipedes, spiders and Toxicofera (clade of venomous snakes and lizards) lizards. In contrast to the rapid evolution of venom in evolutionarily younger lineages, we report an unusually high conservation of venom in centipedes and spiders, despite their long evolutionary histories. Moreover, molecular evolutionary assessments of toxin-encoding genes distributed across the tree of life, has unraveled a surprisingly strong influence of negative selection on the venoms of ancient animals. Our findings reveal contrasting trajectories of venom evolution in ancient and evolutionarily young clades, and emphasize the significant roles of purifying and episodic selections in shaping animal venoms. Further, these results enabled the postulation of a new model of venom evolution that captures their evolutionary dynamics, and the rise and fall in evolutionary rates of animal venoms.
Despite the fact that several centipede and spider toxins are capable of exhibiting a diverse array of pharmacological effects, their venoms remain poorly studied. To date, very few studies have examined the evolutionary mechanisms responsible for the diversification of toxins in centipedes [27, 28] and spiders [37–40]. Hence, we assessed the molecular evolutionary regimes of 17 and 10 gene families encoding toxins in the major lineages of centipedes and spiders, respectively. We computed the ratio of non-synonymous (dN) to synonymous (dS) substitutions, called omega (ω), where ω greater than, less than or equal to one is characteristic of positive, negative and neutral selection, respectively. A large proportion of centipede venoms are characterized by β-pore-forming toxins (β-PFT) that are similar to aerolysins and epsilon toxins from bacteria [28]. They are theorized to be responsible for myotoxic and edematogenic activities of centipede venoms [23, 28, 41]. β-PFT has undergone substantial gene duplication and diversification in centipedes [28]. In contrast to venom-encoding genes in evolutionarily younger lineages that continue experiencing positive selection when they diversify via recurrent duplication events, we find that β-PFTs are evolutionarily extremely constrained under negative selection, as indicated by ω smaller one (Table 1). Centipede venoms are also chiefly constituted by cysteine-rich secretory proteins, antigen 5, and pathogenesis-related 1 (CAP) family members and toxins with low-density lipoprotein receptor Class A (LDLA) repeats [28]. While certain CAP proteins in the venoms of centipedes are characterized by trypsin inhibitory activities [42], the precise role of LDLAs remain unknown. We found that both these toxin classes have experienced a strong influence of purifying selection (Table 1).
Scoloptoxin (SLPTX) is a family of cysteine-rich peptides found in the venoms of several centipedes, where different members exhibit a diversity of pharmacological activities [28]. SLPTX1 appears to be similar to insect peritrophic matrix proteins and has been theorized to be one of the most basally recruited toxins in centipedes [28]. Despite its long evolutionary history, this toxin exhibited lower-levels of sequence variations due to the influence of negative selection (Table 2). SLPTX10 and SLPTX15 families were reported to have undergone a functional radiation, where members exhibit neurotoxicity by targeting various voltage-gated ion channels: calcium (Cav), potassium (Kv) and sodium (Nav) ion channels [28]. Similarly, certain SLPTX11 family members are known for their anticoagulant and Kv channel inhibitory activities [28, 43]. We found that even these putatively potent toxins in centipedes were extremely well conserved under the influence of purifying selection (Table 2).
While SLPTX family 13 appears to have convergently adopted an inhibitory cysteine knot (ICK) scaffold, which is characteristic of various potent toxins from scorpions and spiders, SLPTX16 has adopted a Von Willebrand factor type C (VWC)-like domain. These peptides were highly conserved despite their putative role in prey envenoming and long evolutionary histories. Certain taxonomically restricted toxin families, called ‘novel families’ were recently reported in centipede venoms [28]. Only one (‘novel family 6’) amongst the four of these novel families examined was found to have evolved rapidly, while the rest were negatively selected (Table 3). Overall, centipede venom-encoding genes were found to have evolved under the heavy constraints of purifying selection (Fig 2A).
Spiders are known to have originated 416–359 MYA in the Devonian [44]. All spiders, with the exception of a few species, employ venom for predation. However, toxinological research to date has solely focused on characterizing venom from the medically significant species of spiders. Yet, venom from only 0.4% of the currently cataloged spider species have been characterized to date [25]. We determined the rate of evolution of several venom protein superfamilies in a diversity of spider lineages, such as the lethal latrotoxins secreted by widow spiders [Theridiidae: 223–180 MYA [45]]; Kunitz-type serine protease inhibitors and huwentoxins from tarantulas [Theraphosidae: 250–200 MYA [46]]; the magitoxin family from tarantulas and certain funnel-web spiders [Hexathelidae: 250–200 MYA [46]]; sphingomyelinase-D (SMase D) in the medically significant venoms of recluse spiders [Sicariidae: 145+ MYA [46]]; lycotoxin family [47] from wolf spiders [Lycosidae: ~120+ MYA [46]]; and super family E ICKs [48] secreted by tarantulas and brushed trapdoor spiders [Barychelidae: 250–200 MYA [46]]. These venom proteins are secreted in large amounts by the respective spider lineage and are known for a diversity of biochemical activities, such as insecticidal presynaptic neurotoxicity and the ability to stimulate neurotransmitter secretions [latrotoxins: [49, 50]], dermonecrotic properties [SMase D: [51]], Nav channel targeting capability–with some members additionally capable of targeting Cav channels [huwentoxin-1 family: [52, 53]], serine protease inhibition and the ability to block Kv channels [Kunitz toxins: [54]], insect Nav channel targeting [magi-1 family [55]], insect Cav channel targeting [ω-hexatoxins: [56]], insect Calcium activated potassium channel (KCa) targeting [κ-hexatoxins: [57]], and the Nav modulation and Cav blocking capabilities [Super Family E ICKs: [48]]. The computed ω values suggested a greater influence of purifying selection on nine out of ten toxin families examined, highlighting the slower evolution of spider venoms (S1 Table; Fig 2B).
Computed ω values for the vast majority of venom-encoding genes in all ancient lineages examined in this study and in previous studies [18, 26, 38, 58], highlighted the significant role of negative selection, which was in stark contrast to those of evolutionarily younger lineages, such as the advanced snakes and cone snails [S1 Table; [6, 7, 59–62]]. We also evaluated the molecular evolution of venom families from Toxicofera lizards that originated ~166 MYA [63], and thus represent an intermediate state between ancient and recently originated lineages. Although, relative to advanced snakes, these lizards do not rely on venom for predation or defense to the same degree [6], the evolutionary rates of some of their largely secreted venom proteins exhibited rapid evolution as demonstrated by their high number of positively selected sites (Fig 3; S1 Table).
Three (Kallikreins, CRiSPs and crotamines) among the six gene families examined exhibited an evidence for rapid evolution (ω>1 and/or more than 10 positively selected sites), while the remaining were found to be extremely well conserved (S1 Table). Further, we plotted site-wise ω against their respective amino acid position for each of the genes examined. Results indicated that a majority of sites in most venom proteins of ancient lineages evolved under the strong influence of negative selection (Figs 2–5).
In contrast, a large proportion of sites in toxins of evolutionarily young lineages rapidly mutated under the significant influence of positive Darwinian selection (Fig 6). Thus, a stark difference was found in the evolutionary regimes of ancient and evolutionarily young lineages (Fig 7).
Using the mixed effects model of evolution (MEME) several sites that experienced short periods of diversifying selection were also identified in all the examined venomous clades, which indicated that certain sites in these toxin proteins undergo episodic adaptation (Tables 1–3; S1 Table). Considering the long evolutionary histories of these toxin types, we tested for nucleotide substitution saturation (see methods). These tests did not detect saturation in any of the examined datasets (S2 Table).
We performed regression analyses to evaluate the possibility of the length of the toxin determining its rate of evolution by plotting ω values for various toxin types against their respective lengths (S1 Fig). The coefficient of determination (r2) for toxin types in each of the examined lineages suggested an absence of correlation between the length of the toxin and its ω value, indicating that venom proteins have undergone rapid evolution or extreme sequence conservation irrespective of their size. While most conotoxins are of relatively shorter lengths, snake venom components such as three-finger toxins (3FTxs) and Snake Venom Metalloproteinases (SVMPs) are characterized by lengths of 80 and 600 amino acids, respectively. Despite such stark size differences, these toxins evolved rapidly. Similarly, several venom components in ancient lineages were characterized by a range of lengths. For example, most sea anemone and scorpion neurotoxins were of relatively shorter lengths (40–60 amino acids), while several pore-forming toxins were 450–550 amino acids long. Yet, these toxins were found to have evolved extremely slowly under the influence of purifying selection. Our results thus indicated that the stark differences in ω values for venom proteins of ancient and evolutionarily younger lineages did not result from the differences in size.
Animal venoms are assumed to rapidly diversify under the unabated influence of positive Darwinian selection. They have been theorized to undergo a chemical arms race with prey animals, where the evolution of venom resistance in prey and the invention of efficient toxins in the predatory venomous animal exert reciprocal selection pressure [64], as postulated in the Red Queen hypothesis of Van Valen [65]. While the influence of positive selection is widely recognized, the role of purifying selection in shaping animal venoms has rarely been considered. Investigation of a large number of toxin-encoding gene families in this study has revealed a significant influence of negative selection on venom. Whilst positive selection increases the diversity of venom proteins, purifying selection probably aids in preserving the potency of the venom by filtering out mutations that negatively affect toxin efficiency. However, rare mutations that increase the potency of the venom arsenal (e.g., evolution of novel biochemical activity or increased binding efficiency) are likely to be propagated and preserved in the population. In the absence of a conservatory evolutionary force, neutral or positive selection could modify key residues and result in the reduction of potency or, for worse, the complete loss of bioactivity, which could severely decrease the fitness of the animal. Thus, purifying selection pressure appears to be vital for sustaining the potency and, consequently, shaping the animal venom arsenal.
It has been recently demonstrated that PFTs in Cnidaria, which bind to cell membranes and punch holes, evolve under the heavy constraints of negative selection [26, 66]. The lack of variation in this group of toxins, which includes several unrelated toxin types (e.g., aerolysin-related toxins in sea anemones, independently recruited hydralysins in hydroids, actinoporins and jellyfish toxins), was theorized to be a result of their complex multi-subunit packaging [67] and their ability to attack highly conserved molecular targets, such as cell membranes [26]. Toxins that undergo oligomerization in other classes of animals have also been noted to evolve relatively slowly as a result of structural constraints like the need to conserve sites involved in subunit interaction. While most 3FTxs in snake venoms diversify rapidly, κ-3FTxs, which undergo dimerization, were found to accumulate relatively fewer variations [59]. Similarly, toxins that may function in a ‘non-specific’ manner may also experience negative selection. Here, non-specificity of action is defined as the ability to target regions in a structural/biochemical property dependent (e.g., surface electrostatic charge) and target motif independent manner. For example, cytotoxic 3FTxs and β-defensin toxins—two very potent snake venom proteins, induce cytotoxicity by non-specifically binding to negatively charged cell membranes using hydrophobicity [68] and positively charged molecular surface [69], respectively. As a result, unlike most snake venom components, these proteins remain evolutionarily constrained [59, 62]. Similarly, scorpion lipolytic toxins were also theorized to be evolutionarily constrained because of their non-specific mechanism of action [58]. We found that β-PFTs in centipede venoms, which are similar to the aerolysin-like toxins, evolve under the significant influence of negative selection (Table 1). The lack of variation in this group of toxins may suggest that they either undergo oligomerization like their aerolysin homologues in other lineages or the possibility that they may employ a non-specific mechanism of action. A plot of site-specific ω against their respective amino acid positions reveals the extreme conservation of such toxin types that employ unique strategies for causing toxicity in envenomed animals (S2 Fig). As it allows the targeting of a wide variety of animals, the strategy of exerting toxic action non-specifically or by targeting highly conserved molecular sites, appears to be advantageous and follows a contrastingly different evolutionary regime in comparison to toxins that specialize in attacking highly plastic molecular receptors.
A comparison of evolutionary regimes of ancient and evolutionarily younger lineages suggests a fascinating strategy of venom evolution. When venomous animals venture into novel ecological niches, they encounter new types of prey and predatory animals. Consequently, in order to adapt and conquer niches, they would need to fine-tune venom proteins to efficiently target these new animals. Several sites detected as episodically adaptive—i.e., sites that experience short bursts of adaptive selection, in these ancient clades may be reflective of such shifts in ecology. We propose that these earlier periods in the evolutionary history of a venomous species are accompanied by the significant influence of diversifying selection on the venom arsenal, which would expand the range of target sites and/or result in the origination of novel biochemical activities. This is particularly advantageous, since novel toxins generated may facilitate the efficient and rapid incapacitation of newly encountered prey and predatory animals. The period of expansion is followed by longer periods of purification, where the significant influence of negative selection preserves the potency of the toxin. Whenever there is a major shift in ecology or environment, the aforementioned stages of evolution repeat. Thus, we propose that venom-encoding genes mostly employ a ‘two-speed’ mode of evolution, where episodic diversifying selection accompanies the earlier stages of ecological specialization (e.g., diet and range expansion), resulting in the rapid diversification of the venom arsenal, followed by a longer period of purification and fixation that ensure the sustainability of venom potency. The low sequence variation in venom-encoding genes of ancient clades could be reflective of such long periods of purification and fine-tuning. In contrast, advanced snakes and cone snails, being evolutionarily very young, could still be undergoing the period of expansion and, consequently, exhibit a pronounced signature of positive Darwinian selection.
However, it should be noted that the ‘two-speed’ model of evolution is likely applicable to venoms that serve predominantly predatory roles. Due to limited toxin sequence information from venoms that are employed for non-predatory functions (e.g., intraspecific competition in platypus, exclusively defensive roles in fishes, etc.), it remains to be seen whether they too follow our proposed evolutionary model.
To conclude, in addition to unraveling the evolutionary regimes of toxin families in centipedes and spiders, which are amongst the first terrestrial venomous lineages, our findings highlight the pivotal roles of purifying and episodic selections in shaping animal venoms. Our findings enabled the postulation of a new theory of venom evolution in the animal kingdom that emphasizes the dynamic nature of these complex biochemical cocktails.
Toxin homologues were identified in the recently published genome of the coastal European centipede Strigamia maritima [70] by querying amino acid sequences of each toxin type against all six reading frames using the tblastn tool [71].
Translated nucleotide sequences were aligned using MUSCLE 3.8 [72]. The best-fit model of nucleotide substitution for individual toxin datasets was determined according to the Akaike’s information criterion using jModeltest 2.1 [73] and model-averaged parameter estimates were used for the reconstruction of trees. Phylogenetic trees were built using PhyML 3.0 [74], where node support was evaluated with 1,000 bootstrapping replicates. Maximum-likelihood (ML) models [75] implemented in Codeml of the PAML package [76] were utilized to identify the influence of natural selection on toxin families[6]. As no a priori expectation exists, we compared likelihood values for a pair of models with different assumed ω distributions: M7 (β) versus M8 (β and ω) [77]. Only when the alternate model (M8) shows a better fit than the null model (M7) in the likelihood ratio test (LRT), are its results considered significant. LRT is estimated as twice the difference in ML values between the nested models, and is compared with the χ2 distribution with the appropriate degree of freedom—the difference in the number of parameters of the two models. Further, we used the Bayes empirical Bayes (BEB) approach [78] in M8 to detect amino acids under positive selection by calculating the posterior probability (PP) that a particular site belongs to a given selection class (neutral, conserved, or highly variable). Sites with PP ≥ 95% of belonging to the ‘‘ω > 1 class” are inferred to be positively selected. HyPhy’s [79] FUBAR approach [80] was used to detect sites evolving under pervasive diversifying and purifying selection pressures. MEME [81] was also employed to identify episodically diversifying sites. Sequence alignments used for selection assessments have been made available as a zipped file (S1 File; see S3 Table for accession list). Nucleotide substitution saturation was tested using DAMBE 5.5.9 [82] using the recommended protocol [83].
|
10.1371/journal.pbio.1001473 | The Oxytricha trifallax Macronuclear Genome: A Complex Eukaryotic Genome with 16,000 Tiny Chromosomes | The macronuclear genome of the ciliate Oxytricha trifallax displays an extreme and unique eukaryotic genome architecture with extensive genomic variation. During sexual genome development, the expressed, somatic macronuclear genome is whittled down to the genic portion of a small fraction (∼5%) of its precursor “silent” germline micronuclear genome by a process of “unscrambling” and fragmentation. The tiny macronuclear “nanochromosomes” typically encode single, protein-coding genes (a small portion, 10%, encode 2–8 genes), have minimal noncoding regions, and are differentially amplified to an average of ∼2,000 copies. We report the high-quality genome assembly of ∼16,000 complete nanochromosomes (∼50 Mb haploid genome size) that vary from 469 bp to 66 kb long (mean ∼3.2 kb) and encode ∼18,500 genes. Alternative DNA fragmentation processes ∼10% of the nanochromosomes into multiple isoforms that usually encode complete genes. Nucleotide diversity in the macronucleus is very high (SNP heterozygosity is ∼4.0%), suggesting that Oxytricha trifallax may have one of the largest known effective population sizes of eukaryotes. Comparison to other ciliates with nonscrambled genomes and long macronuclear chromosomes (on the order of 100 kb) suggests several candidate proteins that could be involved in genome rearrangement, including domesticated MULE and IS1595-like DDE transposases. The assembly of the highly fragmented Oxytricha macronuclear genome is the first completed genome with such an unusual architecture. This genome sequence provides tantalizing glimpses into novel molecular biology and evolution. For example, Oxytricha maintains tens of millions of telomeres per cell and has also evolved an intriguing expansion of telomere end-binding proteins. In conjunction with the micronuclear genome in progress, the O. trifallax macronuclear genome will provide an invaluable resource for investigating programmed genome rearrangements, complementing studies of rearrangements arising during evolution and disease.
| The macronuclear genome of the ciliate Oxytricha trifallax, contained in its somatic nucleus, has a unique genome architecture. Unlike its diploid germline genome, which is transcriptionally inactive during normal cellular growth, the macronuclear genome is fragmented into at least 16,000 tiny (∼3.2 kb mean length) chromosomes, most of which encode single actively transcribed genes and are differentially amplified to a few thousand copies each. The smallest chromosome is just 469 bp, while the largest is 66 kb and encodes a single enormous protein. We found considerable variation in the genome, including frequent alternative fragmentation patterns, generating chromosome isoforms with shared sequence. We also found limited variation in chromosome amplification levels, though insufficient to explain mRNA transcript level variation. Another remarkable feature of Oxytricha's macronuclear genome is its inordinate fondness for telomeres. In conjunction with its possession of tens of millions of chromosome-ending telomeres per macronucleus, we show that Oxytricha has evolved multiple putative telomere-binding proteins. In addition, we identified two new domesticated transposase-like protein classes that we propose may participate in the process of genome rearrangement. The macronuclear genome now provides a crucial resource for ongoing studies of genome rearrangement processes that use Oxytricha as an experimental or comparative model.
| Oxytricha trifallax is a distinctive ciliate [1]—an ancient lineage of protists named for their coats of cilia. Like all ciliates, Oxytricha has two types of nuclei: a micronucleus, a germline nucleus that is largely transcriptionally inactive during vegetative growth, and a macronucleus, which is the transcriptionally active somatic nucleus [2]. Compared to the micronucleus, Oxytricha's macronucleus is massively enlarged due to ∼2,000-fold [2] amplification resulting from two rounds of DNA amplification [3] during development.
In the model ciliates Oxytricha trifallax, Tetrahymena thermophila, and Paramecium tetraurelia, varying amounts of micronuclear DNA are deleted (including the “internally eliminated sequences,” or IESs, interspersed between “macronuclear destined sequences,” or MDSs) during conjugation or autogamy (two forms of sexual development) to give rise to the information-rich macronuclear genome (Figure 1). A much larger fraction of the Oxytricha micronuclear genome—∼96% of the micronuclear complexity [2]—is eliminated during the macronuclear formation than in the oligohymenophoreans, Tetrahymena and Paramecium (which both eliminate ∼30% of their micronuclear genomes [4],[5]). The most remarkable difference in macronuclear development between Oxytricha and the two oligohymenophoreans is that the micronuclear-encoded MDSs that give rise to the macronuclear chromosomes may be nonsequential, or even in different orientations in the micronuclear genome [6]. Consequently, unlike the oligohymenophoreans, Oxytricha needs to unscramble its micronuclear genome during macronuclear development.
Two fundamental differences distinguish Oxytricha's macronuclear chromosomes from those of Tetrahymena and Paramecium: Oxytricha's chromosomes are tiny (“nanochromosomes,” with a mean length ∼3.2 kb reported in this study), each typically encoding just a single gene with a minimal amount of surrounding non-protein-coding DNA [7], and they are differentially amplified (Figure 2 and 3) [8],[9]. In some cases, alternative fragmentation of macronuclear-destined micronuclear DNA produces different nanochromosome isoforms (Figure 2) 10–12, which may be present at very different levels of amplification (differing by as much as 10-fold [13]). Gene expression and nanochromosome copy number may be moderately correlated [14]. Macronuclear chromosomes in all the model ciliates segregate by amitosis during cellular replication (without a mitotic spindle) [15],—a process that may lead to allelic fixation 17–20. In ciliates with nanochromosomes, major fluctuations of nanochromosome copy number [8] may arise, since copy number is unregulated during normal cellular replication [21]. Theoretical models propose that these fluctuations are a cause of senescence in these ciliates [22]. In contrast to the lack of copy number regulation during cellular replication, both genetic [23]–[25] and epigenetic mechanisms [9],[26] may influence chromosome copy number during sexual development in ciliates.
As a consequence of the extensive fragmentation of the Oxytricha macronuclear genome, each macronucleus possesses tens of millions of telomeres, an abundance that enabled the first isolations of telomere end-binding proteins [27],[28]. Oxytricha also has micronuclear transposons bearing telomeric repeats (C4A4) that resemble those of nanochromosomes. These telomere-bearing elements, or TBE transposons [29], play an important role in macronuclear genome development [30]. The exact site of telomere addition may vary for some nanochromosome ends [31] and is followed by a roughly 50 bp subtelomeric region of biased base composition with an approximately 10 bp periodicity of the bias (Figure 3) [32],[33].
We report on the assembly and analysis of the first Oxytricha macronuclear genome, from the reference JRB310 strain. During and after assembly, we have addressed a number of challenges arising from the unusual structure of this genome, which we discuss. We focus on the most interesting and unique biological characteristics of this genome and place them in the context of the characteristics of the other sequenced ciliate macronuclear genomes.
To assemble the Oxytricha macronuclear genome for the type strain—JRB310 [1]—we chose to build upon three assemblies, from ABySS [34], IDBA [35], and PE-Assembler [36]/SSAKE [37], based on Illumina sequences, and supplemented by a Sanger/454 assembly. To combine these assemblies, we developed a specialized meta-assembly pipeline (see Materials and Methods and Figure S1). Current genome assembly strategies for second-generation sequence data often employ multiple, hybrid strategies to overcome the experimental biases leading to low sequence coverage in particular genomic regions and repetitive DNA [38]. Since Oxytricha's macronuclear genome was expected to have a low repeat content [2], repetitive DNA was expected to be a relatively insignificant issue, and thus even greedy genome assemblers were able to produce useful preliminary assemblies. However, unlike conventional genomes, the Oxytricha macronuclear genome provides assembly challenges by virtue of its fragmented architecture, variable processing (“alternative chromosome fragmentation”), and nonuniform nanochromosome copy number. We resolved these challenges during and after assembly.
The initial 454/Sanger genome assemblies contained a mixture of bacterial DNA, mitochondrial DNA, and up to two additional alleles other than those expected from the strain we originally proposed to sequence (JRB310) due to accidental contamination by a commonly used strain in our lab (JRB510—a complementary mating type), whereas the Illumina assemblies were produced from purified macronuclear DNA from the type strain (JRB310) alone. Given these contamination issues, we built our final assembly with the Illumina assemblies as the primary data source, rather than the 454/Sanger assemblies, to maintain the purity of our final assembly. This excluded virtually all bacterial and mitochondrial contamination in our final assembly since very few contigs in the Illumina assemblies derive from these sources (sucrose gradient purification of macronuclei eliminated almost all such contaminants) that could potentially be extended by the 454/Sanger data. We also kept JRB510 allelic data to a minimum in our final assembly, (i) by preferring best extensions, which were most likely to be from the more similar JRB310 derived contigs or reads, either from the Illumina assemblies or from the Sanger/454 assemblies and raw Sanger data (see Materials and Methods), and (ii) by the sequence consensus majority rule (the most abundant base at each position from the assembled contigs) of the CAP3 assembler [39] used to combine the three Illumina assemblies versus one 454/Sanger assembly during contig construction. The conclusion that our final assembly strategy succeeded in keeping JRB510 allelic information to a minimum is supported by matches to the three pure JRB310 Illumina starting assemblies. Our final assembly is 82.0% covered by identical BLAT matches ≥100 bp long to one of these pure JRB310 assemblies and 92.5% covered by 99.5% identity, ≥100 bp BLAT matches to these assemblies (note that consensus building by CAP3 may result in alternating selection of JRB310 polymorphisms from the original assemblies, and hence even meta-contigs assembled from the pure JRB310 assemblies may have BLAT matches that differ from all the original assemblies).
We chose to keep alleles apart by applying moderately strict criteria for merging during our meta-assembly (e.g., by merging contigs with overlaps at least 40 bp long and ≥99% identical with CAP3 [39]; see Materials and Methods). However, in order to maximize the number of complete nanochromosomes in our assembly, we collapsed some alleles (i.e., producing “quasi-nanochromosomes” derived from two alleles; see “Extensive Genome Homozygosity and High SNP Heterozygosity”). Merging of contigs is also complicated by alternative fragmentation, which affects ∼10% of the nanochromosomes and may either result in collapsing or splitting of nanochromosome isoforms (see “Extensive Alternative Nanochromosome Fragmentation”).
We discriminated between homozygous and heterozygous nanochromosomes after assembly (see the next section). We have not attempted to determine the haplotypes of the heterozygous contigs due to computational complications arising from both alternative nanochromosome fragmentation and variable representation of alleles (which need not be 1∶1).
A comparison of the Oxytricha macronuclear genome assemblies and meta-assembly is given in Table 1 (also see Tables S11–S17 for the progressive improvements in the genome assembly through the successive steps of our meta-assembly approach shown in Figure S1). Since the size selection used in the construction of our paired-end (PE) sequence library results in poor sequence coverage for a span of approximately 160 bp, roughly 100 bp from the telomeric ends (see Materials and Methods), the incorporation of single-end (SE) sequence data allowed ABySS to assemble far more contigs with telomeric sequences than either IDBA or the PE-Assembler/SSAKE assembler combination, neither of which could use SE and PE sequences simultaneously (Table 1). The ABySS assembly is larger (78.0 Mb) than the other assemblies (47.8 Mb for PE-Asm/SSAKE and 57.7 Mb for IDBA) and also more complete, as evidenced by a higher fraction of reads that can be mapped to this assembly (Table 1). The ABySS assembly also incorporates a substantially higher proportion of telomeric reads than the other two assemblers (91.8% for ABySS versus 45.4% for PE-Asm/SSAKE and 42.4% for IDBA) and contains a larger proportion of telomere-containing contigs (66.5% versus 18.3% for PE-Asm/SSAKE and 8.6% for IDBA) and longer contigs (mean length of 2,273 bp for ABySS versus 2,090 bp for PE-Asm/SSAKE and 1,204 bp for IDBA). Consequently, the ABySS assembler produced almost an order of magnitude more full-length nanochromosomes than the other two assemblers. Even though the ABySS assembly incorporates more telomeric reads than the other assemblies, it also excludes a higher proportion of telomeric reads than nontelomeric reads (92% telomeric PE reads versus 98% total PE reads map to the assembly; telomeric reads comprise ∼13% of all the reads). While the ABySS assembly appears to be the most complete, the majority of its contigs (81.3%) are still missing one or both telomeres.
The initial meta-assembly of the ABySS, IDBA, and PE-Asm assemblies yielded a modest improvement in the total number of full-length nanochromosomes relative to ABySS alone, with the ratio of full-length nanochromosomes to contigs increasing from 21% to 24% of the total number of contigs. Since the meta-assembly was still highly fragmented, and our aim was to assemble full-length nanochromosomes with complete genes, we developed a strategy that consisted of two rounds of extension of nontelomeric contig ends and reassembly (see Materials and Methods). This strategy produced an assembly where the majority (77%) of the contigs had both 5′ and 3′ telomeres. For the final meta-assembly, the average contig length (2,982 bp) is longer than that of any of the original assemblies (2,273 for ABySS versus 2,090 for PE-Asm/SSAKE and 1,204 for IDBA) and the read coverage is as high as or higher than the most complete ABySS assembly (98.0% coverage for PE reads for ABySS and the final assembly; 88.2% SE read coverage for ABySS and 88.5% SE read coverage for the final assembly). However, a larger fraction of telomeric reads than nontelomeric reads still do not map to the final assembly (e.g., 9.0% of telomeric PE reads versus 2.0% of all PE reads do not map to the assembly), indicating that some telomeric regions are still missing from the final assembly.
Since there are currently no published effective population size estimates for Oxytricha trifallax, we wanted to obtain an estimate from allelic diversity of the macronuclear genome. Furthermore, current estimates of effective population size for other free-living model ciliates, Paramecium and Tetrahymena, differ [40]–[42], so additional estimates from other species will be necessary to determine if there are general trends in population size within ciliates.
Given our assembly conditions, we expect many allelic sequences to co-assemble, but visual inspection of reads mapped to the final assembly suggested that a substantial fraction of the genome is homozygous. A trivial explanation for this observed homozygosity is that the micronuclear genome regions (MDS) that form the nanochromosomes are homozygous, however it is also possible that a combination of other factors may be responsible for some of the observed homozygosity. These factors include both nanochromosomal allelic drift, arising from stochastic nanochromosome segregation during amitosis (during normal cellular replication) and nanochromosomal allelic selection, both of which could lead, in principle, to haplotype fixation (also known as “allelic assortment”), as well as allelic biases introduced during conjugation.
Some well-studied macronuclear nanochromosomes, including the 81 locus [43], and the nanochromosomes encoding the telomere end-binding proteins α and β (Contig22209.0 and Contig22260.0) are homozygous in the Oxytricha JRB310 strain upon which our reference macronuclear genome assembly is based. Knowledge of the fraction of homozygous and heterozygous nanochromosomes is also necessary to obtain a reasonable estimate of macronuclear genome size. To determine nanochromosomal homozygosity, we focused on nonalternatively fragmented nanochromosomes in order to avoid both ambiguous alignments and possible false identification of heterozygosity due to the presence of alternative telomere locations. Of the nonalternatively fragmented nanochromosomes, 66% (7,487 out of 11,297) had no substantial BLAT [44] nonself matches (≥100 bp and ≥90% identical; default BLAT parameters) to any other contig in the final assembly (“matchless” nanochromosomes).
Matchless nanochromosomes with variants present at ≥0.5% of the positions were considered heterozygous (see Materials and Methods for our precise definition of heterozygosity); otherwise they were considered to be homozygous. Note that our read mapping cutoff may reduce polymorphism estimates by filtering out reads with more polymorphisms (see Text S1, “Read Mapping Rationale”). These criteria overestimate low frequency variants at lower coverage sites and underestimate higher frequency variants at lower coverage sites (Figure S2) but comprise a small proportion of all the sites (e.g., 11.8% of sites identified as heterozygous are at 20–40× coverage). By these criteria, the well-characterized Oxytricha actin I [45],[46] nanochromosome (Contig19101.0) was correctly classified as heterozygous, with variants at 0.89% of the examined positions (12/1,350). Mismapped reads do not affect the identification of the heterozygous sites in this nanochromosome, since only two of the reads mapped to this nanochromosome map to any of the other contigs. For the actin I nanochromosome, the frequency of allelic variants varies from just over 5% (three positions) to 13.1% (mean 8.7%) corresponding to a roughly 1∶11 ratio of the minor∶major allele. Of the matchless nanochromosomes, 63% have sequence variants identified at 0%–0.5% of positions; hence, 37% of matchless nanochromosomes are classified as heterozygous by this criterion. Read mapping appears to be adequately sensitive to detect most SNPs (single nucleotide polymorphisms), since the mean number of reads per bp mapped to heterozygous and homozygous nanochromosomes is almost identical (see “Nanochromosome Copy Number Is Nonuniform”).
Approximately 42% of nanochromosomes are homozygous when the proportion of putative homozygous nanochromosomes is calculated from the total number of matchless and matched nanochromosomes. The high levels of macronuclear genome homozygosity agree with preliminary observations of micronuclear sequence data for this strain (Chen et al., unpublished), suggesting that the majority of nanochromosomal homozygosity may derive from homozygosity in the micronuclear genome, rather than other possible factors (allelic assortment and/or developmental biases). Once the micronuclear genome is more complete, it will be possible to assess how much these factors have contributed to the observed homozygosity. Nevertheless, the abundance of homozygous nanochromosomes in the final assembly (∼42%) suggests that the wild-isolate JRB310 [47] may actually be substantially inbred and that this inbreeding arose at its source. Deleterious inbreeding effects may contribute to the complexity of Oxytricha trifallax mating types [47]. It will be interesting to determine whether the two “promiscuous” Oxytricha strains (JRB27 and JRB51 [47]) that mate with the broadest set of other mating types are less inbred.
Sequence polymorphisms are abundant in the Oxytricha macronuclear genome: excluding the homozygous nanochromosomes (which may have arisen from inbreeding), the mean SNP heterozygosity is 4.0% (SD = 1.8%; Figure 4). From mapped reads, for heterozygous matchless nanochromosomes, mean SNP heterozygosity is 3.1% (SD = 1.3%), and for heterozygous matched nanochromosomes, it is 4.6% (SD = 1.9%; Figure 4). For alignments of heterozygous matched nanochromosomes (with BLAT matches ≥100 bp and ≥90% identical to each other) produced by MUSCLE (for nanochromosome pairs where one of the nanochromosomes is no more than 10% longer than the other and for alignments with ≤15% differences), mean SNP heterozygosity is 3.0% (SD = 2.5%). These estimates of SNP heterozygosity indicate that assembly has masked a substantial amount of allelic variation. Similar statistics were obtained for the subset of the final assembly's nanochromosomes that were present in the pure JRB310 strain ABySS assembly (e.g., heterozygous matchless mean SNP heterozygosity is 2.8%, SD = 1.2%, and for MUSCLE alignments of heterozygous matched nanochromosomes, mean heterozygosity is 3.2%, SD = 2.9%). Hence it is unlikely that any residual JRB510 strain allelic information in our final assembly has had a considerable effect upon our estimates of heterozygosity. Nevertheless, these are first estimates from a complex genome assembly, complicated by homozygosity due to potential inbreeding, and hence inferences based upon them (including our subsequent population size estimate) should be treated with caution until better estimates from the micronuclear genome and additional strains become available.
At 4-fold synonymous sites from heterozygous nanochromosomes with matches (see Materials and Methods), the mean SNP heterozygosity is 8.3% (SD = 9.4%). This underestimates sequence diversity at 4-fold synonymous sites, since pairwise alignments of contigs underestimate SNP heterozygosity at all sites (see previous paragraph). If we apply a correction for the missing SNP heterozygosity based on our overall estimate of SNP heterozygosity, we obtain an estimate of 11.1% mean SNP heterozygosity at 4-fold synonymous sites [8.3%×(4.0%/3.0%)]. This 4-fold synonymous site SNP heterozygosity is very high—even higher than the current SNP heterozygosity record holder, Ciona savignyi, which has 8.0% 4-fold synonymous site mean SNP heterozygosity [48]. These high levels of SNP heterozygosity suggest that Oxytricha trifallax has a large effective population size.
Assuming a mutation rate μ ∼10−9 per base per generation, as in Snoke et al. 2006 [42], for nucleotide diversity at 4-fold synonymous sites and π4S = 4Neμ, we estimate an effective population size of 2.6×107. This effective population size is on the same order estimated for P. tetraurelia (using silent site diversity, πS, which will yield a smaller population size estimate than one based on π4S) [42]. However, this estimate of the P. tetraurelia effective population size may be an overestimate due to incorrect classification of species within the Paramecium aurelia species complex and may be closer to the order of 106 [40]. In contrast, the T. thermophila effective population size is estimated to be considerably smaller than Oxytricha's, at Ne = 7.5×105 for πS = 0.003 (with μ = 10−9) [41],[42].
In laboratory culture conditions, Oxytricha trifallax tends to replicate asexually and rarely conjugates (resulting in meiotic recombination). Conjugation in the laboratory is induced by starvation as long as cells of compatible mating types are available. However, we do not know the frequency of conjugation relative to replication of Oxytricha trifallax in its natural environment. The relationships between the frequency of asexual reproduction, and additional population genetic factors arising from asexuality, such as the variance of asexual and sexual reproductive contributions, are complex and can result in increases or decreases in estimates of effective population size [49],[50]. As a result, effective population size estimates for Oxytricha should be treated with caution until these factors are better understood.
As indicated for the well-characterized actin I locus, which has a roughly 1∶11 ratio of minor∶major allelic variant, there may be substantial deviations from the expected 1∶1 ratio for the two possible allelic variants at each site. For matchless nanochromosomes, we found that the distribution of median nanochromosomal variant frequency (i.e., the median frequency of putative allelic polymorphisms) is bimodal, with one mode close to the expectation at 40%–45% and the other at 5%–10% (Figure 5; since the lower peak is bounded by the cutoff we chose to assess variants, the true lower peak may be lower than this). The bimodality of this distribution persists even if we only consider nanochromosomes where the mean coverage of the variant sites is high (e.g., at mean coverage of ≥110× at variant sites). Though some deviation from the 1∶1 variant ratio might result from allele-specific read mapping biases [51], given the relatively relaxed read mapping parameters we used (see Materials and Methods, “Read Mapping and Variant Detection”), the two variant frequency modes differ too much to explain the lower mode's existence. Instead, the deviation from the expected ratio may indicate that allelic assortment has occurred, or that there are developmentally specific allelic biases. Since a high proportion of nanochromosomes deviate substantially from the 1∶1 expected allelic ratio, it is also possible that allelic assortment has occurred for some nanochromosomes, which may contribute to the observed abundance of homozygous nanochromosomes. Nanochromosomes that deviate the most from the expected 1∶1 allelic ratio tend to have lower mean SNP heterozygosities, which likely reflects the diminished ability to detect SNPs with more distorted variant frequencies (Figure 5).
We desired an estimate for the total haploid Oxytricha trifallax macronuclear genome size since it is unknown. To obtain a reasonable estimate, we needed to determine the extent of redundancy in our genome assembly. As judged by visual inspection of our original assemblies, the main sources of redundancy are (i) the two alleles from the partially diallelic genome (see “Extensive Genome Homozygosity and High SNP Heterozygosity”), (ii) alternative nanochromosome fragmentation (see “Extensive Alternative Nanochromosome Fragmentation”), (iii) erroneous base calling that may result from high copy number regions and relatively abundant sequencing errors, and (iv) paralogous genes. The assembly with the most redundancy—from ABySS (Figure S3)—has approximately half of its contigs with nonself matches that are identical or almost identical (matches that are ≥100 bp and ≥99% identical). Visual inspection of the ABySS assembly revealed that much of the redundancy arose from the combined effect of high copy number DNA and sequencing errors. Our assembly strategy eliminated most of the redundancy from erroneous base calling, because it collapsed regions that are nearly identical. A small quantity of additional redundancy may have also been introduced by the inclusion of non-reference (JRB510) allelic sequences from the Sanger/454 genome assemblies, though the strategy we used prefers the inclusion of reference allelic sequences (see Materials and Methods, “Princeton Illumina Assembly and TGI Sanger/454 Assembly Integration”). Though some redundancy remains in our final genome assembly, this is counteracted by ∼1/4 of the nanochromosomes that have had their alleles collapsed during the assembly (see “Extensive Genome Homozygosity and High SNP Heterozygosity”).
Given the ∼42% estimate of nanochromosomal homozygosity, we estimate that the haploid number of nanochromosomes is ∼15,600 [from 15,993+1,279 two- and multitelomere contigs and 5,303 single-telomere contigs (Table 1); we also estimate that ∼10% of nanochromosomes are alternatively fragmented (see “Extensive Alternative Nanochromosome Fragmentation”)]. With a mean nanochromosomal length of ∼3.2 kb, we estimate that the haploid macronuclear genome size is 50 Mb, which is similar to earlier experimental estimates [52],[53].
Traditional assessments of genome completeness are not very meaningful in Oxytricha because they usually measure genomes of uniform coverage with relatively long chromosomes. In two key ways, the Oxytricha macronuclear genome assembly is more similar to a de novo transcript assembly than to a conventional genome assembly: it contains multiple nanochromosome isoforms produced by alternative nanochromosome fragmentation (see “Extensive Alternative Nanochromosome Fragmentation”), and it is an assembly of nonuniformly amplified DNA (see “Nanochromosome Copy Number Is Nonuniform”). Unlike RNA transcripts, nanochromosome levels remain relatively stable during asexual growth [22], and variation of nanochromosome copy number is considerably lower than that of transcripts, so we are able to completely sample the genome's DNA over time.
Simple genome metrics indicate that our assembly is largely complete. Firstly, we have sequenced the genome to a substantial depth: we have >62× haploid coverage of the genome assembly by Illumina 100 bp PE reads and >48× haploid coverage by SE reads. Secondly, nearly all high-quality reads map to our final assembly (98% of high quality PE reads) and the majority of contigs (71.3%) represent complete nanochromosomes, with only 5.1% of the contigs missing both telomeres and 23.6% missing one telomere (Table 1). Finally, our 50 Mb haploid genome assembly size estimate is similar to an earlier estimate of ∼55 Mb for the DNA complexity of the Oxytricha macronucleus [2].
To assess genome completeness we analyzed the completeness of two specific, functionally related gene sets—encoding ribosomal proteins and tRNAs—and one general gene data set in Oxytricha. All of these measures of completeness indicate that the macronuclear genome assembly is essentially complete. Firstly, the final genome assembly contains all 80 of the standard eukaryotic ribosomal proteins (32 small subunit and 48 large subunit proteins). Secondly, the Oxytricha macronuclear genome has a haploid complement of ∼59 unique tRNA nanochromosomes (including a selenocysteine tRNA on Contig21859.0). These tRNAs are sufficient to translate all of its codons if wobble position anticodon rules [54] are accounted for. As judged by searches of tRNAdb [55], codons without cognate tRNAs in Oxytricha are either absent or rare in other eukaryotes. Furthermore, with the exception of a Tetrahymena glycine tRNA that has a CCC anticodon [56], Oxytricha's tRNAs share the same anticodons as Tetrahymena.
We also assessed the completeness of the macronuclear genome by searches of predicted proteins against 248 “core eukaryotic genes” (CEGs: defined by KOGS [57] based on the complete protein catalogs of H. sapiens, D. melanogaster, C. elegans, A. thaliana, S. cerevisiae, and S. pombe [58]). Using a strategy similar to that used to assess completeness during analyses of ncRNAs in Oxytricha [59], and based on the CEGMA analysis strategy [58], we searched for all the core proteins using a match coverage of ≥70% of the mean CEG sequence length, accumulated over the span of the query CEG sequence matches from BLASTP (BLAST+ [60]; with at least one of the matches for each query having an E-value≤1e-10). Of our predicted proteins 231 had substantial sequence similarity to the core eukaryotic protein sequences. The numbers of core eukaryotic proteins we found in the latest Tetrahymena (223) and Paramecium (230) gene predictions were similar to those found in Oxytricha (within the limits of the sensitivity of the searches we used and possible gene prediction failures). However, since CEGS are defined for just five genomes of animals, fungi, and one plant and exclude a diversity of other eukaryotes, the true set of CEGs may be somewhat smaller than this. Given the great evolutionary divergences of ciliates from these eukaryotes (possibly in excess of 1.5 billion years ago [61]), it is also possible that the BLAST criteria employed by CEGMA are not sufficiently sensitive to detect more distant ciliate homologs. This suggests that the predicted proteomes of all three ciliates are largely complete.
Given that the deep divergences of ciliates might prevent detection of their homologs to the remaining 17 CEGS without matches, we attempted more sensitive searches at the domain level using HMMER3 [62]. We assigned Pfam domains to each KOG if the domains were best Pfam hits to the majority of the members of each KOG. Using domain searches, 13 of the remaining 17 core proteins had matches in Pfam-A (with domain, full-sequence E-values<1e-3; see Text S1, “Pfam Domains Detected for CEGs Missing in Oxytricha”). With the exception of KOG2531, these CEGs are relatively short, single-domain proteins.
Of the four undetectable CEGs remaining after HMMER3 searches, one, KOG3285, is an ortholog group corresponding to the MAD2 [63] spindle assembly checkpoint (SAC) protein. We were unable to detect homologs of additional SAC proteins such as MAD1 and MAD3, suggesting that these checkpoint proteins have either been lost or that they are very divergent. The Tetrahymena macronuclear genome paper reported the absence of the checkpoint kinase CHK1 [56], but this is difficult to establish unambiguously given both the considerable divergence of ciliate proteins from the model organisms in which this protein was discovered and that the single defining domain for this protein is the widely distributed and extremely common protein kinase domain (PF00069). There is no evidence of a mitotic spindle in ciliate macronuclei [16], so they may not need SAC proteins, unlike conventional nuclei. It is also possible that the lack of these proteins contributed to the evolution of ciliate amitosis. However, micronuclei do appear to undergo a spindle-guided mitosis [2],[16]. Since ciliates need to coordinate the division of multiple nuclei, their nuclear cycle checkpoints may be more complex than most eukaryotes; hence, they may use genes that are nonorthologous to conventional checkpoint genes in nonciliates.
The other three undetectable CEGs—KOG0563, KOG3147, and KOG2653—correspond to three key oxidative pentose phosphate pathway (OPPP) enzymes: glucose-6-phosphate dehydrogenase (G6PD), gluconolactonase (6PGL), and 6-phosphogluconate dehydrogenase (6PGD), which are also missing in Paramecium and Tetrahymena, and hence may have been lost in ciliates (Figure S4). Two of these enzymes—G6PD and 6PGL—were also noted to be missing in Paramecium, Tetrahymena, and Ichthyophthirius [64]. Thus, excluding these three CEGs that appear to be absent from ciliates, Oxytricha's macronuclear genome is only missing one CEG (KOG3285/Mad2).
Together, since the macronuclear genome has 244/245 (99.6%) of the ciliate-restricted CEGs, it (together with the mitochondrial genome [65]) is likely to encode a complete set of genes required for vegetative growth. This is consistent with the observation of amicronucleate Oxytricha species in the wild, which are capable of vigorous replication for hundreds of generations in culture [2],[66], and with temporary dispensability of the micronucleus. Currently TBE transposon genes are the only published examples of micronuclear-limited genes (not encoded on any of the nanochromosomes in our assembly) in Oxytricha. These genes are exclusively expressed during sexual development and appear to be essential for accurate genome rearrangements [30], and hence may only need to be expressed from the micronucleus.
A characteristic feature of the Oxytricha macronuclear genome is the existence of multiple, stable “versions” of nanochromosomes that share genic regions [10]–[12]. “Alternative processing” or “alternative fragmentation” of DNA is analogous to alternative splicing of introns from pre-mRNAs, but unlike alternative RNA splicing, macronuclear DNA is simply fragmented (with telomere addition), rather than joined together. Variable deletion of micronuclear DNA in Paramecium also gives rise to alternatively fragmented macronuclear chromosomes, though it produces much longer multigene chromosomes and this alternative fragmentation is much less frequent than that in Oxytricha [67]–[69]. Initial surveys of Oxytricha fallax nanochromosomes revealed a substantial amount of alternative nanochromosome fragmentation, with 40% (6/15) of the surveyed nanochromosomes alternatively fragmented [11], so we wanted to assess this on a genome-wide scale. We also sought evidence of possible functional relationships between alternative fragmentation and gene expression, since, in principle, alternative nanochromosome fragmentation may affect gene expression by (i) permitting variable amplification of nanochromosome isoforms, thereby affecting basal transcription levels of the genes encoded on these isoforms (see “Nanochromosome Copy Number Is Nonuniform”), (ii) gene truncation, and (iii) affecting regulation of gene expression by modulating which regulatory elements are present on nanochromosomes.
The creation of contigs during assembly merges shorter alternative nanochromosomes into longer isoforms, obscuring telomeric repeats when they contribute a minority of bases, thus making it difficult to identify the alternative isoforms directly from the contig sequences. Therefore, we exploited two sources of raw sequence data to uncover this kind of variation: 454 telomeric read pairs and Illumina telomeric reads (see Materials and Methods). From alternative fragmentation sites predicted by either data source, almost 1/4 of all the nanochromosomes (3,369/14,390) in our final assembly are predicted to be alternatively fragmented (only counting contigs with terminal telomeres ≤100 bp from either end of the contig). We predict 11% (1,909/17,372) and 14% (2,380/17,372) of nanochromosomes are alternatively fragmented from 454 telomeric reads alone and Illumina telomeric reads alone, respectively. Of the nanochromosomes predicted to be alternatively fragmented by Illumina telomeric reads, 63% are also predicted to be alternatively fragmented by 454 telomeric reads, and 68% of the nanochromosomes predicted to be alternatively fragmented by 454 telomeric reads are also predicted to be alternatively fragmented by Illumina telomeric reads. The actual portion of alternatively fragmented nanochromosomes may be closer to 10% since many of the predicted sites are only supported by a few reads (which results in a poor correspondence between the predictions from the two data sources when there are few telomeric reads at a putative alternative fragmentation site; see Text S1, “Classification of Strongly and Weakly Supported Alternative Fragmentation Sites”). We propose that most of the nanochromosomes arising from weakly supported sites with few supporting telomeric reads (e.g., <9 llumina telomeric reads) may represent “developmental noise” or healing of broken nanochromosomes by capping the broken ends with telomeres, rather than functional nanochromosomes.
We may not have recovered some alternatively fragmented nanochromosome isoforms due to limitations of our genome assembly. Since we focused on nanochromosomes with telomeres at both ends, some alternative fragmentation will be missed on nanochromosomes that lack telomeric ends (e.g., the alternatively fragmented 81-Mac locus [10],[11], represented by Contig13637.0.1, is missing both ends). Another possible failure to detect alternative fragmentation is a consequence of the semigreedy nature of our genome assembly strategy, since we stop extending nanochromosomes once we have detected at least one 5′ and at least one 3′ telomeric repeat (see Materials and Methods), which means that we may miss some longer unfragmented nanochromosome isoforms. Consequently, we consider our estimates of the level of alternative fragmentation to be conservative.
Alternative fragmentation sites tend to map between predicted genes in intergenic regions rather than within intragenic regions [Table S1; we use inter-CDS regions rather than intergenic regions since CDS (coding sequence) predictions are more reliable than UTRs (untranslated regions)]. For contigs with single internal alternative telomere fragmentation sites, strongly supported alternative fragmentation sites are 58 times more likely to be located in inter-CDS regions than in intra-CDS regions (per bp of these sequence regions). For nanochromosomes with single-gene predictions, strongly supported alternative fragmentation sites are 27 times more likely to reside within non-CDS regions (i.e., introns, UTRs, subtelomeric regions, or regions with no predicted gene), than within CDSs (Table S2). Strongly supported, noncoding alternative fragmentation sites typically have more telomere-containing reads than do coding alternative fragmentation sites for both single (mean 213 versus 95 reads) and two-gene [mean 186 (intergenic region) versus 116 reads] nanochromosomes.
For strongly supported alternative fragmentation sites predicted by Illumina telomeric reads, 74% (1,208/1,622) of alternatively fragmented nanochromosomes have one site of alternative fragmentation (giving rise to two nanochromosome isoforms: a long unfragmented form and a shorter fragmented isoform), and 21% of alternatively fragmented nanochromosomes have two sites of alternative fragmentation (similar statistics were obtained from strongly supported sites predicted from 454 telomeric reads). This means that typically only a few possible nanochromosome isoforms are produced for each of our assembled nanochromosomes and also that most alternative fragmentation is “directional,” giving rise to only one of the two possible single shorter isoforms. The most extreme example has seven alternative fragmentation sites predicted from the Illumina telomeric reads (at most nine from 454 telomeric reads) for Contig14329.0 (GenBank Accession: AMCR01001519.1).
The observation of directional alternative fragmentation suggests there must either be differential amplification of particular isoforms or degradation of specific forms following excision. The higher amplification levels of alternatively fragmented nanochromosomes relative to nonalternatively fragmented nanochromosomes (see next section) provides support for the first model but does not exclude the second. In the future, it will be interesting to determine how these fragmentation signals relate to chromosome amplification, since the timing of DNA fragmentation correlates with nanochromosome copy number in Euplotes [25].
The longest isoform of the most extreme case of alternative fragmentation we discovered (Contig14329.0) is about 8 kb long with eight distinct protein-coding regions. This contig has 15 predicted telomere addition sites (TASs) (nine 5′ and six 3′ sites relative to the contig orientation in the assembly) from the 454 telomeric reads (with 11 strongly supported sites, including the two terminal sites), giving rise to up to 14 distinct nanochromosome isoforms from the same 8.1 kb region (Figure 6). An alternative fragmentation site at ∼6,000 bp is weakly supported by 454 telomeric reads but strongly supported by Illumina telomeric reads, suggesting that this site largely gives rise to longer nanochromosome isoforms that the 454 telomeric reads are less likely to detect. Every one of the alternative fragmentation sites predicted from Illumina telomeric reads, with the exception of a single weakly supported site at 4,767 bp, was corroborated by 454 telomeric reads within 100 bp of the site. The 454 telomeric reads suggest that each of the seven intergenic regions in this contig is a site of alternative fragmentation (no Illumina telomeric reads map to the site between genes 5 and 6). Consistent with the genome-wide pattern, this contig's alternative fragmentation sites typically reside between, not in the middle of, genes. For the 454 telomeric reads, only a small portion (2/15 sites) of the fragmentation sites are predicted to be in coding regions, and these sites are weakly supported, whereas the Illumina telomeric reads do not predict any sites in coding sequence regions.
To experimentally validate the predicted extreme fragmentation of Contig14329.0, we performed Southern hybridization (Figure S5) on the same vegetative Oxytricha JRB310 macronuclear DNA sequenced by Illumina. With two exceptions, our Southern analysis confirmed all tested nanochromosomes and identified four novel isoforms: the full length ∼8 kb isoform A, and isoforms P, Q, and R (Figure 6, Figure S5; Text S1, “Examination of Discrepancies Between Predicted and Experimentally Determined Alternative Fragmentation Isoforms of the Highly Fragmented Contig14329.0”). Since the process of generating 454 telomeric reads included a size selection (see Text S1, “Whole Nanochromosome Telomere-Based Library Construction”), it is unsurprising that sequencing missed the longer isoforms we were able to detect by Southern hybridization (A, P, Q, and R, at 8.1 kb, 6 kb, 6.5 kb, and 4.5 kb long, respectively).
The eight genes encoded on these alternatively fragmented nanochromosomes are (1) an RNAse HII domain containing protein (Pfam: PF01351), (2) a dsDNA-binding domain (PF01984) protein, (3) a Tim10/DDP family zinc finger domain protein (PF02953), (4) a protein with no significant BLASTP (to GenBank NR; E-value<1e-3) or Pfam matches (E-value<1e-3), (5) a COPI-associated protein domain (PF08507) protein, (6) an uncharacterized conserved protein (DUF2036) protein (PF09724), (7) another protein with no significant BLASTP or Pfam matches, and (8) a translation initiation factor eIF3 subunit (PF08597) protein. From the domain annotations, no obvious functional relationship amongst these genes is evident. From Figure 6, it can be seen that the representation of genes 2, 3, 4, and 8 in the different nanochromosomal isoforms is greater than for the remainder of the genes. Two of the shortest Oxytricha proteins (encoded by genes 2 and 3; see also Text S1, “Analysis of Short Protein and ncRNA-Encoding Nanochromosome”) are encoded on the most abundant nanochromosomal isoforms. Remarkably, in contrast to the surrounding, heterozygous DNA encoding genes 1–3 and gene 8, the ∼4.2 kb DNA region encoding genes 4–7 appears to be completely homozygous, suggesting the possibility that these regions derive from different micronuclear sources.
In contrast to the oligohymenophorean ciliates, which typically have uniformly amplified macronuclear genomes [56],[69], there is considerable variation in nanochromosome copy number in Oxytricha. The distribution of copy number for nonalternatively fragmented nanochromosomes is right-skewed and is restricted around a mean relative copy number of 0.94, with ∼90% of the nanochromosomes contained within a relative copy number range of 0.12–1.76 centered on the mean (Figure 7A). It is possible that some lower copy number nanochromosomes may not have completely assembled since the combined depth of sequence coverage is <120× and lower bound copy number estimation is constrained by the >62× coverage of the PE reads. Mindful of these limitations, within the sequenced JRB310 clonal population of cells, nanochromosome copy number does not appear to vary as much as gene transcription. The most highly amplified nanochromosome, encoding the 18S, 5.8S, and 28S rRNA (Contig451.1), has a copy number that is ∼56× the mean of nonalternatively fragmented nanochromosomes, yet its transcripts typically yield more than 90% of the RNA in our non-poly(A)-selected RNA-seq samples. There is a roughly 2-fold difference between the most highly amplified nanochromosome and the next most highly amplified nanochromosome, encoding the 5S rRNA (Contig14476.0/Contig17968.0; quasi-allelic contigs).
Since the method we used to estimate nanochromosome copy number combines reads from both possible alleles for heterozygous nanochromosomes, it is necessary to map the reads sensitively to avoid exclusion of reads and to minimize incorrect mapping in order to obtain accurate estimates (see “Genome Homozygosity and SNP Heterozygosity”). Our mapping procedure seems to be appropriate for matchless nanochromosomes, since there is no substantial difference in copy number distributions for homozygous and heterozygous matchless nanochromosomes (mean copy number of 0.93, SD = 0.61, and 0.97, SD = 0.67, respectively; Figure 7A). However, for heterozygous nanochromosomes with matches, the mean nanochromosome copy number is lower (0.81; SD = 0.59; Figure 7A) than for matchless nanochromosomes. This is likely because some of the nanochromosomes with matches exhibit higher heterozygosity regions than matchless heterozygous nanochromosomes (>6% mean SNP heterozygosity; see “Genome Homozygosity and SNP Heterozygosity”) and the mapping criteria (≥94% read identity to the mapped contig) eliminated some of the more heterozygous reads.
To assess nanochromosome copy number of alternatively fragmented versus nonalternatively fragmented nanochromosomes, we examined the relationship between the number of telomeric reads and the number of nontelomeric reads per bp of the nanochromosomes (see Materials and Methods). We found that there was a good correlation between telomeric reads from either end of the nonalternatively fragmented nanochromosomes (Figure S6B) with r = 0.90. However, there are examples where the number of reads from each nanochromosome end differs substantially (e.g., Contig22209.0 and Contig5780.0 from Table S3). This may indicate the failure to extend the ends of some nanochromosomes completely or that the ends derive from the DNA of the nonreference strain (JRB510) and are relatively divergent with few reads from the reference DNA mapped to them (e.g., Contig5780.0). Alternatively, there may be experimental biases that skew the numbers of reads mapped to the two ends (e.g., Contig22209.0, which has JRB310 telomeric reads mapped to the nanochromosome end with fewer reads but no reads extending further, even with relaxed read mapping parameters). The correlation between the number of reads per bp and the number of telomeric reads per nanochromosome is also strong (r = 0.89; Figure S6A), indicating that assessment of telomeric reads alone is appropriate for large-scale analyses of nanochromosome copy number. Furthermore, our estimates of relative nanochromosome copy number, either via reads per bp or the number of telomeric reads per contig, are in good agreement with those obtained by qPCR (Table S3; Figure S7).
For relative nanochromosome copy number measured by telomeric reads, the mean number of telomeric reads per alternatively fragmented nanochromosome with a single (directional) alternative fragmentation site (i.e., only two nanochromosome isoforms) is 2.4 times (885 reads, SD = 768 reads) that of nonalternatively fragmented nanochromosomes (363 reads; SD = 290 reads; K-S one-sided test D = 0.59 and p value<1e-9, with the alternative hypothesis that alternatively fragmented nanochromosome copy number>nonalternatively fragmented nanochromosome copy number; Figure 7B). It follows that the DNA of the shorter alternative nanochromosome isoforms is even more highly amplified than that of nonalternatively fragmented nanochromosomes. The greater amplification of alternatively fragmented nanochromosomes relative to nonalternatively fragmented nanochromosomes supports a model of net overamplification of specific alternatively fragmented nanochromosomes isoforms rather than a model of net destruction. The higher amplification of alternatively fragmented nanochromosomes may indicate a commensal DNA relationship between two genes, arising when one of the genes benefits from the amplification signal of a more highly amplified nanochromosome isoform bearing another gene. This relationship requires no functional association between the genes on alternatively fragmented nanochromosomes, consistent with our general observations (e.g., no specific functional associations between nonribosomal genes and ribosomal genes on alternatively fragmented nanochromosomes).
For nonalternatively fragmented nanochromosomes, the ribosomal protein-encoding nanochromosomes are ∼3.9× more highly amplified than nonribosomal protein nanochromosomes, and tRNA-encoding nanochromosomes are ∼3.6× more highly amplified than non-tRNA-encoding nanochromosomes (Figure 7C; for ribosomal versus nonribosomal nanochromosomes: K-S one-sided test D = 0.64 and p value<1e-9, with the alternative hypothesis that ribosomal nanochromosome copy number>nonribosomal nanochromosome copy number; for tRNA versus non-tRNA nanochromosomes: K-S one-sided test D = 0.62 and p value<1e-6, with the alternative hypothesis that tRNA nanochromosome copy number>non-tRNA nanochromosome copy number). Similarly, the ribosomal protein- and tRNA-encoding nanochromosome isoforms arising from alternative fragmentation are typically overamplified relative to the isoforms that encode other genes (50/54 alternatively fragmented ribosomal nanochromosomes and 25/28 alternatively fragmented tRNA nanochromosomes; Figure S8).
Given the modest variation in nanochromosome copy number, most notably the limited overamplification of nanochromosomes encoding highly expressed genes (rRNAs, tRNAs, and ribosomal proteins), even if a strong correlation exists between nanochromosome copy number and transcription levels, copy number may only be a modest contributor to the final RNA and protein expression levels. Regulation of expression at the transcriptional/posttranscriptional level may be essential to buffer the variation in DNA copy number that arises during extended periods of vegetative growth.
Oxytricha nanochromosomes range in length from ∼500 bp to 66 kb, with a mean size of ∼3.2 kb (Figure 8). Few nanochromosomes were assembled at either extremity of the length distribution, with just 32 shorter than 600 bp long and 61 longer than 15 kb, consistent with observations of macronuclear DNA on electrophoretic gels [2],[70]. While the mean length of two-telomere nanochromosomes in the final Oxytricha macronuclear genome assembly is ∼3.2 kb (Table 1), the true average length of nanochromosomes is shorter than this because the longest isoform of alternatively fragmented nanochromosomes is the one that tends to be assembled. On electrophoretic gels, Oxytricha nanochromosomes are visibly longer than those of Euplotes [2],[71], which we propose is primarily a consequence of the lack of alternative fragmentation in Euplotes (inspection of mapped reads to our preliminary Euplotes crassus assembly indicated no signs of alternative fragmentation; unpublished data). The longest isoforms of alternatively fragmented nanochromosomes average 5.0 kb (SD = 2.4 kb), while nonalternatively fragmented nanochromosomes have a mean length of 3.0 kb (SD = 2.4 kb; Figure 8). The mean length of the shortest nanochromosome isoforms produced by alternative fragmentation is 2.4 kb (SD = 1.6 kb). For single-gene nonalternatively fragmented nanochromosomes, the mean nanochromosome length is 2.2 kb (SD = 1.0 kb).
The shortest assembled nanochromosome (Contig20269.0) is a mere 248 bp, excluding the telomeric sequences. Though we were unable to identify any ORFs or any ncRNAs on this nanochromosome by RFAM searches, we found two matching RNA-seq PE reads, suggesting that there is expression from this nanochromosome. The shortest nanochromosome (Contig19982.0) with a known protein is 469 bp (excluding the telomeres) and encodes a 98 aa ThiS/MoaD family protein, while the shortest ncRNA-bearing nanochromosome we found is 540 bp (excluding telomeres) and encodes tRNA-Gln(CUG) (see Text S1, “Analysis of Short Protein and ncRNA-Encoding Nanochromosomes”). Searches for shorter possible nanochromosomes in the Illumina and Sanger reads did not reveal additional plausible nanochromosome candidates (Text S1, “Reads Containing Both Putative Telomeric Repeats Are Not Genuine Nanochromosomes”).
The longest nanochromosomes (>15 kb) typically encode a single large structural protein (Table S4), such as dynein heavy chain proteins (e.g., Contig354.1). None of the 20 longest nanochromosomes are alternatively fragmented. Seven of these 20 nanochromosomes contain multiple predicted genes (up to a maximum of four); however, all but one of these gene predictions are oriented head-to-tail, consistent with the possibility that their predictions may have been incorrectly split. Hence most of the longest nanochromosomes are likely still single-gene nanochromosomes. One ∼20 kb nanochromosome (Contig289.1) does indeed contain multiple genes, since it encodes a Pkinase domain (PF00069) protein on the opposite strand to two predicted PAS_9 domain (PF14326) proteins (though these latter two proteins may also be incorrectly split). Six of the longest nanochromosomes encode single proteins with no detectable Pfam domains (Pfam-A 26; independent E-value<0.01) but all have BLASTP NCBI non-redundant database (nrdb) matches (E-value<1e-10), typically to large proteins (>2,000 aa).
The longest nanochromosome (Contig7580.0) is 66 kb (65,957 bp; excluding telomeres) and encodes a single giant protein (“Jotin,” after a Norse giant) with BLASTP best hits to Titin-like genes in the NCBI nrdb (see Text S1, “Characterization of the Jotin Protein”). We note that this single-gene nanochromosome is comparable in size to the entire, relatively large and gene-rich ∼70 kb Oxytricha mitochondrial genome [65], which was largely eliminated by the sucrose gradient isolation of macronuclei (see Materials and Methods). The Oxytricha Jotin ORF is 64,614 bp. AUGUSTUS predicts four short introns (117, 151, 77, and 63 bp), two of which are supported by and one of which conflicts with RNA-seq reads. This gene's entire coding sequence is well supported by pooled RNA-seq reads (covered from end to end).
The gene prediction software, AUGUSTUS, predicted complete genes on 15,387 of the complete nanochromosomes we surveyed (96%) and 91% of the final assembly's contigs. Examination of three developmental time points (0, 10, and 20 h after initiation of conjugation) confirms transcription of 97% of Oxytricha nanochromosomes (94% of all contigs). AUGUSTUS predicts genes on 94% of nanochromosomes with expression evidence.
Most Oxytricha nanochromosomes (80%) contain single genes, consistent with earlier studies (Figure S9) [2],[33]. Alternatively fragmented nanochromosomes tend to encode more genes per nanochromosome: only 15% of alternatively fragmented nanochromosomes have single gene predictions, versus 90% of all nonalternatively fragmented nanochromosomes. Roughly half (48%) of multigene nanochromosomes have alternative fragmentation. All nanochromosomes with five or more (maximum eight) predicted genes are alternatively fragmented (Figure S9), and only two nonalternatively fragmented nanochromosomes encode four genes. The nanochromosome with the largest number of separate gene products (Contig8800.0; ∼6.8 kb) is alternatively fragmented, with a shorter ∼3.5 kb nanochromosome isoform that encodes 12 C/D snoRNAs [59] and a putative protein-coding gene encoded by the remainder of the full-length isoform.
Key properties of Oxytricha's gene predictions are consistent with a pilot survey [33], including relative AT-richness (34% GC) with noncoding regions that are more AT-rich than coding regions (e.g., introns are 23.6% GC), and 1.6 introns per gene (Table 2). Oxytricha gene lengths (mean length 1,839 bp excluding UTRs) are similar to those predicted for Tetrahymena [56].
Functional differences between the model ciliates may have evolved in numerous ways, given their tremendous divergence. Here we focus on two key differences: absence/presence of protein domains in specific ciliates and expansions of protein families at the level of protein domain. We were particularly interested in comparing the protein domains present in either ciliates with gene scrambling (Oxytricha) or that lack evidence of gene scrambling (Paramecium, Tetrahymena, and Euplotes; Figure 2; Tables S5 and S6; also see Table S7 for genes found in Paramecium and Tetrahymena that are absent in Oxytricha), since many species-specific proteins appear associated with macronuclear genome differentiation. Examples include the transposases in Oxytricha (micronuclear-limited TBE transposases) [30], Tetrahymena and Paramecium (piggyBac transposase—“PiggyMac”) [72],[73], and the Paramecium RNA binding Nowa proteins [74].
Since we were interested in DNA rearrangement, we searched for differences in the nucleic acid binding and nucleic acid metabolism domain content between the ciliates with and without evidence of extensive gene scrambling (see Materials and Methods). We identified 43 such nucleic-acid-related domains that are present in Oxytricha (“Oxytricha-specific” domains) but absent from both Tetrahymena and Paramecium (Table S5 and Table S6).
For additional results, see Text S1, Figures S1–S30 and Tables S1–S28.
The unique architecture of the Oxytricha macronuclear genome expands our perspective on the limits of genome organization. We summarize our main findings below.
The Oxytricha macronuclear genome now enables both comparative genomics in the same cell with its micronuclear precursor, as well as comparative macronuclear genomics with other species that possess a nanochromosome architecture and more divergent model organisms with no or less genome fragmentation. Broader taxonomic sampling of other ciliate macronuclear and micronuclear genomes will greatly enhance evolutionary studies of nuclear development and genome rearrangement. In conjunction with the macronuclear genome, transcriptome data provide the first tantalizing glimpses into sweeping cellular changes during nuclear development and merit more investigation. Specific protein studies will be necessary to identify key genome rearrangement players from the extensive candidate list of development-specific genes. To facilitate these and other studies, the Oxytricha macronuclear genome, which is available both in GenBank (AMCR00000000) and at oxy.ciliates.org, will continue to incorporate future refinements in the genome assembly, gene predictions and annotations.
Briefly, to obtain macronuclear DNA for Sanger and 454 sequencing, Oxytricha trifallax strain JRB310 was cultured in inorganic salts medium according to an established protocol [104] with Chlamydomonas reinhardtii and Klebsiella oxytoca as food sources. The JRB310 cells we used here are likely to have undergone less than 200 divisions since they were originally isolated and have been raised from cultures with two intervening encystments. Oxytricha cells were harvested by filtering through several layers of gauze to remove large particles, and then a 15 µm Nitex membrane was used to concentrate cells and remove bacteria and small contaminants.
The harvested cells were washed by low-speed centrifugation through a 0.25 M sucrose solution, then lysed in 0.25 M sucrose and 0.5% Nonidet P-40. This lysis disrupts the cell membrane, leaving nuclei intact. Nuclei were then spun through 0.25 M sucrose twice to remove bacteria, mitochondria, and other cell debris. Most micronuclei were also removed in this process. DNA was extracted using the AquaPure genomic DNA isolation kit (Bio-Rad) following the manufacturer's protocol.
To obtain pure macronuclear DNA for Illumina sequencing, Oxytricha trifallax strain JRB310 was cultured in inorganic salts medium and starved for 3 d at 4°C to allow consumption of most of the food source (Chlamydomonas reinhardtii) in culture. Cells were harvested by filtering through several layers of gauze to remove large particles. Then, a 10 µm Nitex membrane was used to concentrate cells and remove small contaminants.
We collected a macronuclear fraction in 40% sucrose from a standard sucrose gradient centrifugation protocol designed to separate macronuclei and micronuclei [105]. We then purified the macronuclei an additional time, by passing them through a 70% sucrose gradient at 12,000 rcf for 10 min. DNA was isolated from the macronuclei with a NucleoSpin tissue DNA isolation kit (Machery-Nagel) according to the standard protocol for cultured cells and then RNAse A treated prior to preparation of the Illumina libraries. Since considerable streaking of the DNA was evident from electrophoretic gels, excess salt was suspected in the samples and so the DNA was precipitated in ethanol (>8 h) at 4°C, then centrifuged at 16,000 rcf for 30 min, and washed twice, for 10 min in 70% ethanol, before resuspension in the kit's elution buffer.
Genomic shotgun libraries were prepared for different size fractions of Oxytricha macronuclear DNA using standard methods employed at The Genome Institute (TGI) for both Sanger and 454 sequencing (see Text S1, “Preparation of Nanochromosome DNA for Sanger/454 Sequencing” to “454 Sequencing of DNA from Nanochromosome Size Fractions”), while a special method was developed for the construction and 454 sequencing of paired telomeric ends (Text S1, “Whole Nanochromosome Telomere-Based Library Construction”). Both PE and SE Illumina libraries were prepared at Princeton University from pure JRB310 Macronuclear DNA using standard Illumina kits (Text S1, “llumina Genomic Library Construction and Sequencing”).
We developed a meta-assembly method (Figure S1) to build a reference genome assembly that is primarily derived from Illumina sequence data (Princeton Illumina assembly) but also takes advantage of an earlier Sanger/454 hybrid assembly (TGI 2.1.8 assembly). The data that were used for each of the assemblies are summarized in Table S8.
gmapper version 2.1.1b [111] was used to map reads to the final genome assembly in SE mapping mode with default parameters, and then filtered to contain read pairs that had both members matching with ≥94% identity to the assembly (further details about the read mapping are provided in Text S1, “Read Mapping Rationale”).
To identify potential heterozygous sites that were hidden by the majority rule applied in calling contig consensi during assembly, we identified SNPs (base substitutions and not indels) at positions with ≥20× read coverage and ≥5% frequency for telomere-masked, PE reads mapped to nanochromosomes both with (“matched”) and without (“matchless”) non-self BLAT matches (≥100 bp and ≥90% identical; default parameters), from VarScan (version 2.2.8, with a minimum variant frequency of 0.001) [112] output processed by a custom Python script. We pairwise aligned heterozygous “matched” nanochromosomes with MUSCLE (default parameters; for nanochromosome pairs where one of the nanochromosomes is no more than 10% longer than the other) and estimated heterozygosity for these nanochromosomes for alignments that were ≤15% identical. SNP data can be obtained from http://trifallax.princeton.edu/cms/raw-data/genome/mac/assembly/combined_assembly/snps/varscan_snps.tar.gz/view.
SNP heterozygosity at 4-fold synonymous sites was determined from 649 coding sequence pairs, corresponding to 1,298 matched nanochromosomes, aligned with MACSE (with parameters “-Xmx1000 m” and “-d 6”) [113] with no more than 5% gaps in each of the aligned sequences.
After splitting out and removing the adaptors used in the circular telomere-capturing constructs (see Text S1, “Whole Nanochromosome Telomere-Based Library Construction” and Figure S28 for the procedure used to produce these constructs), we selected all 454 PE reads with telomeric sequence repeats on either end and hard-masked all the telomeric repeats (matching the regular expression [AC]*CCCCAAAACCCC) with a single “N,” then selected all pairs where both reads were ≥30 bp long (719,566 pairs in total) and were terminated by telomeric repeats. We then mapped all the 454 telomeric PE reads to our final genome assembly with gmapper version 2.0.2 [111] in paired mode and the following parameters: -r 50; -p col-bw; -I 0,30000 (-r was set to 50 to accommodate the high indel error rate of 454 reads). Illumina telomeric reads, masked in the same manner as the 454 telomeric reads, were mapped with gmapper version 2.1.2b with default parameters, then filtered so that both members of the pair were ≥94% identical to the contig to which they mapped. Next we identified both 5′ and 3′ TASs for each contig in 200 bp windows around sites with maximal telomeric read coverage (this provides a lower bound estimate of the number of TASs, since these sites may span at least a couple hundred bases). Predicted sites are available at http://trifallax.princeton.edu/cms/raw-data/genome/mac/assembly/combined_assembly/telomere_addition_sites/. Alternative fragmentation sites were classified as strongly supported if they had ≥10 supporting Illumina telomeric reads and weakly supported if they had fewer matching reads than this (additional details about this classification are provided in Text S1, “Classification of Strongly and Weakly Supported Alternative Fragmentation Sites”).
Putative alternative nanochromosome isoforms were predicted based on 454 telomeric read pairs that provide a link between the ends of nanochromosomes, with any read pair ending 100 bp up- or downstream of each site providing a link. This provides a minimum estimate of the number of alternative nanochromosome isoforms produced by each locus and cannot predict longer alternative nanochromosome isoforms (much larger than 5 kb) due to the size selection limits on the initial sequence constructs. Predicted nanochromosome isoforms, with the number of reads supporting each isoform, can be found at http://trifallax.princeton.edu/cms/raw-data/genome/mac/assembly/combined_assembly/454_alt_forms.txt/view.
Relative nanochromosome copy numbers were estimated for nanochromosomes ≥1,800 bp long, from the total number of telomereless paired reads mapped in the intervening, nonsubtelomeric interval 600 bp from either end of each nanochromosome (http://trifallax.princeton.edu/cms/raw-data/genome/mac/assembly/combined_assembly/copy_number/copy_num_sam_filter6.nonsubtelomeric.txt/view). We excluded these subtelomeric regions, since the experimental protocol used to generate the reads lead to uneven coverage (e.g., see Figure 6 and Figure S8), which may in turn lead to poor estimates of nanochromosome copy number for shorter nanochromosomes. The total number of mapped reads was normalized by the total nanochromosome length minus the combined 1,200 bp subterminal interval. We also estimated relative copy number from the number of telomeric reads mapped to each nanochromosome (http://trifallax.princeton.edu/cms/raw-data/genome/mac/assembly/combined_assembly/copy_number/copy_num_sam_filter6.teloreads.unrestricted.txt/view).
RNA was isolated for five developmental time points (0, 10, 20, 40, and 60 h postmixing of JRB310 and JRB510 cells for conjugation) and used to create RNA-seq libraries with the Ovation RNA-Seq System (NuGEN Technologies, Inc. San Carlos, CA). Details of the RNA-seq library construction and sequencing are provided in Text S1, “RNA Isolation, NuGEN cDNA Synthesis and Illumina Sequencing.” To produce spliced mapped reads, RNA-seq data were mapped with BLAT [44] (“-noHead -stepSize = 5 -minIdentity = 92”) and then postprocessed to remove mapping artifacts (see Text S1, “RNA-Seq Mapping and Read Counting” for further details).
Gene predictions were produced by AUGUSTUS (version 2.5.5) [114],[115] using mapped RNA-seq data as “hints” for predictions (details about the training and prediction are provided in Text S1, “Gene Prediction”).
The final genome assembly has been deposited in GenBank with accession number AMCR00000000. Note that there are 20,162 contigs in GenBank, rather than the 22,450 reported for our final assembly, as some contigs were removed (e.g., if they were too short after vector trimming). Tables S8, S9, S10 and Text S1 provide links to the other assemblies and raw data.
For additional methods, see Text S1, Figure S28 and Tables S28, S29.
|
10.1371/journal.pgen.1001183 | Common Genetic Variants and Modification of Penetrance of BRCA2-Associated Breast Cancer | The considerable uncertainty regarding cancer risks associated with inherited mutations of BRCA2 is due to unknown factors. To investigate whether common genetic variants modify penetrance for BRCA2 mutation carriers, we undertook a two-staged genome-wide association study in BRCA2 mutation carriers. In stage 1 using the Affymetrix 6.0 platform, 592,163 filtered SNPs genotyped were available on 899 young (<40 years) affected and 804 unaffected carriers of European ancestry. Associations were evaluated using a survival-based score test adjusted for familial correlations and stratified by country of the study and BRCA2*6174delT mutation status. The genomic inflation factor (λ) was 1.011. The stage 1 association analysis revealed multiple variants associated with breast cancer risk: 3 SNPs had p-values<10−5 and 39 SNPs had p-values<10−4. These variants included several previously associated with sporadic breast cancer risk and two novel loci on chromosome 20 (rs311499) and chromosome 10 (rs16917302). The chromosome 10 locus was in ZNF365, which contains another variant that has recently been associated with breast cancer in an independent study of unselected cases. In stage 2, the top 85 loci from stage 1 were genotyped in 1,264 cases and 1,222 controls. Hazard ratios (HR) and 95% confidence intervals (CI) for stage 1 and 2 were combined and estimated using a retrospective likelihood approach, stratified by country of residence and the most common mutation, BRCA2*6174delT. The combined per allele HR of the minor allele for the novel loci rs16917302 was 0.75 (95% CI 0.66–0.86, ) and for rs311499 was 0.72 (95% CI 0.61–0.85, ). FGFR2 rs2981575 had the strongest association with breast cancer risk (per allele HR = 1.28, 95% CI 1.18–1.39, ). These results indicate that SNPs that modify BRCA2 penetrance identified by an agnostic approach thus far are limited to variants that also modify risk of sporadic BRCA2 wild-type breast cancer.
| The risk of breast cancer associated with BRCA2 mutations varies widely. To determine whether common genetic variants modify the penetrance of BRCA2 mutations, we conducted the first genome-wide association study of breast cancer among women with BRCA2 mutations using a two-stage approach. The major finding of the study is that only those loci known to be associated with breast cancer risk in the general population, including FGFR2 (rs2981575), modified BRCA2-associated risk in our high-risk population. Two novel loci, on chromosomes 10 in ZNF365 (rs16917302) and chromosome 20 (rs311499), were shown to modify risk in BRCA2 mutation carriers, although not at a genome-wide level of significance. However, the ZNF365 locus has recently independently been associated with breast cancer risk in sporadic tumors, highlighting the potential significance of this zinc finger-containing gene in breast cancer pathogenesis. Our results indicate that it is unlikely that other common variants have a strong modifying effect on BRCA2 penetrance.
| After more than a decade of clinical testing for mutations of BRCA1 and BRCA2, there remains considerable uncertainty regarding cancer risks associated with inherited mutations of these genes. This variable penetrance is most striking for BRCA2 [1]–[4], and it affects medical management [5]. Women with the same BRCA2 mutation may develop breast, ovarian or other cancers at different ages or not at all [6]. In a segregation analysis of families identified through breast cancer cases diagnosed before age 55, the residual familial clustering after accounting for BRCA1 and BRCA2 mutations could be explained by a large number of low penetrance genes with multiplicative effects on breast cancer risk [7], [8]. A candidate gene approach in BRCA2 mutation carriers led to the discovery of loci that modify the penetrance of BRCA2 mutations, such as RAD51 135 G>C [9] and perhaps CASP8 [10], [11] and IGFBP2 [12], if replicated. To investigate whether other common single nucleotide polymorphisms (SNP), copy number variants (CNV), or copy number polymorphisms (CNP) modify penetrance for BRCA2 mutation carriers, we undertook a two-staged genome-wide association study (GWAS) in BRCA2 mutation carriers from the international Consortium for Investigators of Modifiers of BRCA1/2 (CIMBA) and other international studies. We hypothesized that an agnostic search for breast cancer loci in an enriched population of BRCA2 mutation carriers, the first among this high risk population, would provide greater power than a sporadic population of equal number, and would yield associations specific to BRCA2 carriers and/or the general population.
In stage 1, genotype data were available for 899 young (<40 years) affected and 804 older (>40 years) unaffected carriers of European ancestry after quality control filtering and removal of ethnic outliers (Figure S1). A total of 592,163 filtered SNPs genotyped using the Affymetrix Genome-Wide Human SNP Array 6.0 platform passed quality control assessment. In stage 1, comparison of the observed and expected distributions (quantile-quantile plot: Figure S2) showed little evidence for an inflation of the test statistics (genomic inflation factor λ = 1.01), thereby excluding the possibility of significant hidden population substructure, cryptic relatedness among subjects or differential genotype calling between BRCA2 affected and BRCA2 unaffected carriers. Multiple variants were found to be associated with breast cancer risk (Figure S3): 3 SNPs had p<10−5 and 39 SNPs had p<10−4. The most significant association () was observed for FGFR2 rs2981582 (Table 1), a variant previously shown to be associated with increased risk of BRCA2-related breast cancer [13]. A positive association was also observed with rs3803662 (Table 1), near TOX3, which has also been associated with sporadic breast cancer risk [13].
Using the stage 1 data, we also performed a GSEA as implemented in MAGENTA [14] to evaluate whether a functionally-related set of genes relevant to BRCA2 function (Table S1) was enriched for relative risk associations (see Statistical Methods). The 59 genes selected are related to the Fanconi anemia pathway [15] as well as other pathways reported in the literature to regulate or interact with BRCA1/2 [16]. These showed no enrichment of associations with the breast cancer risk (p = 0.56). In addition, eight of 125 known cancer susceptibility alleles identified by previous GWAS of other cancers [17] were associated with BRCA2 modification in the current study, a number not greater than expected (Kolmogorv-Smirnov p = 0.60) by chance alone. Of the 113 most significantly associated SNPs (p<10−3) in our study, three showed significant association (p<0.05) with BRCA1-associated breast cancer risk in a complimentary GWAS [18].
In the combined stage 1 and stage 2 results, four independent SNPs (pairwise ) were associated with increased risk of breast cancer risk with p-values<10−4 (Table 1). Previously identified breast cancer susceptibility loci [13], [19], [20] had the most significant associations among BRCA2 mutation carriers (FGFR2: per allele and TOX3: per allele ). Novel loci, rs16917302 on chromosome 10 and rs311499 on chromosome 20, had HRs in stage 2 that were in the same direction as those observed for stage 1 (Figure 1, Table 1), but were smaller in magnitude (HR = 0.67 (95% CI:0.56–0.80) vs. 0.85 (95% CI: 0.70–1.04) for rs16917302; HR = 0.60 (95%CI:0.50–0.78) vs. 0.84 (95%CI: 0.67–1.06) for rs311499) perhaps reflecting a “winner's curse” effect” [21]. The associations for these SNPs were not statistically significant in stage 2 (Table 1). In the combined stage 1 and stage 2 dataset, the C allele of rs16917302 was associated with lower risk of breast cancer (per allele HR = 0.75, 95% CI 0.66–0.86; ; Table 1), and the C allele of rs311499 was associated with a reduced risk (per allele HR = 0.72, 95% CI 0.61–0.85; ; Table 1). A full list of stage 2 results can be found in Table S2. Using the combined stage 1 and stage 2 data, there was no evidence that the HR for SNP rs16917302 changes with age (p = 0.63), but there was some evidence that the per-allele HR for rs311499 may increase with age (p = 0.034).
We also examined the association of both high-frequency CNPs and low-frequency CNVs to case-control status using the stage 1 data. After performing standard quality control measures including a minor allele frequency (MAF) threshold of 5%, we identified 191 polymorphisms with reliable genotypes. No associations were found between CNVs and the phenotype; there was no inflation or deflation of the test statistic, and the best p-value was . We similarly assessed less common CNPs, and found neither the overall burden of events (or any subclass thereof, such as large deletions overlapping genes) nor any specific locus associated with breast cancer risk (Figure S4).
Because of the prior evidence of significant LD extent around the 6174delT (c.5946delT) founder mutation in the Ashkenazi Jewish population [22], we explored the potential excess sharing of the genome compared to the BRCA2 region in both Ashkenazi Jewish and non-Jewish European ancestries. Using GERMLINE [23], shared segments of greater than 5 cM were computed based on the imputed genotype dataset. In the BRCA2 region, we observed a significant excess of sharing amongst both Ashkenazi (n = 304) and non-Jewish (n = 1331) individuals compared to samples from an autism study (n = 808) suggesting common founders for BRCA2 mutations. Examining sites across the genome every 2.5 cM (excluding telomere and centromere regions), we observed possible pairs share segments greater than 5 cM that on average 0.005% (u = 50.17, s.d = 55.5, max = 491) for non-Jewish individuals and 0.12% (u = 141.11, s.d = 57.32, max = 525) for Ashkenazi Jewish individuals. Comparing cases and controls, we did not observe a significant difference in number of pairs of samples sharing segments greater than 5cM across the genome excluding chromosome 13. That is, there was no evidence of overall excess sharing across the genome other than for the BRCA2 locus within the Ashkenazi Jewish and non-Ashkenazi Jewish populations in the study.
In this GWAS of BRCA2 mutation carriers, the first in this high risk population, we found previously identified breast cancer susceptibility loci modified risk of BRCA2-associated breast cancer with similar magnitude of association. Although FGFR2 (rs2981575) was the only locus to reach genome-wide statistical significance, novel loci, rs16917302 and rs10509168 were each associated with breast cancer risk.
rs16917302 is located on chromosome 10, in the zinc finger protein 365 gene (ZNF365). A recent multistage GWAS of 15,992 sporadic breast cancer cases and 16,891 controls also observed an inverse association (per allele OR = 0.82, 95% CI 0.82–0.91, ) between breast cancer risk and rs10509168, a SNP 18kb from rs16917302 (pairwise ) and located in intron 4 of ZNF365 [24]. Of the 3,659 cases and 4,897 controls in phase 1 of that study, imputation revealed that the locus identified in our BRCA2 study, rs16917302, was significantly associated with risk for breast cancer (p = 0.02) (Easton DF, personal communication). The second novel SNP in the current study, rs311499, is located on chromosome 20, within a region containing several possible candidate genes including GMEB2, SRMS, PTK6, STMN3, and TNFRSF6. The functional significance of both of these regions with breast carcinogenesis is unknown; further research is warranted.
There was some evidence that the HR associated with rs311499 may change with age. We also observed that the stage 1 HR for this SNPs was larger in magnitude compared to the stage 2 HR, consistent with a winner's curse effect [21]. Since stage 1 of our experiment included mostly BRCA2 mutation carriers diagnosed at a young age, and stage 2 mutation carriers diagnosed an older age, the “winner's curse” and age-specific effects are confounded and may be difficult to distinguish. Fitting the age-dependent HR model for SNP rs311499 using the stage 2 data yielded no significant variation in the HR by age (p = 0.47), but the sample size for this analysis was relatively small. Future larger studies should aim to clarify this.
Mutations in known genes (BRCA1, BRCA2, TP53, CHEK2, PTEN, and ATM) explain only 20–25% of the familial clustering of breast cancer; the residual familial clustering may be explained by the existence of multiple common, low-penetrance alleles (‘polygenes’) [25]. Perhaps because the majority of BRCA2-associated breast tumors are estrogen receptor (ER)-positive, as are the majority of non-hereditary breast cancers [26], risk alleles for sporadic breast cancer are more likely to be modifiers of risk of BRCA2-associated hereditary breast cancer. Of the seven GWAS-identified breast cancer-associated SNPs examined in a BRCA2 background [13], [19], [20], SNPS in FGFR2 (rs2981575), TOX3 (rs3803662), MAP3K1 (rs889312), and LSP1 (rs3817198) have been shown to modify BRCA2 penetrance, in contrast with BRCA1 tumors, in which only two of these same SNPs (based on a 2 degrees of freedom model) modified risk of these largely ER-negative tumors [26]. As previously noted [13], [20], the stage 1 HRs among BRCA2 mutation carriers, reported here, were nearly identical to odds ratio estimates observed in sporadic breast cancer studies, consistent with a simple multiplicative interaction between the BRCA2 mutant alleles and the common susceptibility SNPs. If replicated, the two additional SNPs identified here would only explain about 1.7% of the variance in breast cancer risk among BRCA2 mutation carriers. Taken together, the combined effects of all the common and putative risk modifiers in this study only account for ∼4% of the variance of BRCA2 mutations, compared with 1.1% for the single RAD51 135 G>C variant, which is rare and biologically-linked to BRCA2 function, as shown by candidate gene studies [9]. Thus, the common alleles that modify risk in BRCA1 and BRCA2 backgrounds appear to have comparable associated risks in sporadic ER-positive and ER-negative tumors, respectively [18]. While individual SNPs are unlikely to be used to guide radiographic screening and risk-reducing surgical strategies, the combined effect of these SNPs may ultimately be used for the tailor management of subsets of BRCA mutation carriers [5].
While we took great efforts to collect all of the possible known BRCA2 mutation carriers, there were insufficient numbers to stratify by race and BRCA2 mutations with the exception of BRCA2*6174delT mutations. Due to the small numbers of women of non-European ancestry who have participated in the individual studies represented here, the current analysis was based only on women who had genetic backgrounds consistent with HapMap CEU samples. While we expect that SNPs identified among women of European ancestry might also be applicable to women of other genetic backgrounds, additional research in these populations will be needed. Similarly, the observed associations represented across all types of mutations, and specifically a weighted average of BRCA2*6174delT and non-delT mutations. It is possible that the observed associations may only modify the penetrance of specific BRCA2 mutations due to differential effects on function or differences in genetic background. Our analysis was stratified on the basis of the most common BRCA2 mutation, BRCA2*6174delT, which is prevalent in individuals with an Ashkenazi Jewish ancestry. Large numbers of mutation carriers will be necessary to calculate mutation-specific estimates. In addition, there was a drop-out of SNPs in the two phases of this study. While we were able to achieve a representative coverage of the genome, it is also possible that additional studies using denser arrays may provide further information.
As expected, we observed associations with some of the major common genetic variants seen in genome-wide scans of breast cancer in a non-BRCA1/2 mutation background. However, we found no evidence for loci with stronger effects than FGFR2. Although we observed an association with a novel locus at ZNF365 that appears also to be a risk factor for sporadic breast cancer, overall, our results suggest that there are no common variants with major effects (i.e., OR>2.0) that are specific in BRCA2 carriers. Similarly, in a recent report of SNPs from sporadic breast cancer GWAS genotyped in a restricted set of BRCA1/2 carriers [27], loci in LOC134997 (rs9393597: per allele HR = 1.55, 95% CI 1.25–1.92, ) and FBXL7 (rs12652447: HR = 1.37, 95% CI 1.16–1.62, ) were associated with BRCA2 breast cancer risk with p-values weaker than FGFR2 reported here (per allele ), although the magnitudes of the associations were slightly stronger than FGFR2 (HR = 1.28). Although these SNPs were not in our genotyped panel of SNPs at stage 1, imputation results indicate that SNP rs9393597 has a p-value of 0.008 and SNP rs12652447 a p-value of 0.04 for association with breast cancer risk for the BRCA2 mutation carriers in our stage1. However, there is substantial overlap between our study and the study of Wang et al. [27].
Replication in larger datasets will be necessary to precisely estimate the magnitude of the associations of suspected loci identified from our study, candidate gene analysis [10]–[12], and other selection approaches [27]. It is of interest, however, that when utilizing an agnostic approach in BRCA2 mutation carriers in this study, the major determinants of risk variation in mutation carriers are those that also modify risk in subsets of sporadic, BRCA1/2 wild type, breast cancer. However, it remains possible that unique variants with smaller effects, or rarer variants (not evaluated in this experiment), may be specific modifiers of breast cancer risk in BRCA2 carriers. Their detection would require study populations much larger than the current analysis, which is presently the largest such cohort assembled.
|
10.1371/journal.pntd.0002253 | Schistosoma japonicum Soluble Egg Antigens Attenuate Invasion in a First Trimester Human Placental Trophoblast Model | Schistosomiasis affects nearly 40 million women of reproductive age, and is known to elicit a pro-inflammatory signature in the placenta. We have previously shown that antigens from schistosome eggs can elicit pro-inflammatory cytokine production from trophoblast cells specifically; however, the influence of these antigens on other characteristics of trophoblast function, particularly as it pertains to placentation in early gestation, is unknown. We therefore sought to determine the impact of schistosome antigens on key characteristics of first trimester trophoblast cells, including migration and invasion.
First trimester HTR8/SVneo trophoblast cells were co-cultured with plasma from pregnant women with and without schistosomiasis or schistosome soluble egg antigens (SEA) and measured cytokine, cellular migration, and invasion responses.
Exposure of HTR8 cells to SEA resulted in a pro-inflammatory, anti-invasive signature, characterized by increased pro-inflammatory cytokines (IL-6, IL-8, MCP-1) and TIMP-1. Additionally, these cells displayed 62% decreased migration and 2.7-fold decreased invasion in vitro after treatment with SEA. These results are supported by increased IL-6 and IL-8 in the culture media of HTR8 cells exposed to plasma from Schistosoma japonica infected pregnant women.
Soluble egg antigens found in circulation during schistosome infection increase pro-inflammatory cytokine production and inhibit the mobility and invasive characteristics of the first trimester HTR8/SVneo trophoblast cell line. This is the first study to assess the impact of schistosome soluble egg antigens on the behavior of an extravillous trophoblast model and suggests that schistosomiasis in the pre-pregnancy period may adversely impact placentation and the subsequent health of the mother and newborn.
| Approximately 40 million women of childbearing age suffer from schistosome infection globally at any given time. Multiple studies in rodent models, as well as a few reports in humans, suggest that schistosome infection results in poor pregnancy outcomes. We have previously shown that antigens released from schistosome eggs result in a pronounced pro-inflammatory response in syncytialized third trimester trophoblasts. Herein, we examine the effect of schistosome egg antigens on a first trimester trophoblast cell line, an accepted model for early placental development. Not only is the pro-inflammatory response recapitulated in this model system, but we also observed a decrease in migration and invasion of trophoblast cells after exposure to these antigens. Both migration and invasion are key aspects in early placental development, and inadequate invasion has been implicated in pregnancy-related diseases such as growth restriction and preeclampsia. This study is the first to examine the impact of schistosome antigens on early placental development, and may have implications for the subsequent health of both the pregnancy and the child.
| Schistosomes are parasitic worms endemic to many parts of Africa, South America and Southeast Asia. They represent a significant disease burden in endemic regions, and have been estimated to be responsible for as many as 13–15 million disability-adjusted life years (DALYs) lost per year, with the true number potentially much higher [1]. Of the estimated 200 million people worldwide infected at any one time, approximately 40 million are women of reproductive age [2], [3]. A 2002 World Health Organization policy statement recommended the use of praziquantel in pregnant and lactating women [4], however many women still experience multiple cycles of pregnancy and lactation with schistosomiasis. This is due to the fact that in many regions of the world, pregnant and lactating women are still routinely excluded from treatment initiatives due to the Federal Drug Administration Class B designation that praziquantel still carries, as well as barriers to praziquantel acquisition and distribution.
Data from our laboratory and others have demonstrated poor reproductive outcomes in rodents and humans in the context of schistosomiasis. In rodent models, schistosome infection has profound impacts on birthweight and litter size [5]–[7]. We have previously shown that schistosome infection in a population of pregnant women residing in The Philippines is positively associated with increased risk for chorioamnionitis and increased pro-inflammatory cytokines in both maternal and cord blood [8]. Although very few adult worms and/or eggs are thought to directly traffic to the tissues of the maternal-fetal interface [9], [10], the residency of adult worms in the mesenteric vasculature and lodging of eggs in the liver allow for continuous secretion of antigens directly into the blood stream of the host. These antigens are known to traffic to, and cross, the human placenta, and have been found in fetal circulation [11]–[17]. We hypothesized that these antigens may have a direct effect on the cells at the maternal-fetal interface, and in the studies described herein, have chosen to focus specifically on processes specific to extravillous trophoblast cells.
Many events occur early in gestation that can have profound effects on the subsequent health of the fetus from gestation into adulthood. A lack of data pertaining to schistosomiasis during this critical window of development prompted us to utilize the first trimester cell line, HTR8/SVneo, to investigate the influence of schistosome infection on early events of pregnancy. Initial investigation was performed with co-culture of HTR8 cells and plasma collected from pregnant women infected with schistosomiasis or matched controls. To isolate the schistosome soluble egg antigen (SEA) specific effect on trophoblasts from any contribution of the host response, we next evaluated the direct impact of SEA on HTR8 cells. Events critical to placentation, including cytokine production, cellular migration and invasion were all assessed in an in vitro setting.
For the HTR8 human plasma co-culture experiment, written informed consent was obtained from each participant, and the study was approved by the institutional review boards at Rhode Island Hospital and the Philippines Research Institute of Tropical Medicine.
We used the immortalized first trimester cell line HTR8/SVneo, originally obtained from a human pregnancy terminated in the first trimester, and displaying properties of invasive extravillous cytotrophoblast cells [18]. Cells were maintained at 37°C with 5% CO2 in 1∶1 DMEM/F-12 media (Invitrogen, Grand Island, NY) supplemented with 1% L-glutamine (Invitrogen), 1% penicillin/streptomycin (Invitrogen) and 5% fetal bovine serum (Atlanta Biologicals, Lawrenceville, GA). All experiments were performed with the addition of SEA (25 µg/ml) in complete media for 24 h or media only control, unless otherwise noted. This dose was chosen based on dose response curves performed previously in our laboratory using purified primary trophoblast cells [19].
Schistosoma japonicum SEA was generously donated by Dr. Chuan-Xin Yu (Jiangsu Provincial Institute of Schistosomiasis Control, Wuxi, Jiangsu, China) after having been prepared according to standard procedure [20]. The SEA was prepared under endotoxin free conditions, with all reagents and equipment used to isolate the SEA from collected livers being LPS free prior to use. Preparations were evaluated for contaminating endotoxin using an FDA-cleared LAL-assay (Lonza Group, Basel, Switzerland). Endotoxin levels for all SEA preparations used were <6 EU/mg protein, which, in our culture conditions, is at least 1000-fold lower than levels that have been previously shown to influence human trophoblast cells [21].
Following the treatment period, media from HTR8/SVneo cells was collected and levels of multiple cytokines, chemokines, and fibrotic markers were assessed on a bead-based platform (BioPlex, Bio-Rad, Hercules, CA) using a sandwich antibody-based assay as previously described [22]. Cytokines evaluated included interleukin (IL)-1β, IL-6, interferon (IFN)-γ, tumor necrosis factor (TNF)-α, IL-4, IL-5, IL-10, IL-13, IL-12, IL-8 and IL-2. These specific cytokines were measured, as they were all components of a multiplex analysis developed and validated in our laboratory. The majority of these cytokines have been reported to be expressed by the placenta, although expression is highly variable depending on culture conditions, gestational age and disease status [23]–[26]. In addition, we measured levels of tissue inhibitor of metalloproteinases (TIMP)-2, TIMP-4, insulin-like growth factor binding protein (IGFBP)-5, matrix metalloproteinase (MMP)-9, tenascin C, syndecan 1, Fas ligand, osteopontin, TIMP-1, connective tissue growth factor (CTGF), macrophage inflammatory protein (MIP)-1α, MMP-1, IGF-1, MMP-8, monocyte chemotactic protein (MCP)-1, and TIMP-3. These analytes were developed into multiplexed assays due to their importance in schistosomiasis-associated and idiopathic pulmonary fibrosis [27]–[29]. However, they are also widely implicated in invasion and remodeling at the maternal-fetal interface, underscoring the similarities between these two processes.
For the human plasma assays, we utilized plasma collected from pregnant women at 32 weeks gestation residing in Leyte, the Philippines, an area endemic for S. japonicum. The study population and sample collection has been described elsewhere [8], with socioeconomic status (SES), gravida, parity, body mass index (BMI), smoking status and age determined via questionnaire [30]. Schistosomiasis and co-infections (Ascaris lumbricoides, Trichuris trichuria, and hookworm) were determined from stool samples using the Kato Katz method. Infection intensities for each were determined using the WHO guidelines [31]. From this larger cohort of 150 women, we selected 29 women infected with schistosomiasis (20 lightly infected [1–99 eggs per gram], 9 moderate infection [100–399 epg]), and 29 uninfected women matched to the infected women for SES, co-infections, gravida, parity, gestational age, BMI, smoking status and maternal age (Table 1). Schistosomiasis was evaluated as a nominal (yes/no) variable due to the low numbers of women with moderate-high infection intensities. Serum was collected at the only pre-natal study visit (32 wks gestation).
HTR8/SVneo cells were cultured to 80% confluency in complete media before being cultured for 48 hours in serum free media with the addition of 10% plasma from the aforementioned pregnancies. Following 48 h incubation, trophoblast culture supernatants were collected and analyzed for cytokine production as described above.
HTR8/SVneo cells were cultured to 100% confluence in complete media. Once completely confluent, a scratch was made across the well using a sterile pipette tip. The underside of each of the wells was cross-hatched for reference. Cells were briefly washed with PBS in order to remove all detached cells after scratching, and cultured in serum-free media with the addition of SEA (25 µg/ml) for 48 h. Phase contrast images of the denuded region were taken at 0 h, 24 h and 48 h after scratch formation using an Olympus IX70 inverted tissue culture microscope (Olympus Corp., Tokyo, Japan). The same region of each well was imaged at each time point, using the cross-hatching as reference. The area free of cells was quantified using ImageJ software (NIH, Betheseda, MD). The denuded area at 24 h and 48 h for a specific well was expressed as a percentage of the denuded area that had been present in that well at 0 h, thus controlling for well-to-well variation in original scratch sizes.
MTT assays were done on HTR8/SVneo cells in parallel to the migration assays. MTT (Sigma Aldrich, St. Louis, MO) was added to each well and the cells were incubated for 4 h at 37°C in a humidified environment. Media and MTT were aspirated from each well, MTT solvent (4 mM HCl, 0.1% Nonidet P-40, in isopropanol) added, and the plate incubated at 25°C in the dark with rotation for 15 minutes. Absorbance for each well was read at 560 nm and 630 nm.
HTR8/SVneo cells at 80% confluency were treated with SEA (25 µg/ml) for 24 h in complete media before being gently trypsinized, washed with complete media and resuspended in serum-free media. 25,000 cells/well were plated on matrigel-coated transwell inserts with an 8 µm pore size (Corning, Tewksbury, MA). The bottom chamber contained complete HTR8 media. Following 48 h incubation, cells and matrigel were gently removed from the top of the transwell, and those cells that had invaded through the matrigel, traversed the pores, and reached the bottom of the transwell were stained with hematoxylin (Sigma Chemical). Stained cells were visualized and counted using an Olympus BH-2 microscope (Olympus Corp.).
Data analysis was performed using JMP v.10 (SAS Institute, Cary, NC). All data were evaluated for normality using the Shapiro-Wilk test. Those experiments for which all data were normally distributed were further evaluated with ANOVA and t-tests, with means ± SEM reported. For data that was not normally distributed, Wilcoxon Signed Rank analyses were performed, with data reported as median ± IQR. Specifically, cytokine production by HTR8 cells was compared between cells exposed to uninfected plasma and those exposed to infected plasma (Figure 1) as well as cells cultured with media alone and media with SEA (Figures 2 and 3). Similarly, HTR8 migration and invasion were compared between cells cultured with media alone and those with SEA addition to the media (Figures 4 and 5). Statistical significance was considered as P<0.05.
For these experiments, we selected a sub-set of plasma samples collected from women at 32 weeks gestation as part of a previous study [8], matching samples from 29 infected women with 29 samples from uninfected women on key potential confounding covariates (Table 1). HTR8/SVneo cells were cultured in serum-free media supplemented with 10% plasma collected at 32 weeks gestation, and allowed to remain in culture for 48 h. Media from HTR8/SVneo cells cultured with plasma from schistosome-infected women had significantly higher levels of the pro-inflammatory cytokines IL-6 and IL-8 (29%, P<0.02 and 42%, P<0.01, respectively), compared with cells cultured in the presence of plasma from uninfected pregnant women (Fig. 1). Levels of both IL-6 and IL-8 in the plasma alone were in most cases undetectable, with the highest level of either cytokine across all plasma samples being 32 pg/ml (data not shown). These data indicate that schistosomiasis results in the production of some factor(s), present in maternal circulation, that stimulate a pro-inflammatory cytokine response by first trimester trophoblasts.
Given that SEA are found in the circulation of infected individuals and are known to cross the placental barrier, we treated HTR8/SVneo cells with purified SEA (25 µg/ml) in culture. Within 24 h of culture, HTR8 cells treated with SEA secreted higher levels of IL-6 (3.2-fold, P = 0.01), IL-8 (1.5-fold, P = 0.02) and the pro-inflammatory chemokine, MCP-1 (also known as CCL2; 1.7-fold, P 0.04) into the culture media, as compared to HTR8/SVneo treated with media alone (Fig. 2). Not only do first trimester cell models respond to SEA with a pro-inflammatory signature, they also secrete higher levels of a chemo-attractant protein that may help recruit specific immune cells to the placenta, exacerbating the inflammatory reaction initiated by SEA at the maternal-fetal interface.
In addition to cytokine analysis, media from HTR8/SVneo cells treated with SEA for 24 h in culture were assessed for altered levels of a number of fibrotic markers. Of these, TIMP-1 production was increased in media from HTR8 treated with SEA, compared to media from cells with no SEA exposure (Fig. 3). In contrast, production of CTGF, CCL18, TIMP-3 and fibronectin all showed no difference in levels secreted from HTR8/SVneo cells exposed to SEA compared to those with media alone (data not shown). Given the important role of TIMP-1 in the inhibition of MMPs, SEA may influence the ability of these first trimester cells to remodel and migrate into the maternal uterine wall.
We performed in vitro wound assays to assess migration of the first trimester cell line, HTR8/SVneo, following exposure to SEA. There was little difference in cell migration at 24 h (Fig. 4a). By 48 h however, HTR8 cells in culture with SEA had filled only 57±8% of the denuded area compared to 92±6% for untreated cells (P<0.01, Fig. 4). Cell proliferation was measured using MTT assay in all wells, however no differences were observed between those wells with SEA added compared to those with media alone indicating the wound closure was not due to a proliferative effect of SEA treatment (data not shown). Together, these data suggest that migration of first trimester trophoblast cells is decreased in the presence of SEA.
We next assessed the ability of the HTR8/SVneo cell line to invade through matrigel, a model of extracellular matrix, and a transwell insert after being treated with SEA, using a standard invasion assay [32]. Cells were pretreated with SEA for 24 h to minimize any direct effect of SEA on the matrigel. Enumeration of the cells that had traversed the matrigel and the pores of the transwell after an additional 48 h incubation showed 2.7-fold lower absolute cell numbers in those wells that had been pre-treated with SEA compared to the HTR8 cells that received media alone for the initial 24 h (P = 0.04, Fig. 5). These data indicate that SEA inhibits the migratory and invasive properties of the first trimester cell line, HTR8/SVneo.
Despite a 2002 WHO recommendation that pregnant and lactating women be considered for inclusion in treatment programs [33], pregnant women with schistosomiasis are still excluded in many regions pending further evaluation of praziquantel's safety during pregnancy. Data regarding the impact of schistosome infection on human pregnancy is rather scant, although we have reported an increase in pro-inflammatory markers in maternal, placental, and newborn compartments of pregnancies complicated by schistosomiasis, as well as increased rates of acute subchorionitis in these women [8]. Several human studies have evaluated the role of schistosomiasis during pregnancy [11], [34], [35]. Two observational studies reported lower birth weights among infants from infected mothers. However, methodological issues, including lack of control for important potential confounders such as socioeconomic status and maternal nutritional status [35], and potential selection bias [11] make interpretation difficult. A recently completed RCT evaluating treatment of schistosomiasis during pregnancy in Uganda reported no change in birth weight among women treated for schistosomiasis during the second trimester and untreated controls [34]. Any effect(s) of schistosomiasis on early placentation however would not be detected because treatment occurred late in gestation. Thus, questions regarding the influence of schistosome infection during the first trimester of pregnancy in humans remain largely unanswered.
Although direct trafficking of the adult worm or schistosome eggs to the maternal-fetal interface is thought to be a rare event [9], schistosomiasis is known to produce a distinct antigenic signature in the circulation of infected individuals, including the presence of high levels of soluble egg antigens (SEA) which can cross the placental barrier [12]. Previously, we demonstrated that SEA can cause pro-inflammatory cytokine production in primary human trophoblast cells taken at term and allowed to syncytialize in vitro [19]. However, placentation is a dynamic process requiring trophoblast populations distinct in time and differentiation lineage to behave in specific and unique ways. In this manner, a syncytialized term trophoblast is responsible for very different functions (i.e. nutrient transfer, cytokine and hormone production) than a first trimester invasive trophoblast cell. In this report, we have focused on the effect of SEA on first trimester trophoblast cells, using the cell line HTR8/SVneo, as these cells represent one of the best model systems available for studying the behavior and characteristics of invasive, extravillous trophoblast cells [36].
As we have previously reported in term syncytialized trophoblasts, HTR8 cells exposed to SEA for 24 h (25 µg/ml) exhibited a pro-inflammatory cytokine signature. These findings echoed the pro-inflammatory signature we observed in HTR8 cells exposed to plasma from pregnant women infected with schistosomiasis, compared to plasma from pregnant, uninfected controls. A potential limitation of these experiments is that HTR8 cells were cultured with maternal plasma collected during the third trimester because the original study from which these samples originated did not enroll pregnant women until 32 weeks gestation [8]. We do not expect this to influence either the validity or generalizability of our results as schistosomiasis infection status and its consequent host response should not be altered during gestation.
Although little is known regarding the impact of localized pro-inflammatory cytokines during the first trimester, they have been suggested to contribute to reduced migration and invasion of trophoblast cells, increased migration of innate immune cells to the maternal fetal interface, and, at very high levels, are postulated to play a role in preterm delivery and/or miscarriage [21], [37], [38].
Another major role of extravillous trophoblast cells, particularly in the first trimester, is to remodel and invade deep into the maternal endometrium. This process is tightly regulated, and failure to invade to the appropriate degree has been associated with the development of a number of gestational diseases, most importantly preeclampsia, preterm birth and low birth weight [39], [40]. Failure to identify any difference in the cellular metabolic activity (MTT assay as a surrogate for cell proliferation) between the untreated and SEA-exposed cells supports the idea that SEA is not simply cytotoxic to HTR8 cells. Rather, HTR8 cells display increased production of TIMP-1, inhibition of cellular migration and decreased levels of invasion through matrigel when exposed to SEA. Of the fibrosis-associated molecules measured, TIMP-1 is arguably the most relevant to trophoblast migration/invasion [41], [42]. These data suggest that the process of placentation could be compromised in pregnancies complicated by schistosomiasis during the first trimester. To our knowledge, there have been no studies examining the effect, if any, of schistosomiasis on gestational diseases such as preeclampsia.
Our data are consistent with previous work from our laboratory regarding schistosome induced pro-inflammatory cytokine production across different models of trophoblast cells [19]. Outside of pregnancy, we and others have related these responses to nutritional, hepatic and hematologic morbidities in infected individuals [43]–[46]. Surprisingly, maternal schistosomiasis during pregnancy elicits a pro-inflammatory response detectable in the neonate [8], and neonates exposed to maternal schistosomiasis during pregnancy display a more robust response to antigenic challenge and have elevated levels of antibodies against schistosome antigens at birth than their unexposed counterparts [13], [47]. Together, these results suggest that maternal schistosomiasis may influence the outcome of initial pediatric schistosome infections acquired during early childhood.
The finding that SEA may modify invasion of extravillous trophoblasts and alter the cytokine milieu at the maternal fetal interface lends support to an aggressive treatment approach for women of reproductive age, such that they enter pregnancies infection free. It should also be noted that studies which have, and will, evaluate the efficacy of praziquantel given after the first trimester (ClinicalTrials.gov, registered study number NCT00486863), may not capture the full benefit of treatment as it relates to early placentation processes. Studies regarding the incidence of gestational diseases such as preeclampsia in the context of high schistosome prevalence are warranted and may shed additional light on the impact of schistosomiasis on the early development of the human placenta.
|
10.1371/journal.ppat.1004840 | Hepatitis D Virus Infection of Mice Expressing Human Sodium Taurocholate Co-transporting Polypeptide | Hepatitis D virus (HDV) is the smallest virus known to infect human. About 15 million people worldwide are infected by HDV among those 240 million infected by its helper hepatitis B virus (HBV). Viral hepatitis D is considered as one of the most severe forms of human viral hepatitis. No specific antivirals are currently available to treat HDV infection and antivirals against HBV do not ameliorate hepatitis D. Liver sodium taurocholate co-transporting polypeptide (NTCP) was recently identified as a common entry receptor for HDV and HBV in cell cultures. Here we show HDV can infect mice expressing human NTCP (hNTCP-Tg). Antibodies against critical regions of HBV envelope proteins blocked HDV infection in the hNTCP-Tg mice. The infection was acute yet HDV genome replication occurred efficiently, evident by the presence of antigenome RNA and edited RNA species specifying large delta antigen in the livers of infected mice. The resolution of HDV infection appears not dependent on adaptive immune response, but might be facilitated by innate immunity. Liver RNA-seq analyses of HDV infected hNTCP-Tg and type I interferon receptor 1 (IFNα/βR1) null hNTCP-Tg mice indicated that in addition to induction of type I IFN response, HDV infection was also associated with up-regulation of novel cellular genes that may modulate HDV infection. Our work has thus proved the concept that NTCP is a functional receptor for HDV infection in vivo and established a convenient small animal model for investigation of HDV pathogenesis and evaluation of antiviral therapeutics against the early steps of infection for this important human pathogen.
| Currently 15 million people worldwide are infected by hepatitis D virus (HDV). HDV is the smallest virus known to infect human. With co-infection of its helper hepatitis B virus (HBV), viral hepatitis D is considered as the most severe form of viral hepatitis. No specific anti-HDV drugs are available; antivirals against HBV do not ameliorate hepatitis D. We report mice expressing a human bile acids transporter sodium taurocholate co-transporting polypeptide (NTCP) in the liver support HDV infection, providing a useful model for studying antivirals against HDV and understanding how the simplest virus interacts with a host in vivo. Our transcriptome analyses of livers of infected mice have unveiled interaction landscape of HDV and the hosts, laying a foundation for further studies.
| Hepatitis D virus (HDV) is the smallest virus known to infect humans with a single—stranded, circular RNA genome of about 1.7 kilobases in length. It is enveloped by surface proteins from its helper hepatitis B virus (HBV) and undergoes a unique replication cycle via an intermediate, antigenomic RNA [1,2]. Prevalence of HDV infection remains high in many areas around the world despite the implementation of vaccine programs against HBV. Currently 15 million people are infected by HDV among the 240 million chronic HBV carriers. Viral hepatitis D is considered as one of the most severe forms of human viral hepatitis. However, there are no specific antivirals available for clinical treatment of the infection and antiviral therapies against HBV do not ameliorate hepatitis D. The mechanisms that determine whether an individual clears HDV spontaneously or becomes chronically infected are unclear [3,4]. Understanding HDV infection and developing antivirals against HDV have been hampered by the lack of reliable cell culture systems and convenient small animal models susceptible for efficient HDV infection.
Sodium taurocholate co—transporting polypeptide (NTCP), a multiple transmembrane transporter predominantly expressed in the liver [5], was recently identified as a common entry receptor for HDV and HBV [6]. This finding has enabled cell culture systems that support HDV and HBV infection in vitro. For instance, exogenous expression of human NTCP (hNTCP) rendered HDV infection of multiple mammalian cell lines, while successful HBV infection has only been achieved in hNTCP—expressing human hepatoma cells [6–9]. It is reasonable to speculate that additional human hepatocyte—specific factors are required for HBV infection of mice, however, transgenic expression of hNTCP may confer susceptibility of mouse hepatocytes to de novo HDV infection, which may provide a much—needed convenient small animal model for investigation of HDV pathogenesis and evaluation of antiviral drugs against HDV in vivo. In addition, as no other virus is simpler than HDV yet still can infect mammals, studying HDV infection in a susceptible mouse model may also help to illuminate how an animal reacts to the invading of the smallest pathogen.
We report herein that transgenic mice expressing hNTCP in hepatocytes, designated as hNTCP—Tg, support de novo HDV infection. Active HDV genome replication in the livers of infected mice was demonstrated by the presence of antigenomic RNA and edited RNA species. Infection kinetic studies revealed that HDV infection of hNTCP-Tg mice was acute and age—dependent. The infection was efficiently blocked by monoclonal antibodies specifically recognizing the critical regions of HBV envelope proteins. In our efforts toward unraveling the mechanism underlying the resolution of HDV infection in the hNTCP-Tg mice, we obtained evidence suggesting that adaptive immunity was not required for the clearance of HDV infection in the mouse model. Instead, HDV infection of hNTCP-Tg mice induced a type-I interferon (IFN) response that might have contributed to the suppression of HDV replication. Intriguingly, HDV infection could also be efficiently cleared in hNTCP-Tg type I interferon receptor 1 (IFNα/βR1) null mice. RNA—seq analyses of liver transcriptome of the HDV infected hNTCP-Tg and hNTCP-Tg/IFNα/βR1-/- mice revealed that, in addition to known interferon—stimulated genes (ISGs), HDV infection was also associated with up—regulation of novel cellular genes yet uncharacterized for antiviral activity.
We and others previously demonstrated that HDV infection is restricted by murine Ntcp in cell cultures, various mammalian cells complemented with human— but not mouse NTCP supported HDV infection [7,8]. To generate a mouse model for HDV infection, we created C57BL/6 mouse lines expressing hNTCP with a C—terminal tag (C9), driven by a mouse albumin enhancer/promoter (Fig 1A). Mice were screened for the transgene by real—time PCR with primers specific for the human NTCP, and the expression levels of hNTCP mRNA in the liver were examined by real—time PCR after reverse transcription (qRT-PCR). One hNTCP transgenic (hNTCP-Tg) C57BL/6 mouse line that was confirmed for germline transmission of the hNTCP transgene by Southern blot analysis (Fig 1B) and expression of high level human NTCP mRNA in liver (Fig 1C—left) was selected for breeding and subsequent experiments. The expression level of hNTCP transgene was similar to that of the endogenous murine Ntcp (Fig 1C-right). There was no significant difference in the expression of hNTCP transgene between the homozygotes and heterozygotes at both mRNA level quantified by qRT-PCR (Fig 1C—right) and protein level examined by Western blot using antibodies against the C9 tag (Fig 1D). Immunofluorescent staining of liver sections with a C9—tag specific antibody showed hNTCP only expressed in the hepatocytes of hNTCP transgenic mice (Fig 1E).
Homozygous and heterozygous hNTCP-Tg C57BL/6 mice were both tested for their susceptibility to HDV infection. The viral infection efficiency in hNTCP-Tg mice of 9–10 days old (n = 10) correlated with the dose of inoculating virus and was independent of the hNTCP transgene homozygosity or gender of the transgenic mice (Fig 2A). At the highest HDV dose (5×1010 genome equivalents, GEq) tested, no infection was observed in the wild—type littermates (n = 6) that shared the same genetic background, microbiota and environment with the hNTCP-Tg mice. In the hNTCP-Tg mice, approximately 3% hepatocyte was being infected as examined by immunofluorescent staining positive for HDV delta antigens and the infection occurred in randomly scattered hepatocytes (Fig 2B). No significant liver histopathological changes were observed in the infected mice. There was only modest apoptosis, which was about 0.8% of the total cells, in the liver of HDV infected mice (S1 Fig). The HDV infection of hNTCP-Tg mice could be efficiently blocked by monoclonal antibody 2D3 specifically recognizing the pre-S1 domain of HBV large envelope protein (n = 6) [6], but not a control antibody (1C10) recognizing the core protein of HBV (n = 6) (Fig 2C). Similarly, a monoclonal antibody 17B9 targeting the S domain of the HBV envelope [6] that presumably attaches with the heparan sulfate proteoglycan on hepatocytes [10] also greatly reduced HDV infection in the mice (n = 5) (Fig 2D). These results show that HDV infection of hNTCP-Tg mice depends on both the pre-S1 and the S region of HBV envelop proteins, and suggest these animals are useful for evaluating inhibitors against HDV entry. HDV undergoes a unique replication cycle via an intermediate, antigenomic RNA [11]. In the hNTCP-Tg but not wild—type littermates, both genomic and antigenomic HDV RNA were readily detectable by Northern blot analysis (Fig 2E), indicating HDV effectively replicated in the hNTCP-Tg mice.
We next tested the susceptibility of the hNTCP-Tg mice to HDV infection at different age. Interestingly, while challenge of hNTCP-Tg mice younger than 17 days by intraperitoneal (i.p.) injection resulted in marked HDV infection, as indicated by the presence of approximately 1000 copies of HDV RNAs per cell (~106 copies/20ng liver total RNA) at 9 days post infection in the livers of mice (S2A Fig), challenge of the transgenic mice older than 4 weeks with HDV failed to establish effective infection (S2B and S2C Fig), although these mice efficiently expressed hNTCP in the livers regardless of their genotype of being homozygote or heterozygote of the transgene. Together these results demonstrate that hNTCP transgenic mice support HDV entry and RNA replication in hepatocytes in vivo in an age—dependent manner.
Because HDV infection of hNTCP-Tg mice should only result in a single—round infection of hepatocytes, it is interesting to know how the host immune system responds to the virus infection. In addressing this question, we first determined the kinetics of HDV replication in the liver of transgenic mice. hNTCP-Tg mice were infected by about 1×1010 GEq of HDV at 9 day after birth and sacrificed on 2, 6, 10, 14, and 18 days post infection. Intrahepatic HDV RNA reached peak level around 6 day post infection and then declined from the peak by approximately 1000 fold within next 12 days (Fig 3A). These observations suggest that HDV infection of the hNTCP-Tg mice is transient.
To explore host factors responsible for controlling the HDV infection in vivo, we first investigated the role of adaptive immunity. Specifically, a hNTCP-Tg scid mouse line was established by crossing hNTCP-Tg with adaptive immunity deficient Prkdcscid mice which bear a premature stop codon in the Prkdc gene whereby the differentiation of lymphoid cells was disrupted in these mice [12]. Similar to that observed in hNTCP-Tg mice, HDV infection was cleared rapidly in hNTCP-Tg mice with a homozygous Prkdc gene mutation (hNTCP / Prkdcscid) (n = 8) (Fig 3B), indicating that adaptive immunity is dispensable for viral clearance in these mice. Interestingly, on day 6 after viral inoculation, intrahepatic HDV RNA level in the hNTCP+/+/Prkdcscid mice was significantly lower than that of the hNTCP+/+ mice (Fig 3C); perhaps a higher level of innate immunity activity in the Prkdcscid mice may have contributed to the rapid HDV clearance at early time [13,14].
We further examined HDV infection in hNTCP-Tg mice with deficiency in IFNα/βR1 (hNTCP-Tg/ IFNα/βR1 -/-), which were established by crossing hNTCP-Tg with IFNα/βR1 null mice. It’s known that IFNα/βR1 is essential for type I IFN-mediated signal transduction and its deficiency greatly reduces host antiviral activities in general [15,16]. Consistent with this notion, the efficiency of HDV infection in hNTCP-Tg/IFNα/βR1 -/- mice was significantly higher than that in the normal hNTCP-Tg mice examined by qPCR (Fig 4A) and immunofluorescent staining of HDV delta antigen (S3A Fig),with about 10 folds increase of HDV RNA copies and 3–5 folds of more delta antigen positive cells, respectively. However, to our surprise, HDV infection was also efficiently cleared within 2 weeks in the hNTCP-Tg/IFNα/βR1 -/- mice (n = 14) (Fig 4B), suggesting that clearance of HDV infection in the transgenic mice may also be achieved via type I IFN-independent mechanism(s). In order to further investigate the mechanisms of HDV clearance in the infected mice, we performed RNA—seq analysis to capture the transcriptomic landscapes in the liver of different hNTCP-Tg mouse lines upon HDV infection (n = 3 in each group). The cellular factors that mediate the type I interferon antiviral response are ISGs [17]. We first compiled a list of 583 mouse ISGs based on the previously reported datasets [18,19], and analyzed their expression fold changes in the infected mice (S1 Table). Comparing to the mock—infected hNTCP-Tg mice, hNTCP-Tg mice infected by HDV exhibited an elevated level of ISGs in the liver sample (S3B Fig). A dot plot for ISG expression fold change is shown in Fig 4C. Ifit1, Ifi44, Rsad2, Ccl7, Slfn1, Isg15, Mx1, Tgtp1, Gbp3, Ifit3, Ifit2, Ddx60, Oasl1, Zbp1, Oasl2, Cxcl10 and Irf7 were among the most significantly up—regulated ISGs in the hNTCP-Tg mice comparing to the wild—type mice similarly inoculated with HDV (Fig 4C, left). In marked contrast, no ISGs were significantly induced in HDV—infected hNTCP-Tg/IFNα/βR1 -/- mice, although the virus infection was also efficiently cleared in those mice (Fig 4C, right). Together these results indicate that although HDV infection of hNTCP-Tg mice induces a type I IFN response, which may subsequently suppress the replication of the virus, other cellular factors may mediate IFN—independent clearance of HDV infection.
To identify cellular factors mediating IFN—independent innate immune response against HDV, a systematic investigation using hierarchical clustering analysis of total 7802 genes expressed in the livers of infected mice was performed. The transcriptome analysis revealed multiple genes were up—regulated in HDV inoculated hNTCP-Tg/IFNα/βR1 -/- mice. Among them, 22 genes (Gm26130, mt-Ts2, Mgst2, Cyp7a1, Vps45, etc.) were clustered as a hot block next to a block containing Irf7 gene that is the master regulator for the induction of type I interferon during viral infection (Honda et al., 2005) (S4 Fig). Interestingly, most of these 22 genes apparently do not have a known function in the host anti—pathogen process or innate immunity. Together, these results unveiled the interaction landscape of HDV and the hosts, and they also indicate that multiple ISGs in hNTCP-Tg and additional novel cellular factors identified in hNTCP-Tg/ IFNα/βR1 null mice may be relevant to the clearance of HDV infection in the animals.
HDV RNA editing is a crucial step late in theHDV lifecycle for switching from viral RNA replication to packaging. HDV RNA editing was detected in mouse liver injected with copies of HDV DNA genome using hydrodynamics—based transfection in vivo [20]. It has been shown by in vitro studies that HDV RNA editing is controlled by ADAR1 which is also an ISG [21] [22]. Using the novel mouse models, we examined the levels of ADAR1 and HDV editing upon the viral infection in vivo. The expression level of both ADAR1 variants, ADAR1S and ADAR1L, was induced in hNTCP-Tg mice upon the infection (Fig 4D). Consistent with ADAR2 not being an ISG, HDV infection did not induce ADAR2 level in hNTCP-Tg mice. Intriguingly, the degree of HDV RNA editing was comparable in hNTCP-Tg/IFNα/βR1-/- and hNTCP-Tg mice (S5A and S5B Fig), hence it is tempting to speculate that the baseline expression of ADAR1 in these mice may be sufficient for HDV RNA editing. We further examined the degree of viral RNA editing as a function of time upon de novo HDV infection of hNTCP-Tg mice. The result showed the degree of viral RNA editing in the animals increased from day 6 to day 18 after infection; nonetheless, during the entire experiment period, only a small fraction of viral RNA was edited with the highest ratio of 4.6% on day 18 after viral infection (S5C Fig).
HDV is enveloped by glycoproteins derived from HBV [23], and it shares species restriction with its helper virus HBV at entry level. Mice are not a natural host for both HDV and HBV and do not support de novo infection by either virus. In this study, we demonstrated that hNTCP-Tg C57BL/6 mice can be infected by HDV. Unlike HDV, HBV infection may be restricted by unknown host factors in mice as successful HBV infection has only been achieved in hNTCP expressing hepatoma cells from human but not mouse [7,8]. It is speculated that in addition to hNTCP that facilitates HBV entry, other human hepatocyte—specific factors may be required to enable the mouse hepatocytes to support efficient formation of HBV cccDNA, which is essential for establishment of HBV infection [24,25]. Nevertheless, our work reported herein provides strong genetic evidence suggesting that hNTCP is a functional receptor for HDV infection in vivo.
Both human— and mouse NTCP can co—transport bile salts from circulation into hepatocytes with sodium [26]. In agreement with the results from cell—based studies suggesting that mouse Ntcp is not a functional receptor for HDV, the wild—type neonatal C57BL/6 mice which express mouse Ntcp at a level similar to that of hNTCP in the hNTCP-Tg littermates, did not support HDV infection. In contrast, mice bearing hNTCP, irrespective of the sex or the homozygosis of hNTCP transgene, were readily susceptible to HDV infection. Intriguingly, the receptor binding pre-S1 lipopeptide was shown to be able to bind to mouse hepatocyte in vitro [27] and in vivo [28], and to mouse NTCP albeit at a lower efficiency as elucidated by us [7]. The mRNA level of hNTCP was comparable to that of endogenous mNTCP in the hNTCP-Tg mice. It is unclear from current study whether the endogenous mNTCP competes for the pre-S1 domain mediated HDV interaction with hNTCP and thereby negatively affects the viral infection. Apparently mNTCP did not exert a trans—dominant negative effect on hNTCP in the transgenic mice. It will be interesting to compare HDV infection efficiency between hNTCP-Tg mice and mice with their endogenous NTCP genes replaced by hNTCP. Nonetheless, HDV effectively replicated in the liver of hNTCP transgenic mice, Northern blot analysis readily detected antigenomic RNA that is the intermediate and a diagnostic marker of HDV replication. Moreover, although quantifying the degree of RNA editing was limited by the resolution of the RNA editing assays, the study showed that HDV underwent evident RNA editing, an event essential for production of large delta antigen and switch to viral assembly, in the hNTCP-Tg mice during the in vivo infection. Importantly, the HDV infection of hNTCP-Tg mice could be effectively blocked by monoclonal antibodies recognizing either pre-S1 or S domain of HBV envelope proteins, suggesting that interactions between pre-S1 and hNTCP as well as S and heparan sulfate proteoglycans on hepatocytes are essential for HDV infection in vivo.
In contrast to immune—deficient uPA/SCID mice implanted with human hepatocytes, which have been used for studying HDV infection and drug candidate evaluation [29,30], hNTCP transgenic mice and their derivatives are heritable, easier to handle and more consistent among individual animals. They can serve as valuable and convenient models for evaluating antivirals, in particular HDV entry inhibitors. Because entry of HBV and HDV are both mediated by envelope proteins of HBV, hNTCP-Tg mice thus may also be used for evaluating HBV entry inhibitors using HDV as a surrogate. Moreover, by crossing with other mice bearing well—defined mutation(s) of various immunodeficiency and large—scale analysis of liver transcriptome, they created an unprecedented opportunity for in—depth studies of HDV viral infection and the host immune defense against HDV infection in vivo. In fact, our work reported herein has already revealed several unique characteristics of HDV infection in the transgenic mice.
First, we showed that neonatal but not adult hNTCP-Tg C57BL/6 mice supported readily detectable HDV infection in the liver. It is known that significant differences exist between adults and neonates in innate as well as adaptive immunity. For example, the neonatal innate immune system is biased against the production of pro—inflammatory cytokines [31] and dendritic cells (DC) may be immature until about 5 weeks of age [32]. In addition, age—dependent susceptibility of mice to virus infection has been reported for many different viruses and frequently the infection efficiency is related to the mouse genetic background [33–35]. It remains to be tested whether the HDV infection efficiency differs by age in other mouse strain(s). In addition, effects of intrahepatic immunity maturation and other age—dependent physiological changes on the susceptibility of HDV infection can also not be ruled out [36,37].
Second, concerning the infection efficiency of HDV infection in mice in vivo, we showed that inoculating hNTCP-Tg C57BL/6 mice at 9 days after birth with 3.3X1010 mge of HDV resulted in about 3% cells infected by the virus as indicated by the immunofluorescent staining of the delta antigen. It was reported that passage of HDV to woodchucks chronically infected by WHV could infect 10 to 40% hepatocyte, depending on whether the inoculated virus was first or second passage of HDV in woodchucks [38]. However, as there was no HBV infection in hNTCP-Tg mice, HDV only underwent single round infection in the mouse model reported here. The observed infection rate in the animals may also be affected by other factors, such as the route of inoculation and variations among HDV preparations. Direct intravenous injection of the virus may increase the infection rate in hNTCP-Tg mice, but it was not feasible for the 9-days animals.
Third, we observed that HDV infection of hNTCP-Tg mice was transient, irrespective to the status of their immune competency (with Prkdcscid or IFNα/βR1-/-). Previous studies reported that no significant liver histopathological changes were found in HDV transgenic mice [39,40] or in experimental HDV infected chimpanzees [41]. Reports on experimental HDV inoculation into neonatal or SCID mice with WHV enveloped HDV, which took advantage of HDV ribonucleoprotein’s compatibility with WHV envelops thereby bypassing the species restriction at entry level of HDV, showed a transient infection in the livers of infected animals [42]. We showed herein that the receptor mediated, de novo infection of HDV was cleared in about two weeks in the hNTCP-Tg mice upon viral inoculation with no obvious liver pathological changes. Interestingly, only few cells positive for HDV delta antigen were found to be TUNEL positive in the liver samples of infected mice. More studies, ideally using hNTCP-Tg mice deficient in hepatic apoptosis or necrosis, are needed to clarify whether HDV infection results in the death of the hepatocytes in the mice. Surprisingly, intrahepatic HDV RNA was cleared after infection at comparable kinetics among normal, homozygous Prdkcscid and IFNα/βR1-/- hNTCP-Tg mice; this suggested that the clearance of HDV infection in hNTCP-Tg mice was either due to the activation of innate immune response or accumulation of large delta antigen that suppressed HDV RNA replication. However, the latter hypothesis is not supported by a recent finding that HDV mono—infection of immune deficient (SCID/beige) mice transplanted with human hepatocytes persisted intrahepatically for more than 6 weeks [29]. It will be interesting to further investigate the underlying mechanisms controlling the apparently different outcome of the HDV infection in the hNTCP transgenic versus the human hepatocyte—transplanted mouse models, for example whether the difference is due to the activity of NK cells presented in the hNTCP-Tg/Prdkcscid but not in the xenotransplanted SCID/beige mice. Another possible explanation of the discrepancy between the two models is that HDV infection may induce production of cytokines or other soluble factors by non—parenchymal cells that species—specifically inhibit HDV replication in hepatocytes of the infected mice.
Forth, two lines of independent evidence presented in this study strongly suggest that HDV infection of hNTCP-Tg mice induces a type I IFN response suppressing HDV replication in hepatocytes. Firstly, intrahepatic HDV RNA in IFNα/βR1 null hNTCP-Tg mice is about 10-fold higher than that in normal hNTCP-Tg mice. In addition, dozens ISGs, among which some have been characterized for their antiviral activities against various other viral infections, for example Mx1, Ifit1, Isg15, Ifi44, Ddx60, Oasl (Liu et al., 2012; Schoggins et al., 2011) and Irf7 were up-regulated upon HDV inoculation in hNTCP-Tg mice. This is the first report demonstrating that HDV infection induces an early activation of IFN response in vivo. Of note, Hartwig et al. showed that treatment of cultured cells with IFNα increased both ADAR expression levels and RNA editing [43] and enhanced HDV RNA editing was shown to increase the expression level of large delta antigen which could further restrict the HDV replication in cells [44]. It will be interesting to further investigate how the type I IFN response restricts HDV replication in hepatocytes in vivo.
Finally, liver RNA-seq analyses of HDV-infected normal and IFNα/βR1 null hNTCP-Tg mice also revealed that expression of additional cellular genes was associated with HDV infection. The majority of these genes have not been characterized for activity in immune response or antiviral infection. Interestingly, 22 genes including Gm26130, a snoRNA gene, and Cyp7a1, the rate-limiting enzyme in the synthesis of bile acid from cholesterol via the classic pathway, and several mitochondrial tRNAs are clustered in the analysis of 7802 liver genes of the infected mice. Although further experiments are needed to dissect the possible antiviral roles of these molecules, it is tempting to speculate that at least some of them may function in parallel or in addition to the known ISGs, and be relevant in HDV viral clearance. Of note, as the smallest virus known to infect humans, HDV encodes only one protein (delta antigen), which modulates viral replication through interaction with cellular DNA-dependent RNA polymerases and other host factors [1,2]. Studying of HDV infection in hNTCP-Tg mouse model thus opens a unique door for understanding how an animal reacts to invasion by the smallest viral pathogen.
In summary, our studies of HDV infection in hNTCP-Tg mice not only proved that NTCP is a functional receptor for HDV infection in vivo and hNTCP-Tg mouse is a useful model for studying antivirals against the infection, and they also shed new light on the interaction between HDV and host immunity, and laid a foundation for future investigation toward better understanding the pathogenesis of HDV infection.
All animals were housed in the animal facility of the National Institute of Biological Sciences (NIBS), Beijing. Animal experiments were conducted following the National Guidelines for Housing and Care of Laboratory Animals and performed in accordance with NIBS institutional regulations after approved by the institution's Institutional Animal Care and Use Committee (IACUC). The protocol number is NIBS-0012.
The human NTCP gene with a C9 tag [6] was cloned into a vector with an expression cassette driven by mouse albumin enhancer/promoter. The recombinant plasmid was linearized and introduced into the pronuclei of C57BL/6NCrlVr mouse zygotes. PCR primers for identifying hNTCP transgene are hNTCP-F (5′- GGATAGGGATCCGCCACCATGGAGGCCCACAACGCG-3′) and hNTCP-BGH-R (5′-ATTTCCCTCGA GCCATAGAGCCCACCGCAT-3′).
Fox Chase SCID® (CB17/Icr-Prkdcscid/IcrlcoCrlVr, homozygous for the severe combined immune deficiency spontaneous mutation Prkdc) mice were from the Vital River, Beijing, China; Interferon (alpha and beta) receptor 1 knockout (B6.129S2-Ifnar1tm1Agt/Mmjax) mice [15] backcrossed to C57BL/6 for at least 5 generations were from the Jackson Laboratory, Maine, USA. hNTCP transgenic mice homogeneous for Prkdc mutation or null for IFNα/βR1 were obtained by cross breed hNTCP transgenic mice with the corresponding immune deficient mice, respectively. The genotypes of the mice were determined by PCR with DNA isolated from mouse tail. Animals were hosted in an SPF mouse facility and all animal experiments were conducted following the national guidelines for housing and care of laboratory animals and performed in accordance with institutional regulations after approval by the IACUC at National Institute of Biological Sciences, Beijing.
RNA from mouse liver or kidney was extracted using TRIzol® Reagent (Invitrogen). The total RNA was reverse transcribed into cDNA with PrimeScriptTM RT-PCR Kit (Takara), cDNA obtained from 20ng RNA was used for real time PCR assay. See S1 Text. Supplemental experimental procedures for details. Western blot analysis for the expression of hNTCP was performed by using liver samples collected from hNTCP transgenic or wild-type mice, total 30 μg protein was loaded and 10μg/ml 1D4 antibody (Santa Cruz Biotech) was used for detecting the C9 tag fused at the C-terminus of the hNTCP transgene. Mouse GAPDH was used as a loading control.
HDV was produced as previously described by using two plasmids transfection in Huh-7 cells [6]. Mice were inoculated with purified HDV by intraperitoneal (i.p.) injection. To minimize the influence of variables, in each experiment (usually presented as one panel of a figure in the manuscript), mouse littermates were injected with HDV from same viral preparation. Mouse liver samples were homogenized in liquid nitrogen immediately after collection, and then lysed by TRIzol® reagent. The total RNA was reverse transcribed into cDNA with PrimeScriptTM RT-PCR Kit (Takara), cDNA from 20ng RNA was used for real time PCR assay. See S1 Text. Supplemental experimental procedures for details.
Liver tissue RNA was extracted using TRIzol® reagent. 2μg total RNA was electrophoresed through formaldehyde-containing 1% agarose gels, blotted onto a nitrocellulose membrane (Hybond-C Extra, Amersham), and hybridized with digoxigenin (DIG)-labeled RNA probes for HDV genome, HDV antigenome, or mouse GAPDH, respectively. For detecting HDV RNA editing, Nco I restriction digestion of PCR-amplified cDNA derived from HDV RNA was used. 300 ng total RNA was reverse transcribed using random hexamers, followed by PCR using primers specific to HDV cDNAs. PCR products were purified and subjected to overnight digestion with restriction enzyme Nco I (NEB). The total digestion products were separated by 4% polyacrylamide gel electrophoresis (PAGE), and the gel was stained with silver nitrate. In independent experiments, HDV RNA editing was also examined using the [32P] dCTP labeling method as reported by Casey et al [45] or quantified by microcapillary electrophoresis analysis using an Agilent 2100 bioanalyzer. See S1 Text. Supplemental experimental procedures for details.
RNA-seq analysis was conducted using the total RNA of livers from 3 mock-inoculated hNTCP transgenic (hNTCP+/-) mice, and HDV-inoculated mice including three hNTCP+/-, three hNTCP+/-/IFNα/βR-/-, and three wild-type mice. Mice were inoculated on day 9 after birth, and sacrificed 6 days after the inoculation. RNA-seq was performed using Illumina Genome Analyzer IIx system. Sequence data was deposited at Sequence Read Archive (SRA) of the NCBI under BioProject PRJNA236433. See S1 Text. Supplemental experimental procedures for details.
|
10.1371/journal.pcbi.1006716 | State-aware detection of sensory stimuli in the cortex of the awake mouse | Cortical responses to sensory inputs vary across repeated presentations of identical stimuli, but how this trial-to-trial variability impacts detection of sensory inputs is not fully understood. Using multi-channel local field potential (LFP) recordings in primary somatosensory cortex (S1) of the awake mouse, we optimized a data-driven cortical state classifier to predict single-trial sensory-evoked responses, based on features of the spontaneous, ongoing LFP recorded across cortical layers. Our findings show that, by utilizing an ongoing prediction of the sensory response generated by this state classifier, an ideal observer improves overall detection accuracy and generates robust detection of sensory inputs across various states of ongoing cortical activity in the awake brain, which could have implications for variability in the performance of detection tasks across brain states.
| Establishing the link between neural activity and behavior is a central goal of neuroscience. One context in which to examine this link is in a sensory detection task, in which an animal is trained to report the presence of a barely perceptible sensory stimulus. In such tasks, both sensory responses in the brain and behavioral responses are highly variable. A simple hypothesis, originating in signal detection theory, is that perceived inputs generate neural activity that cross some threshold for detection. According to this hypothesis, sensory response variability would predict behavioral variability, but previous studies have not born out this prediction. Further complicating the picture, sensory response variability is partially dependent on the ongoing state of cortical activity, and we wondered whether this could resolve the mismatch between response variability and behavioral variability. Here, we use a computational approach to study an adaptive observer that utilizes an ongoing prediction of sensory responsiveness to detect sensory inputs. This observer has higher overall accuracy than the standard ideal observer. Moreover, because of the adaptation, the observer breaks the direct link between neural and behavioral variability, which could resolve discrepancies arising in past studies. We suggest new experiments to test our theory.
| The large majority of what we know about sensory cortex has been learned by averaging the response of individual neurons or groups of neurons across repeated presentations of sensory stimuli. However, multiple studies in the last three decades have clearly demonstrated that sensory-evoked activity in primary cortical areas varies across repeated presentations of a stimulus, particularly when the sensory stimulus is weak or near the threshold for sensory perception [1–3], and have suggested that this is an equally important aspect of sensory coding as the average response [4–6]. Variability is thought to arise from a complex network-level interaction between sensory-driven synaptic inputs and ongoing cortical activity, and single-trial response variability is partially predictable from the ongoing activity at the time of stimulation. A large body of work has focused on characterizing this relationship between notions of cortical “state” and sensory-evoked responses [7–13], establishing some simple models of local cortical dynamics [14]. Less is known about the impact of this relationship for downstream circuits (though see [15,16]).
As an example, consider the detection of a sensory stimulus, which has been foundational in the human [17–22] and non-human primate psychophysical literature [23,24] and serves as one of the most widely utilized behavioral paradigms in rodent literature [25–27]. In an attempt to link the underlying neural variability to behavior, the principal framework for describing sensory perception of stimuli near the physical limits of detectability is signal detection theory [28]. A key prediction of signal detection theory is that, on single trials, detection of the stimulus is determined by whether the neural response to the stimulus crosses a threshold. Particularly large responses would be detected but smaller responses would not, so variability in neural responses would lead to, and perhaps predict, variability in the behavioral response. From the perspective of an ideal observer, if variability in the sensory-evoked response can be forecasted using knowledge of cortical state, the observer could potentially make better inferences, but in traditional (state-blind) observer analysis, the readout of the ideal observer is not tied to the ongoing cortical state.
In this work, using network activity recordings from the whisker sensitive region of the primary somatosensory cortex in the awake mouse, we develop a data-driven framework that predicts the trial-by-trial variability in sensory-evoked responses in cortex by classifying ongoing activity into discrete states that are associated with particular patterns of response. The classifier takes as inputs features of network activity that are known to be predictive of single-trial response from previous studies [9,14], as well as more complex spatial combinations of such features across cortical layers, to generate ongoing discrete classifications of cortical state. We optimize the performance of this state classifier by systematically varying the selection of predictors. Finally, embedding this classification of state in a state-aware ideal observer analysis of the detectability of the sensory-evoked responses, we analyze a downstream readout that changes its detection criterion as a function of the current state. We find that state-aware observers outperform state-blind observers and, further, that they equalize the detection accuracy across states. Downstream networks in the brain could use such an adaptive strategy to support robust sensory detection despite ongoing fluctuations in sensory responsiveness during changes in brain state.
To directly assess the relationship between ongoing cortical activity and variability in the sensory-evoked cortical response, we recorded extracellular activity across layers of cortex in the awake head-fixed mouse. Specifically, spontaneous and sensory-evoked local field potentials (LFPs) were recorded using a 32-channel laminar array targeted to the region of the primary somatosensory cortex corresponding to facial vibrissae (S1 barrel cortex, Fig 1A). Mice were subjected to brief single-whisker deflections (11 recordings in 6 mice; average 438 (196–616) trials per recording). Intrinsic optical signal imaging was performed to locate the barrel column corresponding to the stimulated whisker. The sensory stimulus (Fig 1B, top) was a computer-controlled punctate deflection in the caudo-rostral plane (see Methods), designed to emulate velocity transients observed during ‘stick-slip’ events in rodents whisking across surfaces [29–31]. Cortical layers were assigned based on the trial-averaged spatial profile of the sensory-evoked LFP and current source density (CSD) responses, with layer 4 centered on the largest evoked response in the LFP and the large, early current sink in the CSD (Fig 1C, see Methods). Putative boundaries between layer 4 and layer 2/3 or layer 5 were based on published laminar dimensions [32].
Sensory-evoked responses in layer 4 were variable: mean amplitude of the response in layer 4 was a negative dip of 0.81 mV (+/- 0.35 across recordings, N = 11), and the standard deviation of evoked response size across trials was 0.45 mV (average SE across recordings, N = 11). We examined the impact of such variability on the detectability of sensory inputs in the framework of ideal observer analysis, which is conceptually presented in Fig 1D. In this scenario (Fig 1D, top row), a sensory stimulus W takes on one of two possible values: “+” in the case that a sensory input was present and “−” in the absence of a sensory input. Neural activity (x, see Methods), either spontaneous (W is “−”) or generated by the stimulus (W is “+”), was variable across trials and described by a conditional distribution P(x|W). The task of a downstream network, imagined here as an ideal observer, was to determine from the neural activity whether or not there was a stimulus. In the classical signal detection framework, this was envisioned as the observed activity arising from one of two distributions, P(x|W = "−") or P(x|W = "+"). The ideal observer ascribes activity above a chosen threshold (to the right of dashed red line) as belonging to P(x|W = "+"), and thus concludes that a stimulus was present, and otherwise as belonging to P(x|W = "−"), and thus the stimulus was deemed absent. In this work, we considered an alternative perspective, which is that the ideal observer was “state-aware.” That is, we considered the case in which the response distribution P(x|W, s) depended upon ongoing activity (“state,” s) as well as the stimulus (Fig 1D, blue, bottom row). In this case, the discrete state (s^) is classified from the recorded, ongoing cortical activity, which is subsequently used by the state-aware observer to set the detection threshold independently for each state.
To illustrate how this framework operates, we show a set of example trials from a single recording in Fig 1E–1H. Two of these examples of layer 4 LFP activity show responses to a whisker input and one is a segment of spontaneous activity (Fig 1E). Across the stimulus-evoked responses, we observed significant variability in the overall size of sensory-evoked response (Fig 1F). Moreover, one of the evoked responses (Fig 1F, top row) is smaller than a spontaneously occurring LFP event (Fig 1F, bottom row). We also note that pre-stimulus LFP activity is quite different across these recordings. The goal of the state classifier is to find consistent relationships between features of the ongoing activity and the details of the single-trial sensory-evoked response. Assuming for the moment that it can do so, the state classifier would classify the pre-stimulus activity for these responses (Fig 1E) into separate states (s^ = 1 or 2, Fig 1G). When tasked with detecting or rejecting stimuli on the basis of the LFP response (Fig 1H), an ideal observer sets a single threshold for detection, which causes it to fail to detect a true sensory response while generating a false alarm on the spontaneous fluctuation (Fig 1H, black). In contrast, the state-aware observer sets its detection criterion separately for each state. For this example, the threshold may be lowered for state 1 and raised for state 2, thus the sensory-evoked responses would be detected but the spontaneous fluctuation would be rejected (Fig 1H, blue).
The state-aware observer thus has two distinct stages: state classification and sensory stimulus detection. The idea is that, by adapting its criterion for detection in accordance with the expected response, the state-aware observer will more reliably detect sensory-evoked responses and reject spontaneous fluctuations. Overall, the success of this strategy depends on a state classifier that predicts variation in the future sensory-evoked response, so we first optimized classification models with the goal of identifying the most relevant features of ongoing activity for the prediction of the details of the sensory-evoked responses. We then use this framework to classify ongoing activity into states and compare traditional (state-blind) and state-aware observers to determine how using this prediction to adjust detection strategy impacts overall detection performance.
The foundation upon which the state-aware observer is constructed is a prediction of the sensory-evoked cortical response. This prediction is based on classifying elements of the ongoing, pre-stimulus activity into discrete “states,” and the goal is to find the features of ongoing activity and the classification rules that generate the best prediction of sensory-evoked responses. Treating this as a discrete problem was a methodological choice motivated by the rationale that such an approach could find rules that are not linear in the features of ongoing activity and could lend more flexibility in the rules relating features of ongoing activity to variability in the response. The features of ongoing activity include the power spectrum of pre-stimulus LFP and the instantaneous “LFP activation” (Fig 2A). To describe sensory-evoked responses, we define a parameterization of the LFP response using principal components analysis (Fig 2B). The state classifier is a function that takes as inputs features of pre-stimulus LFP and produces an estimate of the principal component (PC) weights and thus of the single-trial evoked response (Fig 2C). In the following sections, we describe this process in detail.
Next, within the general class of pre-stimulus features considered–power ratio and LFP activation–we optimized several choices: the range of frequencies used to compute the power ratio; the cortical depth from which the ongoing LFP signal is taken; and possible combinations of LFP signals across the cortical depth. Changes in pre-stimulus features resulted in changes in the boundaries between states, and ultimately in changes in prediction performance.
First, we varied the bounds of the low-frequency range (“L range”, Fig 3A). The increase in fVE was on average 0.09 ± 0.05 (N = 11 recordings) (Fig 3B; classifier boundaries shown in S2 Fig), with a significant increase in 10 of 11 recordings (Fig 3C, asterisks). We found that the optimal L range could extend to frequencies up to 40 Hz (Fig 3C), with the median bounds of the optimal L being from 1 to 27 Hz.
Using for each recording the power ratio based on the optimized range of low-frequency power (Fig 3), we next determined where along the cortical depth the most predictive activity was and whether taking spatial combinations of LFP activity could improve the prediction. Note that in this analysis, the channel for the stimulus-evoked response was held fixed (L4) and thus the parameterization of the evoked response using principal components did not change, but the pre-stimulus channel was varied. For each recording, we thus built a series of classifiers, using single- and multi-channel LFP activity from across the array (Fig 4A, S3 Fig), which again were optimized for prediction of the single-trial L4 sensory-evoked response.
Classifiers built from a single channel of LFP performed best when the channel was near L4 (Fig 4B, single example; Fig 4C, average profile). Because the LFP represents a volume-conducted signal, we also examined the current source density (CSD) [34–36], estimated on single trials using the kernel method [37]. There was no improvement in fVE using CSD to build classifiers (fVE difference, CSD minus LFP: -0.07; range: (-0.12, -0.01)).
For each recording, we defined an optimal classifier channel based on the spatial profile of fVE for single-channel predictors (Fig 4B; S3 Fig). In the “pair” combination, we paired the optimal classifier channel with each of the other possible 31 channels (Fig 4B; green dashed line). We optimized the classifier in the 3-dimensional space defined by power ratio (on the optimal channel only) and LFP activation from each of the two channels and compared the fVE to that obtained using the optimal classifier channel only (Fig 4D). We found no improvement in the prediction using the pair combination compared to using the optimal channel alone (Fig 4D, mean fVE difference: 0.00 ± 0.01; 0/11 recordings with significant change, pair vs. single) or using more complex combinations of channels (S3 Fig).
To summarize, we optimized classifiers based on pre-stimulus features to predict single-trial sensory-evoked LFP responses in S1 cortex of awake mice. We found that the classifier performance was improved by changing the definition of the power ratio (L/W) such that the low-frequency range (L) extended from 1 Hz to 27 Hz, depending on the recording, which differed from the range typically used from anesthetized recordings in S1 (1–5 Hz) [8,9]. We also found that the most predictive pre-stimulus LFP activation was near layer 4.
After establishing a clear enhanced prediction of the single-trial stimulus-evoked response within the LFP by considering the pre-stimulus activity, we investigated the impact of this relationship on the detection of sensory stimuli from cortical LFP activity using a state-aware ideal observer analysis. We first considered a simple matched-filter detection scheme [38] in which the ideal observer operated by comparing single-trial evoked responses to the typical shape of the sensory evoked response (Methods, Detection). The matched filter was defined by the trial-average evoked LFP response, and this filtered the raw LFP (Fig 5A) to generate the LFP score (Fig 5B). For the state-blind observer, a detected event was defined as a peak in the LFP score that exceeded a fixed threshold (Fig 5B, stars). The LFP score distributions from time periods occurring during known stimulus-evoked responses and from the full spontaneous trace were clearly distinct but overlapping (Fig 5C), and detected events (Fig 5B, stars) included both “hits” (detection of a true sensory input) and “false alarms” (detection of a spontaneous fluctuation as a sensory input).
Next, using the state classifier constructed in the first half of the paper, we analyzed the performance of a state-aware observer on a reserved set of trials, separate from those used for fitting and optimizing the state classifiers (Methods). Specifically, using the optimized state classifier (Figs 3 and 4), we continuously classified “state” at each time point in the recording (Fig 5D). The state-aware observer detects events exceeding a threshold, which changed as a function of the current state (Fig 5E). Instead of a single LFP score distribution, we now have one for each predicted state (Fig 5F), leading to many possible strategies for setting the thresholds for detecting events across states.
In general, the overall hit rate and false alarm rate will depend on hits and false alarms in each individual state (Fig 6A and 6B for single example; S4 and S5 Figs show all recordings), as well as the overall fraction of time spent in each state (Fig 6A, inset). We walk through the analysis for a single example, selected as one of the clearest examples of how state-aware detection worked. While this example recording shows a relatively large improvement, it is not the recording with the largest improvement, and, moreover, the corresponding plots for all recordings are shown in S4 and S5 Figs. To compare between traditional (state-blind) and state-aware observers, we compared hit rates at a single false alarm rate, determined for each recording as the false alarm rate at which 80%-90% detection was achieved by a state-blind ideal observer. To select thresholds for the state-aware observer, we systematically varied the thresholds in state 1 and state 3, while adjusting the state-2 threshold such that average false alarm rate was held constant. For each combination of thresholds, we computed the overall hit rate (Fig 6C). For the example recording highlighted in Fig 6, the state-aware observer (hit rate: 96%) outperformed the traditional one (hit rate: 90%). This worked because the threshold in state 3 could be increased with very little decrease in the hit rate (Fig 6B), and this substantially decreased the false alarm rate in state 3 (Fig 6A). Because the overall false alarm rate is fixed, this meant more false alarms could be tolerated in states 1 and 2. Consequently, thresholds in states 1 and 2 could be decreased, which increased their hit rates. Across recordings, we found that the state-aware observer outperformed the state-blind observer in 9 of 11 recordings (Fig 6D; S4 and S5 Figs). Hit rates slightly but significantly increased from a baseline of 81% for the state-blind observer to 84% for state-aware detection, or an average change of +3 percentage points (SE: 3%; signed-rank test, p < 0.01, N = 11).
The overall change in hit rate reflects both the fraction of time spent in each state (some fixed feature of an individual mouse) and the changes in state-dependent hit rates. To separate these factors, we analyzed the hit rate of the state-blind and state-aware observers by computing, for each observer, the hit rate conditioned on each pre-stimulus state (Fig 6E). For this recording, the state-blind observer had very low hit rate in state 1 and high hit rates in states 2 and 3. In comparison, hit rates were similar across the three state for the state-aware observer (Fig 6D). Thus, in state 1 (smallest responses, blue), we observed a large increase in the hit rate depending on whether the observer used state-blind or state-aware thresholds. Averaged across all recordings, the state-1 hit rates increased from 60% to 76%, which is a relative increase of 26% (SE 11%). Because this is weighted by the fraction of time spent in state 1, the overall impact on the hit rate is smaller. Hit rates increased slightly on average in state 2 (+ 2%, SE 4%) and decreased slightly in state 3 (-7%, SE 9%). The net impact of this is that across the majority of recordings, the cross-state range of hit rates for the state-blind ideal observer was much larger than that for the state-aware ideal observer (Fig 6D and 6F; 19%, average state-blind minus state-aware hit rate range in percentage points (SE: 5%); p < 0.01, signed-rank test, N = 11). Thus, while the overall differences between state-aware and state-blind hit rates are modest, the state-aware observer has more consistent performance across all pre-stimulus states than a state-blind observer.
Due to the rapid development of tools that enable increasingly precise electrophysiology in the awake animal, there is a growing appreciation that the “awake brain state” encompasses a wide range of different states of ongoing cortical activity, and that this has a large potential impact on sensory representations during behavior [39–44]. Here, we constructed a framework for the prediction of highly variable, single-trial sensory-evoked responses in the awake mouse based on a data-driven classification of state from ongoing cortical activity. In related work, past studies have used some combination of LFP/MUA features to predict future evoked MUA response [9,14]. We used a similar approach for state classification and response prediction in cortical recordings in the awake animal, extending this to allow complex combinations of ongoing activity in space and different features of the pre-stimulus power spectrum as predictors. We found that simple features of pre-stimulus activity sufficed to enable state classification that yielded single-trial prediction of sensory evoked responses. These predictive features were analogous to the synchronization and phase variables found in previous studies [8,9,14], though we found a revised definition of synchronization was more predictive. In particular, we found that the very low-frequency band of the LFP power spectrum (1–5 Hz) was less predictive of single-trial evoked responses in our recordings than a wider band (e.g. 1 to 27 Hz). This is consistent with findings from a recent study [40] that surveyed the power spectrum of LFP across different behavioral states in the awake animal and demonstrated differences in the power spectrum between quiet and active wakefulness up to 20 Hz. While we have focused on the problem of state classification and prediction from the perspective of an internal observer utilizing neural activity alone, future work could investigate whether the state classifier is also tracking external markers of changes in state, such as those indicated by changes in pupil diameter [42,45], whisking [40], or other behavioral markers in the awake animal.
We fit classifiers for each individual recording rather than pooling responses across animals and recording sessions. The structure of the classification rules was similar across recordings, showing that the relationship between pre-stimulus features and evoked responses is robust. This suggests that a single classifier could be fit, once inputs and outputs are normalized to standard values. This normalization could be accomplished by determining the typical magnitude of LFP sensory responses and rescaling accordingly. Moreover, the ordered structure of the classification rules suggests that a continuous model of state, rather than a discrete model, would have worked as well. To implement as a continuous model, one would fit a regression of the evoked response coefficients using as independent variables LFP activation and power ratio. Judging by the classification boundaries shown in S1 and S2 Figs, keeping only linear terms in activation and power ratio would give a good prediction.
In its current formulation, this framework utilizes only the features of ongoing cortical activity that are reflected in the LFP in order to classify state and predict the evoked LFP response. Both as features underlying the state classifier and as the sensory-evoked response being predicted, LFP must be interpreted carefully, as the details of how underlying sinks and sources combine depend on the local anatomy and population spiking responses [46]. In barrel cortex, the early whisker-evoked LFP response (0 to 25 ms) is characterized by a current sink in L4 initially driven by thalamic inputs to cortex, but also reflecting cortical sources of activity: the evoked LFP is highly correlated with the layer-4 multi-unit activity response [47,48]. We restricted our predictive framework to the high degree of variability in this initial response. It remains to determine how LFP response variability is reflected in the sensory-evoked single-unit cortical spiking activity patterns. Further, regarding LFP as a predictor used by the state classifier, LFP is a mesoscopic marker of cortical state that neglects finer details of cortical state organization. In addition to establishing whether better predictions are made from more detailed representations of cortical state, it is an interesting question how microcircuit spiking dynamics are related to the mesoscopic markers of cortical state, or how much can be inferred about population spiking dynamics from the LFP. Finally, thalamic and cortical activity are tightly linked, and the results presented here may also reflect variations in ongoing thalamic activity. Disentangling thalamic and cortical sources of variability in the evoked response will require paired recordings and perturbative experimental approaches designed to address issues of causality.
In the second part of the paper, we used ideal observer analysis to show that state-aware observers, with oracle knowledge of the spontaneous, ongoing state fluctuations informative of the single-trial sensory-evoked response, can out-perform a state-blind ideal observer. Our analysis relied on classification of the markers of ongoing state. This is not to suggest that this specific estimation takes place in the brain, but instead could potentially be achieved dynamically by a downstream network through the biophysical properties of the circuitry. Theoretically, the gain and threshold of such a readout neuron or network could be dynamically modified on the basis of the ongoing activity as a biophysical manifestation of the adaptive state-aware ideal observer, though the identification of specific mechanisms was beyond the scope of the current study.
We found that the state-aware observer had higher accuracy than the traditional, state-blind observer, but the absolute gain in hit rate (at fixed false alarm rate) averaged across all states was modest. When pre-stimulus states were analyzed separately, however, we found that accuracy in the low-response state was substantially higher for the state-aware observer, where there was a relative increase of 25% in the hit rate for this state. Because small sensory responses are predictable from the ongoing activity, transiently lowering the threshold for detection resulted in more “hits” in the low-response state, while false alarms in high-response states could be avoided by raising the threshold when the state changed. However, the cortical activity was classified to be in this particular state approximately 20% of the time, and thus had a relatively modest impact on the overall performance, averaged across all states. What is not currently known is the overall statistics associated with the state transitions (i.e. distribution of time spent in each state, rate of transitions, etc.) during engagement within perceptual tasks, but in any case, what we observe here is a normalization of detectability across brain states.
For near-threshold sensory perception, the signal detection theory framework asserts that single-trial responses are predictive of perceptual report [28]. While there are many previous studies that seem to support this [49–52], several animal studies have called this into question, showing that primary sensory neural activity does not necessarily co-vary with perceptual report on simple detection tasks [23,25,27]. It is possible that the conflicting findings in the literature are due to behavioral state effects, and that more consistent reports would emerge if the analysis of the neural activity incorporated elements of the state-classification approach developed here. Our results show how single-trial response size can be decoupled from perception, if a downstream network predicts and then accounts for the variability in sensory responses. Moreover, our analysis showed that some states of pre-stimulus activity should be associated with higher or lower performance on a near-threshold detection task, which has been observed in near-threshold detection studies in the rodent [26] and monkey [24]. It should be noted that there is controversy regarding the relevance of primary sensory cortex in simple behavioral tasks [53,54], but this is likely related to the task difficulty [55], where a large body of literature has resolutely shown that processing in primary cortical areas is critical for difficult tasks that increase cognitive load, and we suspect that near threshold stimuli such as those shown here fall in that category.
Many studies have demonstrated a link between pre-stimulus cortical activity and perceptual report on near-threshold detection tasks in humans [17,18,56–59]. Currently, it is not entirely clear how far the parallel in cortical dynamics between the mouse and human can be taken. One challenge is that connecting invasive recordings in the mouse to non-invasive recordings in human studies is non-trivial. Here, at the level of LFP, we observed similarities between species in the interaction between ongoing and evoked activity: the largest evoked responses tended to be preceded by positive deflection in the LFP, and the smallest evoked responses were preceded by negative deflection in the LFP. This relationship, the negative interaction phenomenon, points to a non-additive interaction between ongoing and evoked activity and is also observed in both invasive and non-invasive recordings in humans [33,56,60,61]. Establishing parallels between cortical dynamics on a well-defined task, such as sensory detection, between humans and animal models is an important direction for future studies.
In summary, we have developed a framework for the prediction of variable single-trial sensory-evoked responses and shown that this prediction, based on cortical state classification, can be used to enhance the readout of sensory inputs. Utilizing state-dependent decoders for brain-machine interfaces has been shown to greatly improve the readout of motor commands from cortical activity [62,63], at the very end-stage of cortical processing. Others have raised the possibility of using state knowledge to ‘cancel out’ variability in sensory brain-machine interfaces, with the idea that this could generate a more reliable and well-controlled cortical response [64,65], which would in theory transmit information more reliably. This is intriguing, though our analysis suggests a slightly different interpretation: if downstream circuits also have some knowledge of state, canceling out encoding variability may not be the appropriate goal. Instead, the challenge is to target the response regime for each state. This could be particularly relevant if structures controlling state, including thalamus [66], are upstream of the cortical area in which sensory BMI stimulation occurs. The simple extension of signal detection theory we explored suggests a solution to the problem that the brain faces at each stage of processing: how to adaptively read out a signal from a dynamical system constantly generating its own internal activity.
All procedures were approved by the Institutional Animal Care and Use Committee at the Georgia Institute of Technology (Protocol Number A16104) and were in agreement with guidelines established by the National Institutes of Health. Six nine to twenty-six week old male C57BL/6J mice were used in this study. Mice were maintained under 1–2% isoflurane anesthesia while being implanted with a custom-made head-holder and a recording chamber. The location of the barrel column targeted for recording was functionally identified through intrinsic signal optical imaging (ISOI) under 0.5–1% isoflurane anesthesia. Recordings were targeted to B1, B2, C1, C2, and D2 barrel columns. Mice were habituated to head fixation, paw restraint and whisker stimulation for 3–7 days before proceeding to electrophysiological recordings. Following termination of the recordings, animals were anesthetized (isoflurane, 4–5%, for induction, followed by a euthanasia cocktail injection) and perfused.
Local field potential was recorded using silicon probes (A1x32-5mm-25-177, NeuroNexus, USA) with 32 recording sites along a single shank covering 775 μm in depth. The probe was coated with DiI (1,1’-dioctadecyl-3,3,3′3’-tetramethylindocarbocyanine perchlorate, Invitrogen, USA) for post hoc identification of the recording site. The probe contacts were coated with a PEDOT polymer [67] to increase signal-to-noise ratio. Contact impedance measured between 0.3 MOhm and 0.7 MOhm. The probe was inserted with a 35° angle relative to the vertical, until a depth of about 1000 μm. Continuous signals were acquired using a Cerebus acquisition system (Blackrock Microsystems, USA). Signals were amplified, filtered between 0.3 Hz and 7.5 kHz and digitized at 30 kHz.
Mechanical stimulation was delivered to a single contralateral whisker corresponding to the barrel column identified through ISOI using a galvo motor (Cambridge Technologies, USA). The galvo motor was controlled with millisecond precision using a custom software written in Matlab (Mathworks, USA). The whisker stimulus followed a sawtooth waveform (16 ms duration) of various velocities (1000 deg/s, 500 deg/s, 250 deg/s, 100 deg/s) delivered in the caudo-rostral direction. To generate stimuli of different velocity, the amplitude of the stimulus was changed while its duration remained fixed. Whisker stimuli of different velocities were randomly presented in blocks of 21 stimuli, with a pseudo-random inter-stimulus interval of 2 to 3 seconds and an inter-block interval of a minimum of 20 seconds. The total number of whisker stimuli across all velocities presented during a recording session ranged from 196 to 616 stimuli.
For analysis, the LFP was down-sampled to 2 kHz. The LFP signal entering the processing pipeline is raw, with no filtering beyond the anti-aliasing filters used at acquisition, enabling future use of these methods for real-time control. Prior to the analysis, signal quality on each channel was verified. We analyzed the power spectrum of LFP recorded on each channel for line noise at 60 Hz. In some cases, line noise could be mitigated by fitting the phase and amplitude of a 60-Hz sinusoid, as well as harmonics up to 300 Hz, over a 500-ms period in the pre-stimulus epoch, then extrapolating the sinusoid over the stimulus window and subtracting. A small number of channels displayed slow, irregular drift (2 or 3 of 32 channels) and these were discarded. All other channels were used.
Current source density (CSD) analysis was used for two different purposes: first, to functionally determine layers based on the average stimulus-evoked response, and second, to analyze the pre-stimulus activity (in single trials) to localize sinks and sources generating the predictive signal. We describe the general method used here. Prior to computing the current source density (CSD), each channel was scaled by its standard deviation to normalize impedance variation between electrodes. We then implemented the kernel CSD method [37] to compute CSD on single trials. This method was chosen because it accommodates irregular spacings between electrodes, which occurs when recordings on a particular contact do not meet quality standards outlined above. To determine the best values for the kernel method parameters (regularization parameter, λ; source extent in x-y plane, r; and source extent in z-plane, R) we followed the suggestion of Potworowski (2012) and selected the parameter choices that minimize error in the reconstruction of LFP from the CSD. These parameters were similar across recordings, so for all recordings we used: λ = 0.0316; r = 200μm; R = 37.5μm.
The trial-averaged evoked response was computed on each trial by subtracting the pre-stimulus baseline (average over 200 ms prior to stimulus delivery) and computing the average across trials. The CSD of this response profile was computed as described above. The center of layer 4 was determined by finding the largest peak of the trial-averaged evoked LFP response as well as the location of the first, large sink in the trial-averaged sensory-evoked CSD response. We assume a width of 205 μm for layer 4, based on published values for mice [32].
The matched filter ideal observer analysis [38] is implemented as follows. The score s(t) is constructed by taking the dot product of the evoked responses yt with a filter matched to the average evoked response:
s(t)=yt∙ξ0
This is equivalent to computing the sum
s(t)=∑t'=1Nξ(x(t+t')-x(t))ξ(t')
In the standard encoding model, if η is zero-mean white noise, this gives a signal distribution
P(s)~N(∥ξ0∥,σ2)
where σ2=∥ξ0∥2ση2 and a noise distribution with mean 0. In practice, we do not parameterize the distribution, because η is not uncorrelated white noise, and work from the score distribution directly.
For the state-aware decoder, we use the prediction α^t,k of evoked responses
yt=ξ0+∑k=1NCα^t,kξk+η'
This changes the score to
s(t)=|ξ0|2+∑kα^t,kξk∙ξ0+η'∙ξ0
Typically, one of the first two PCs (ξ1 or ξ2) has a very similar shape to ξ0, while the other one has both positive and negative components (Fig 2, S1 and S2 Figs). For the state-aware threshold, we use state predictions for the component that is more similar to ξ0, as indicated in S1 and S2 Figs.
An event is detected at time t for threshold θ when s(t) > θ is a local maximum that is separated from the nearest peak by at least 15 ms and has a minimum prominence (i.e. drop in s before encountering another peak that was higher than the original peak) of |ξ0|2/2.
|
10.1371/journal.pgen.1008225 | Native American admixture recapitulates population-specific migration and settlement of the continental United States | European and African descendants settled the continental US during the 17th-19th centuries, coming into contact with established Native American populations. The resulting admixture among these groups yielded a significant reservoir of Native American ancestry in the modern US population. We analyzed the patterns of Native American admixture seen for the three largest genetic ancestry groups in the US population: African descendants, Western European descendants, and Spanish descendants. The three groups show distinct Native American ancestry profiles, which are indicative of their historical patterns of migration and settlement across the country. Native American ancestry in the modern African descendant population does not coincide with local geography, instead forming a single group with origins in the southeastern US, consistent with the Great Migration of the early 20th century. Western European descendants show Native American ancestry that tracks their geographic origins across the US, indicative of ongoing contact during westward expansion, and Native American ancestry can resolve Spanish descendant individuals into distinct local groups formed by more recent migration from Mexico and Puerto Rico. We found an anomalous pattern of Native American ancestry from the US southwest, which most likely corresponds to the Nuevomexicano descendants of early Spanish settlers to the region. We addressed a number of controversies surrounding this population, including the extent of Sephardic Jewish ancestry. Nuevomexicanos are less admixed than nearby Mexican-American individuals, with more European and less Native American and African ancestry, and while they do show demonstrable Sephardic Jewish ancestry, the fraction is no greater than seen for other New World Spanish descendant populations.
| The post-Colombian settling of North America brought African, European, and Native American populations into close proximity for the first time. The inevitable admixture among these groups resulted a reservoir of Native American ancestry in modern US populations, outside of traditional Native American groups. Here we characterize that Native American ancestry in a geographically diverse set of African descendant, Western European descendant, and Spanish descendant populations. We show that Native American ancestry in the US population is not monomorphic, strongly related to geography, and suggestive of frequent historical admixture between European settlers and local Native American groups. We also show the presence of a unique, admixed Spanish population in the Southwestern US, the modern Nuevomexicanos, that is distinct from other Spanish descendant groups.
| Native Americans inhabited the area that now makes up the continental US for thousands of years prior to the arrival of the first European settlers. The ancestors of modern Native Americans are thought have arrived in the Americas from Asia, by way of the Bering Strait, in several successive waves of migration [1]. The current model, based on archaeology and comparative genomic studies, holds that the earliest ancestors of Native Americans arrived in the Americas ~23,000 years ago [2]. The earliest evidence for Native Americans in the continental US dates to ~14,000 years ago [3]. The much later arrival of Europeans in the Americas, followed shortly thereafter by Africans who were brought by force via the trans-Atlantic slave trade, had a drastic effect on the demographic makeup of the region. Native American population numbers declined rapidly in the face of continuous immigration, settlement, and conflict, and as a result the modern US population is made up mainly of descendants of European and African immigrants.
Europeans arrived in the Americas more than 20,000 years after the first Native Americans. The first European settlers to reach the continental US were Spaniards led by the conquistador Ponce de León, who claimed Florida for the Spanish crown in 1513 [4]. British settlers arrived more than 70 years later, initially establishing the ill-fated colony of Roanoke in 1585 and later the permanent settlement of Jamestown in 1607 [5]. An estimated 400,000 British had migrated to the US by the end of the 17th century. The first Africans were brought to Jamestown in 1619 by Dutch pirates who traded them to the British settlers as indentured servants [6]. The social status of Africans in the US changed quickly, with slavery first legally sanctioned by 1640. The trans-Atlantic slave trade would eventually bring ~400,000 enslaved Africans to the continental US [7].
The arrival of Europeans and Africans in the Americas, and the conflict that followed, would prove to be catastrophic for the indigenous population. It has been estimated that 10–100 million Native Americans may have died in the first 150 years after Columbus’ arrival in the New World, amounting to a 95% reduction in the population [8]. This massive Native American population decline is mainly attributed to the introduction of European and African endemic infectious diseases–e.g. malaria, measles, and smallpox–for which the indigenous population had little or no immune defense [8, 9].
The story of conflict between Native Americans and European settlers and enslaved Africans, along with the devastating consequences for the indigenous population, is by now well-known. However, there is another, perhaps less appreciated, aspect of the encounter between these population groups that has also had profound consequences for the genetic demography of the Americas. Here, we are referring to the process of genetic admixture, whereby individuals from previously isolated population groups reproduce, resulting in the novel combination of ancestry-specific haplotypes within individual genomes. Admixture has been a fundamental feature of human evolution and migration [10]. Whenever previously isolated human populations meet, no matter what the circumstances, they mix and give rise to individuals with a mosaic of different genetic ancestries.
As European and African descendants settled the continental US, they inevitably came into contact with established Native American populations resulting in admixture and the introduction of Native American genomic sequence into the expanding US population. Accordingly, the genomes of European and African descendants in the US are expected to contain some fraction of Native American ancestry. In other words, a significant reservoir of Native American ancestry currently exists outside of recognized indigenous communities. In this study, we ask how the historical processes of migration and settlement affected the distribution of Native American admixture across the continental US (S1 Fig). We address this question for the three largest genetic ancestry groups in the modern US population: African descendants (AD), Western European descendants (WD), and Spanish descendants (SD).
The first aim of our study was to characterize the major genetic ancestry groups for the continental US based on observable patterns of ancestry and admixture seen for the 15,620 individuals from the Health and Retirement Study (HRS) analyzed here. The Health and Retirement Study data is sponsored by the National Institute on Aging (grant number U01AG009740) and is conducted by the University of Michigan. Having defined the US genetic ancestry groups, we then considered the distribution of Native American admixture within and between ancestry groups and among geographic regions. We provide a detailed description (S1 Text), along with supporting results (S3–S8 Figs, S2 and S3 Tables), of how we defined the three main US ancestry groups–African descendants, Western European descendants, and Spanish descendants–in the Supplementary Material.
The ancestry distribution of HRS individuals among the three largest US genetic ancestry groups is shown in Fig 1. Visual inspection of the continental ancestry fractions seen for members of the three groups supports our approach to genetic ancestry-based classification (Fig 1A). For example, the majority of Spanish descendant individuals show substantially higher Native American ancestry compared to Western European descendants (Fig 1A); the median Native American ancestry for the Spanish descendant group is 38% compared to 0.1% for the Western European descendant group (Fig 1B). In addition, individuals from the Spanish descendant group cluster tightly with the Mexican reference population from the 1KGP, along the second axis between the European and Native American populations in the principal components analysis (PCA) plot of the pairwise genome distances (Fig 1C). It is important to note that we did not use Native American ancestry for the purposes of classification. Rather, European ancestry alone was sufficient to recapitulate known levels of Native American ancestry for Spanish descendants.
Individuals from the African descendant group show medians of 85% African ancestry, 14% European ancestry, and 1% Native American ancestry (Fig 1B). Most of these individuals group along the first PCA axis separating the African and European reference populations. In contrast to the admixed Spanish and African descendant groups, Western European descendants show extremely low levels of admixture with non-European populations, with a median value of 99.8% European ancestry. Given their relatively low numbers (S2 Fig), as well as their relatively late historical arrival in the continental US, we did not consider Asian descendants further in this study.
Individuals assigned to the three main genetic ancestry groups show distinct geographic distributions across the continental US, which are largely consistent with demographic data for the country. The proportion of African descendants is highest in the three southern census regions, Western European descendants in the two north central regions, and Spanish descendants in the Mountain census region, which includes Arizona and New Mexico (Fig 1D).
We compared the patterns and extent of sex-biased admixture among the three US genetic ancestry groups by comparing the continental ancestry fractions–African, European, and Native American–seen for the X chromosomes versus the autosomes. For any given ancestry component, a relative excess of X chromosome ancestry is indicative of female-biased admixture, whereas an excess of autosomal ancestry reflects male-biased admixture [11]. This was only done for admixed individuals that had two or more continental ancestry fractions at >1.5% of the overall ancestry. Almost all individuals from the African and Spanish descendant groups met this criterion, but only a small minority of Western European descendant individuals with Native American admixture did. African and Spanish descendant groups showed marked patterns of sex-biased admixture, whereas the Western European descendants did not show any appreciable evidence of sex-biased admixture (Fig 2). The strongest pattern of sex-biased admixture was seen for Spanish descendants, with female-biased Native American admixture and male-biased European admixture. African descendants show female-biased African ancestry and male-biased European ancestry.
For each US genetic ancestry group, we considered three distinct characteristics of Native American ancestry across the continental US: (1) the relative levels of Native American ancestry genome-wide, (2) the patterns of Native American allele frequencies, and (3) the phylogenetic relationships among US populations based on their Native American ancestry.
As we showed previously, overall Native American ancestry is highest for the Spanish descendant group (median 38%, SD = 20.1), followed by the African descendant (1%, SD = 4.4) and Western European descendant groups (0.1%, SD = 2.7) (Fig 1B). Among all three ancestry groups, the highest levels of Native American ancestry are seen for the West-South-Central (WSC; including Texas), Pacific (PAC; including California), and Mountain (MNT; including Arizona and New Mexico) census regions (Fig 3). Native American ancestry levels show the highest variability among regions for the Spanish descendant group (coefficient of variation [c.v.] = 1.08), followed by the Western European descendant (c.v. = 0.65) and African descendant (c.v. = 0.60) groups.
We characterized the ancestry-specific and genome-wide haplotype heterozygosity (HH) for each of the admixed populations to interrogate how admixture has affected the diversity of the populations (S9 Fig, S4 Table). Where present, the African-specific HH was the highest for each population and the Native American HH was the lowest, consistent with previous observations of present day populations [12]. The genome-wide HH was significantly higher than any ancestry-specific HH for the African descendant populations, consistent with the introduction of novel haplotypes into the already diverse African background. Spanish descendant genome-wide HH was significantly higher than both the European and Native American-specific HH, but lower than African, which contributes only a small fraction of the total ancestry in the present-day Spanish descendant populations. The Western European descendant populations show a relatively very small amount of Native American ancestry; accordingly, the genome-wide HH shows no significant difference from the European HH, but is nevertheless higher than the Native American HH.
We measured the patterns of Native American allele frequencies across the continental US using ADMIXTURE analysis of Native American haplotypes for individuals from the three ancestry groups. Visualization of the ancestry vectors produced by ADMIXTURE shows that the African and Western European descendant groups have patterns that are similar to each other (Fig 4A, top panel; S10 Fig) and distinct from the patterns seen for the Spanish descendant group (Fig 4B, top panel; S11 Fig). Comparing the ADMIXTURE vectors of these two population groups to those of the Spanish descendant populations shows that African descendant and Western European descendant populations are significantly closer to each other than either is to the Spanish descendant populations (S1 Text, S12 and S13 Figs). Furthermore, the African descendant and Western European descendant groups show ancestry patterns that are intermediate to the Canadian and Northern Mexican Native American reference populations, whereas the Spanish descendant group shows Native American ancestry patterns that are more similar to the Mexican reference population, Mexican Native American populations, or the admixed Puerto Rican population. This is consistent with the fact that we use Native American reference populations from outside the US to identify Native American haplotypes in US population groups. There is substantial regional variation in Native American ancestry seen in the Spanish descendant group, with characteristically Mexican patterns seen in the Pacific (PAC) and West South-Central (WSC) regions and a strongly Puerto Rican pattern in the Mid-Atlantic (MA) region. At K = 9, ADMIXTURE is able to resolve the two Northern Mexican Native American reference populations, as well as reveal a unique ancestry in the Spanish descendant population from the Mountain (MNT) region. This population shows a distinct pattern of Native American ancestry under any K (S11 Fig), which we explore in more detail in the following section.
The phylogenetic relationships of the Native American ancestry in modern US populations were inferred by calculating the fixation index (FST) between pairs of populations based on their masked Native American haplotypes (Fig 5). The Canadian and Amazonian Native American reference populations occupy the most distant clades on the phylogeny with the admixed Mexican and Mexican Native American reference populations adjacent to the Amazonian group. African descendant populations from all of the census regions form a single clade, along with Western European descendants from the Southeast region (SE, WD). Western European descendant populations from the West North-Central (WNC, WD) and East North-Central (ENC, WD) regions group most closely with the Canadian Native American reference populations. Western European descendant populations from the Western US (West South-Central (WSC, WD), Pacific (PAC, WD), and Mountain (MNT, WD) regions) are intermediate between the African descendant clade and the Spanish descendant of populations. Spanish descendant populations from most of the US census regions group closely with Mexican populations, with the exception of the Mid-Atlantic region (MA) which groups most closely with the Puerto Rican and Amazonian reference populations. To quantify the affinities of the Native American ancestry in admixed US populations we computed outgroup f3-statistics of the form f3(African; admixed, reference) and D-statistics of the form D(African, admixed; Native American, reference) using the masked Native American haplotypes and AdmixTools [13]. The f3 and D-statistics agree well with the inferred phylogeny (S14 & S15 Figs). Western European descendants from WNC and ENC regions showed the highest affinity for Canadian Native American populations. Consistent with the single clade observed in the phylogeny, African descendant populations showed generally lower affinity to the reference populations. Spanish descendant populations showed the highest affinities for Mexican reference populations, apart from the MA population which showed higher affinity for Amazonian groups.
The ADMIXTURE results for the Spanish descendant group in the Mountain region (MNT) point to the presence of two distinct sub-populations, one of which is clearly of Mexican descent, whereas the second group has a pattern distinct from any other group analyzed here (Figs 4B and 6A). If these two apparent Spanish descendant Mountain sub-populations are considered separately, they form distinct phylogenetic groups (Fig 6B). One group clearly falls into the clade with the other Mexican origin populations (MNT, Mexican), whereas the distinct group is basal to the Mexican clade and intermediate between the Western US and Mexican clades (MNT, Nuevomexicano). The results of the ADMIXTURE and phylogenetic analyses are consistent with historical records indicating the presence of a unique group of Spanish descendants in the American Southwest, known as the ‘Hispanos of New Mexico’ or Nuevomexicanos. This population is descended from very early Spanish settlers to the Four Corners region of the US, primarily New Mexico and southern Colorado, and distinct from Mexican-American immigrants who arrived later [14].
Members of the Nuevomexicano population have maintained a distinct cultural identity for centuries, and the ability to isolate individuals from this group based on analysis of their genotypes allowed us to address open questions related to their ancestry. In addition to characterizing their distinct pattern of Native American ancestry, we also compared the levels of Native American admixture between Nuevomexicanos and the other nearby Spanish descendant groups, which show a Mexican pattern of Native American ancestry. Consistent with previous results [15], we show that Nuevomexicanos have significantly more European ancestry and less Native American ancestry than other Spanish descendant groups from the Western Census regions (Fig 6C). Nuevomexicanos also show significantly lower levels of African ancestry compared to the other Spanish descendant groups.
Nuevomexicano cultural and historical traditions suggest that many of the early Spanish settlers in the region were Conversos, Jewish individuals who ostensibly converted to Catholicism in an effort to avoid religious persecution and pogroms, while secretly maintaining Jewish identity and traditions [16]. We interrogated this idea by comparing the extent of Sephardic Jewish admixture found among individuals with the Nuevomexicano ancestry pattern compared to other Spanish descendant populations. Sephardic Jewish admixture was measured by comparing European haplotypes from Spanish descendant individuals to a reference panel including both European and Sephardic Jewish populations. Nuevomexicanos show elevated levels of matching to Jewish haplotypes compared to Spanish and other European populations, consistent with substantial Converso ancestry among New World Spanish descendant populations [17] (Fig 6d). However, Nuevomexicanos do not show a higher level of Converso ancestry compared to the other New World Spanish descendant populations.
We were able to delineate three predominant genetic US ancestry groups–African descendant, Western European descendant, and Spanish descendant–using comparative analysis of whole genome genotypes from >15,000 individuals from across the continental US. Each of these different groups of people experienced distinct historical trajectories in the US, which we found to be manifested as group-specific patterns of Native American ancestry.
Individuals from the African descendant group show low (Fig 1B) and relatively invariant (Fig 3A) levels of Native American ancestry across the continental US. The patterns of Native American ancestry seen for the African descendant group are also more constant among US census regions compared to individuals from the other two ancestry groups (Fig 4A). With respect to the Native American component of their ancestry, African descendant populations from all US census groups form a single clade, along with the Southeast Western European descendant population (SE, WD) (Fig 5). Considered together, these results point to a most likely scenario whereby African descendants admixed with local Native American groups in the antebellum South. Early admixture with Native Americans in the South was followed by subsequent dispersal across the US during the Great Migration in the early to mid-twentieth century [18]. The genetic legacy of the Great Migration has previously been explored based on overall patterns of African American genetic diversity [19]. Here, we were able to uncover traces of this same history based solely on the relatively low Native American ancestry component found in the genomes of African descendants.
Of the three US ancestry groups characterized here, the Western European descendant group shows the lowest levels of Native American ancestry (Fig 1B), consistent with a large and fairly constant influx of European immigrants to the US along with social and legal prohibitions against miscegenation [20]. Compared to African descendants, individuals from the Western European descendant group show more variant levels of Native American ancestry among US census regions (Fig 3B) along with substantially more region-specific patterns of Native American ancestry (Fig 4A). Their region-specific patterns of Native American ancestry are also reflected in the Native American ancestry-based phylogeny, whereby the Western European descendant populations are related according to their geographic origin across the country (Fig 5). These results point to a historical pattern of continuous, albeit infrequent, admixture between local Native American groups and European settlers as they moved westward across the continental US.
As can be expected, the Spanish descendant group shows by far the highest (Fig 1B) and most variable (Fig 3C) levels of Native American ancestry across the US. Individuals from this group show highly regional-specific patterns of Native American ancestry (Fig 4B), consistent with known demographic trends. For example, analysis of the Native American component of Spanish descendant ancestry is sufficient to distinguish Puerto Rican immigrants from the Mid-Atlantic census region from Mexican Americans who predominate in the western census regions. Perhaps most striking, the patterns of Native American ancestry seen for the Mountain census regions were alone sufficient to distinguish descendants of very early Spanish settlers to the region, the group known as Hispanos or Nuevomexicanos, from subsequent waves of Spanish descendants who arrived later from Mexico.
The three main US ancestry groups are also distinguished by their patterns of sex-biased ancestry in a way that reflects the unique history of each group (Fig 2). Western European descendants show very little evidence for sex-biased ancestry, along with very low levels of overall admixture, compared to the African and Spanish descendant groups. Sex-bias for Spanish descendants is characterized by a strong female-bias for Native American ancestry coupled with European male-biased ancestry. The pattern that we observe here is similar to what has been reported in a number of previous studies and is consistent with the history of male-biased migration to the region dating back to the era of the conquistadors [21, 22]. The African descendant group shows female-biased African ancestry and male-biased European ancestry, a pattern which has also been documented previously and tied to the legacy of slavery and racial oppression in the US [23, 24]. It has not been previously possible to directly compare the extent of sex-biased admixture among the three largest ancestry groups in the US as we have done here. As such, it is interesting to note that the history of the Spanish colonization in Latin America had a stronger impact on sex-biased ancestry than the legacy of slavery in the US.
Our ability to distinguish Nuevomexicanos from the HRS dataset, using their distinct Native American ancestry, allowed us to address a number of open questions and controversies regarding the history and culture of this interesting population. Nuevomexicanos from the American southwest are historically defined as the descendants of early Spanish settlers, those who arrived in the period from 1598 to 1848, as opposed to immigrants from Mexico who arrived the region considerably later. The two distinct patterns of Native American ancestry seen for Spanish descendant individuals from the Mountain census region are very much consistent with this historical definition. The Nuevomexicanos show a pattern of Native American ancestry that is intermediate to the Canadian and Mesoamerican reference populations analyzed here, whereas the Mexican American individuals from the same region are more closely related to Mesoamerican reference populations. This is consistent with early admixture with local Native American groups in the US southwest, for the Nuevomexicanos, versus admixture with Mesoamerican groups in Mexico for the later Mexican immigrants. A more precise characterization of Nuevomexicanos’ Native American ancestry would require access to genomic data from US Native American reference populations, which are not readily available owing to cultural resistance to genetic testing for ancestry among these groups [25].
Historically, Nuevomexicanos have identified strongly with their European (Spanish) ancestry, while downplaying ancestral ties to Native Americans [26]. This tradition of exclusive European identity is rooted in the colonial era when Spanish descendants in the region were preoccupied with the notion of maintaining so-called pure blood, and the local aristocracy identified as Castilian. The Spanish preoccupation with admixture in the Americas was codified into the so-called Sistema de Castas, whereby mixed-race individuals were categorized into a complex hierarchical system, with tangible legal and social implications, based on their parents’ ancestry [27]. Mexicans, on the other hand, have long identified as Mestizo with an explicit recognition of their Native American heritage [28]. Our comparative analysis of genetic ancestry for Nuevomexicanos and Mexican ancestry groups yielded results that are partly consistent with this historical narrative. On the one hand, Nuevomexicanos do have a substantial amount of Native American ancestry, with a median of just under 40% (Fig 6C), which is far more than seen for the African descendant and Wester European descendant groups analyzed here. The fraction of Native American ancestry seen in the Nuevomexicanos is also higher than in several populations in South America (Medellín [29] and Chocó [30], Colombia) and the Caribbean (Cuba, the Dominican Republic, and Puerto Rico [29]). Nevertheless, the Nuevomexicanos have significantly less Native American ancestry, and more European ancestry, than nearby Mexican descendant populations (Fig 6C). Our results are consistent with a recent study that used microsatellite-based ancestry analysis on a much smaller sample of self-identified Nuevomexicanos, who were also found to have higher European ancestry and lower Native American ancestry compared to Mexican Americans [15]. Interestingly, we found that the Nuevomexicanos also have significantly less African ancestry than Mexican descendant populations, which likely reflects higher levels of early African admixture in Mexico [31].
We investigated this apparent differing population history by inferring the timings and proportions of admixture with the TRACTS utility [32]. The best models from the TRACTS analysis indicated a European and Native American admixture 10–11 generations ago, followed shortly by a small African admixture (S16 Fig). All models for the Mexican populations converged on an admixture time of 10–11 generations. The best Nuevomexicano model suggests a slightly older admixture, though with the same ordering, 11–12 generations ago, while the best Nuevomexicano model for 10–11 generations produced a significantly worse model (log-likelihood of -413 vs. -390). Regardless, this suggests that the timing of admixture in the Mexican populations and the Nuevomexicano population was similar, consistent with historical records, while the Native American source populations were different, consistent with their geographical origins. While the admixture timing estimates for these groups are within the range of previous estimates, they are younger than what has been previously reported for Mexican populations [33]. Nevertheless, as can be expected for a Caribbean population, the Puerto Rican descendant MA population showed a much older admixture, ~15 generations ago, very similar to the 1KGP Puerto Rican population (S17 Fig).
Perhaps the most controversial aspect of Nuevomexicano history relates to the influence of Conversos on the culture and traditions of the local community. Conversos are Jewish people who converted to Catholicism under intense pressure from religious persecution in Spain, and elsewhere in Europe, and many Spanish Conversos immigrated to the New World [34]. Despite their forced conversion to Catholicism, some New World Conversos apparently maintained Jewish religious traditions over the centuries since their immigration from Spain. For example, the persistence of rituals and symbols related to Jewish traditions in New Mexico has been taken as evidence for an influential presence of Conversos among the Nuevomexicanos, a position championed by the historian Stanley Hordes[16]. On the other hand, the folklorist Judith Neulander and others have been fiercely critical of this narrative based on what they perceive to be misunderstandings of the origins of many of the cultural traditions tied to Jewish rituals and even deliberate misrepresentations of facts [35]. Neulander’s interpretation relates the notion of Converso identity among Nuevomexicanos back to the colonial assertions of pure Spanish ancestry given that the Sephardim are Spanish and would presumably be loath to marry outside of their religious group [36].
We evaluated the extent of Sephardic Jewish ancestry among Nuevomexicanos, via comparative analysis of their European haplotypes to both European and Sephardic Jewish reference populations, in attempt to assess the genetic evidence in support of the Converso narrative. While we did find more Sephardic Jewish ancestry among Nuevomexicanos compared to Spaniards or other Europeans, they did not show any more Sephardic Jewish ancestry than Mexican descendants from nearby regions (Fig 6D). Our results are consistent with a recent study that used haplotype-based ancestry methods to uncover widespread Converso ancestry in Latin American populations [17]. Taken together, we interpret these results to indicate that, while Nuevomexicanos do in fact have a demonstrable amount of Jewish ancestry, they show no more, or less, Jewish ancestry than other New World Latin American populations. Of course, we cannot weigh in on the strength of evidence for or against the persistence of Jewish cultural traditions among Nuevomexicanos based on our genetic evidence alone. Nevertheless, there does not seem to be anything particularly unusual, at least from the genetic perspective, with respect to the extent of Sephardic Jewish heritage among Nuevomexicanos.
Much of the genetic legacy of the original inhabitants of the area that is now the continental US can be found in the genomes of the descendants of European and African immigrants to the region. In this study, we analyzed signals of Native American genetic ancestry in a comparative analysis of genomes from the three largest US ancestry groups: African descendants, Western European descendants, and Spanish descendants. Our study was enabled by the use of haplotype-based methods for genetic ancestry inference and leveraged a large dataset of whole genome genotypes. This approach allowed for detailed analysis of Native American ancestry patterns even when the per-genome levels of Native American ancestry were quite low. Each of the three genetic ancestry groups analyzed here shows distinct profiles of Native American ancestry, which reflect population-specific historical patterns of migration and settlement across the US. Analysis of the Native American ancestry component for members of these groups allowed for the delineation of region-specific subpopulations, such as the Nuevomexicanos from the American southwest, and facilitated the interrogation of specific historical scenarios.
This study was approved by the Georgia Institute of Technology Central Institutional Review Board, #H17029. Data were provided by third party sources and no additional ethical approval was required.
Whole genome genotype data of US individuals from the Health and Retirement Study (HRS) dataset (n = 15,620) were merged with whole genome sequence variant data from the 1000 Genomes Project (1KGP) [37, 38] (n = 1,718) and whole genome genotype data from the Human Genome Diversity Project (HGDP) [12, 39, 40] (n = 230) (S1 Table). Individual HRS genotypes are provided along with geographical origin data for sample donors from the nine census regions in the continental US. A collection of Native American genotypes from 21 populations across the Americas was taken from a comprehensive study on Native American population history [2] (n = 314). These Native American genotype data were accessed according to the terms of a data use agreement from the Universidad de Antioquia. Whole genome genotype data from 5 populations of Sephardic Jewish individuals (n = 40) were also included as reference populations [41]. The genotypes from HRS individuals were merged with the comparative genomic data sources using PLINK version 1.9 [42], keeping only those sites common to all datasets and correcting SNP strand orientations for consistency as needed. The final merged dataset includes 228,190 SNPs across 17,882 individuals. Pairwise distances between individuals was calculated using the–dist option of PLINK [42], and principal component analysis carried out using the prcomp function of R [43]. The merged genotype dataset was phased using ShapeIT version 2.r837 [44]. SNPs that interfered with the ShapeIT phasing process were excluded from subsequent analyses. ShapeIT was run without reference haplotypes, and all individuals were phased at the same time. Individual chromosomes were phased separately, and the X chromosome was phased with the additional ‘-X’ flag.
The RFMix algorithm [45] is able to accurately characterize the local ancestry of admixed individuals but is prohibitively slow when run on a dataset of the size used here. To reduce the runtime, we modified RFMix version 1.5.4 so that the expectation-maximization (EM) procedure samples from, and creates a forest for, the entire set of individuals rather than each individual. This modified RFMix was run in the PopPhased mode with a minimum node size of five, using 12 generations and the “—use-reference-panels-in-EM” for two rounds of EM, generating local ancestry inference for both the reference and admixed populations. Continental African, European, and Native American populations were used as reference populations. Contiguous regions of ancestral assignment, “ancestry tracts,” were created where RFMix ancestral certainty was at least 95%. Genome-wide ancestry estimates from the modified RFMix algorithm closely correlate with those from ADMIXTURE (S18 Fig). The present Native American reference populations may not be close to the actual ancestral Native American populations for all of the HRS regions. To evaluate how a distant reference population would affect the LAI, we carried out the RFMix procedure a second time, but using only East Asian populations as the reference for Native American ancestry. The local ancestry inferred in this was very similar to that inferred when using actual Native American populations as references (S19 Fig), indicating that the choice of reference population does not greatly affect the LAI.
For each admixed population, rephased genotypes from the final output of RFMix were used to compute the haplotype heterozygosity (HH) for both the masked ancestry-specific genomes and for the unmasked whole-genome. Haplotypes were found by considering sets of 5–15 consecutive variants with a maximum recombination rate between any two variants of 0.5 cM/mB as in [46], resulting in 11,816 haplotypes. Significance in HH between ancestry-specific genomes was assessed using a Wilcoxon rank-sum test.
The extent of Sephardic Jewish (Converso) ancestry in individuals from the Spanish descendant group in HRS (as defined in the genome-wide ancestry section below), and Latin American populations from 1KGP, was inferred via ancestry-specific haplotype comparisons with Sephardic Jewish reference populations using the program ChromoPainter2 [10] (kindly provided by Garrett Hellenthal). First, African and Native American haplotypes were masked from the RFMix output. Then, the remaining European haplotypes were compared against genomes from the European reference populations together with the Sephardic Jewish populations. The extent of Jewish ancestry for any individual genome is defined as the ‘copying fraction’ from the Sephardic Jewish populations, where the copying fraction is taken as the fraction of sites with best matches to the Sephardic Jewish reference genomes. It should be noted that this procedure results in a relative fraction of Sephardic Jewish ancestry for all individuals under consideration, which is directly comparable among individuals but likely to be an overestimate of the total ancestry derived from a single source population.
ADMIXTURE [47] version 1.3.0 was used with K = 4 to infer continental ancestry fractions for individuals in the dataset via comparison with reference populations from Africa, Europe, the Americas, and East Asia. Sub-continental ancestry was inferred independently for each of the three major continental ancestry components–African, European, and Native American–using an ancestry-specific masking procedure that we developed as previously described [30]. This procedure relies on the local continental ancestry assignments, along with the re-phased genotypes, generated by RFMix as described above. Sub-continental ancestry was characterized by first masking out two of the three continental ancestries (African, European, and/or Native American) at a time and then analyzing the genomic regions (haplotypes) corresponding to the remaining continental ancestry. For sub-continental ancestry analysis of any given continental ancestry component, only those individuals with at least 1.5% genome-wide ancestry for that same continental group were used. This 1.5% threshold was chosen empirically based on observed ancestry assignments in the reference populations. As this work was focused on Native American ancestry, we chose a threshold higher than the Native American ancestry inferred in any of the European or African reference populations (max = 1.4% in a Spanish individual). While lowering this threshold would likely include a number of additional individuals with genuine Native American ancestry, we chose this stricter cutoff to avoid any possible ambiguity.
We developed a novel machine learning based approach to distinguish Spanish from other (primarily Western) European descendants in the HRS dataset via analysis of European-specific haplotypes. First, ADMIXTURE was run with K = 5 on the RFMix characterized European haplotypes for the HRS individuals to stratify sub-continental European ancestries based on comparison with Northern (Finnish and Russian), Western (French and British), Spanish, and Southern (Italian and Sardinian) European reference populations from the 1KGP and HGDP datasets. The ADMIXTURE results at K = 5 were used as one of the ADMIXTURE components was substantially different between the Spanish and Italian reference populations (Fig 4, S3 and S4 Figs). A Support Vector Machine (SVM) classifier [48] was then trained using the resulting ADMIXTURE ancestry vectors for the European reference populations from the four sub-continental groups: Northern, Western, Spanish, and Southern. The European-specific ADMIXTURE ancestry vectors for the HRS individuals were then classified into one of the four European sub-continental groups defined by the SVM classifier. A confidence threshold of 0.8 was used for sub-continental group assignments in order to minimize the number of misclassified individuals; while a lower threshold would allow for additional individuals to be included, a threshold below 0.7 lead to a higher missassignment rate while validating the classifier. For the purpose of analysis here, we consider two major groups of European descendants in the HRS data set: Spanish descendants (SD) and all others. Non-Spanish HRS individuals with <5% African ancestry are defined as Western European descendant (WD), whereas non-Spanish HRS individuals with at least 20% African ancestry were defined as African descendant (AD). It should be noted that this approach to defining genetic ancestry groups, as opposed to relying on self-identified race/ethnicity groups, is likely to yield ancestry classifications that correspond very well to self-identified race/ethnicity labels for the vast majority of individuals analyzed here. But our African descendant group will not include a small fraction of self-identified African Americans with little or no African ancestry. For example, there are 11 individuals in HRS who self-identify as African American but have no discernable African ancestry. We chose to rely on genetic ancestry, as opposed to self-identified race/ethnicity, in such cases in an effort to be as consistent as possible when delineating the three broad ancestry groups. We discuss this issue at more length in the Supplementary material (S1 Text).
Sex-biased ancestry contributions were inferred by comparing the RFMix characterized fractions of each continental ancestry component on the X chromosomes versus the autosomes as previously described [22, 46]. For each individual genome, and each ancestry component, the normalized difference between the X chromosome ancestry fraction and the autosomal ancestry fraction (ΔAdmix) is defined as:
ΔAdmix=Fanc,total×(Fanc,X-Fanc,auto)/(Fanc,X+Fanc,auto)
where Fanc,total, Fanc,X, and Fanc,auto are the genome-wide, X chromosome, and autosome ancestry fractions, respectively.
We used the RFMix defined Native American haplotypes for individuals from the HRS and reference populations to infer the phylogenetic relationships between populations. Using the masked Native American haplotypes, the FST was found between each population using smartpca from the EIGENSOFT package [49]. The resulting FST distance matrix was used to create a neighbor-joining tree [50] with the program MEGA6 [51]. Clade bootstrap values were calculated by resampling sites from the data, recalculating FST, and counting the occurrences of each clade using prop.part and part.clades of the Ape package [52].
The TRACTS method was used to infer the timing of admixture events with ancestry tracts defined by RFMix [32]. For the admixed Nuevomexicano, Mexican (1KGP), MA Spanish descendant, and Puerto Rican (1KGP) populations, three possible orderings of admixture were evaluated with TRACTS: (1) European, Native American, and African; (2) European, African, and Native American; and (3) African, Native American, and European. For each ordering, TRACTS was used to evaluate possible admixture timing from 14 to six generations ago, in 1000 bootstrap attempts. From the bootstrap attempts, the most likely series of admixture events was chosen to represent the population.
|
10.1371/journal.pcbi.1005360 | Effects of FGFR2 kinase activation loop dynamics on catalytic activity | The structural mechanisms by which receptor tyrosine kinases (RTKs) regulate catalytic activity are diverse and often based on subtle changes in conformational dynamics. The regulatory mechanism of one such RTK, fibroblast growth factor receptor 2 (FGFR2) kinase, is still unknown, as the numerous crystal structures of the unphosphorylated and phosphorylated forms of the kinase domains show no apparent structural change that could explain how phosphorylation could enable catalytic activity. In this study, we use several enhanced sampling molecular dynamics (MD) methods to elucidate the structural changes to the kinase’s activation loop that occur upon phosphorylation. We show that phosphorylation favors inward motion of Arg664, while simultaneously favoring outward motion of Leu665 and Pro666. The latter structural change enables the substrate to bind leading to its resultant phosphorylation. Inward motion of Arg664 allows it to interact with the γ-phosphate of ATP as well as the substrate tyrosine. We show that this stabilizes the tyrosine and primes it for the catalytic phosphotransfer, and it may lower the activation barrier of the phosphotransfer reaction. Our work demonstrates the value of including dynamic information gleaned from computer simulation in deciphering RTK regulatory function.
| Receptor tyrosine kinases are proteins integral to relaying signals from outside the cell to activators inside the cell that stimulate cell growth and development. Therefore, when these proteins show intrinsic activity independent of extracellular signaling, they can frequently cause developmental abnormalities, if the unchecked activity occurs before birth, or cancer, if the unchecked activity occurs later in life. Understanding what causes these proteins to become active upon receiving an extracellular signal will be helpful in pinpointing how they can exhibit activity without the extracellular signal. To study this phenomenon, we examined one receptor tyrosine kinase, FGFR2 kinase, and used computer simulation to identify what conformational changes occur in this protein upon activation. We then identified the function of these conformational changes in enabling the enzyme’s catalytic reaction to occur. Our results demonstrate the value of incorporating simulation data in analyzing the mechanisms of receptor tyrosine kinase activation, and suggest important features of this enzyme that should be considered in future drug development.
| Receptor tyrosine kinases (RTKs) occupy a central role in cellular regulation, acting as intermediaries in relaying signals from extracellular ligands to major signaling pathways in the cell [1–3]. Although the structural elements of RTKs are well-conserved [4], their functions are widely divergent. This is due to the subtle differences in the sequences and dynamic properties of structural elements underlying kinase activity [5]. The similarities between the various RTKs combined with their divergent behaviors presents a unique challenge in designing drugs to target specific RTKs whose constitutive activity has pathologic consequences, without generating off-target effects caused by reduced activity of other kinases [6, 7]. This endeavor has had profound successes [8] but still requires additional effort, particularly with regard to filling the gaps in our structural knowledge of these proteins.
RTKs, like all kinases, have an N-lobe and C-lobe, with the active site generally in the pocket buried between them [4, 9]. In order to avoid pathologic constitutive activity, RTKs have several autoinhibitory mechanisms in place that prevent the substrate from accessing the active site or prevent the phosphotransfer from taking place [10–13]. Some of these regulatory mechanisms involve the extracellular, transmembrane or juxtamembrane domains of the kinase preventing association of two kinase domains and their resultant autophosphorylation. Other mechanisms are contained within the kinase domain itself and involve regulatory regions whose dynamics may either favor or disfavor catalytic activity. One regulatory region is the nucleotide-binding loop, often referred to as the P-loop, at the tip of the N-lobe near the active site, that binds the ATP molecule that donates a phosphate group to the substrate [4, 9]. A second regulatory region is the αC helix that makes contact with the activation loop and often undergoes large movements to form the catalytically active state of the kinase. A third regulatory region, which is usually post-translationally modified to alter its regulatory behavior, is the activation loop. The activation loop usually contains one or multiple tyrosine residues that are available to be phosphorylated by other enzymes or, in many cases, autophosphorylated. This phosphorylation leads to altered dynamics of the activation loop residues resulting in greater catalytic activity of the kinase [14–17].
The fibroblast growth factor receptors (FGFRs) are a superfamily of RTKs that activate the MAP kinase and PI3 kinase pathways [18, 19]. Binding of an activator of the fibroblast growth factor family in concert with heparan sulfate stabilizes the dimerization of two receptors’ extracellular domains, leading in turn to the apposition of the receptors’ intracellular kinase domains. As in other RTKs, the kinase domain contains an activation loop with two adjacent tyrosine residues. Apposition of the kinase domains enables the activation loops to undergo trans-autophosphorylation, rendering the kinases catalytically active and able to perform phosphorylation of tyrosine residues in FGFR kinase substrates including PLC-γ [20–22] and additional sites on FGFR kinases [23–25]. In this work, we focus on the FGFR2 kinase, for which a wealth of experimental structural information is available, including crystal structures of the wild type kinase [25, 26], of mutant kinases [15, 26, 27], and NMR chemical shift data [15]. Previous work suggests that the FGFR2 kinase activation loop toggles between two states, inactive and active, and that mutation of activation loop residues can perturb the balance between these two states to increase the time that the kinase is in the activated state, even without phosphorylation [15]. Crystal structures illustrate several structural changes that occur when FGFR2 kinase is activated. These include rearrangement of the activation loop, a small rotation of the N-lobe toward the C-lobe, and dissolution of a network of hydrogen bonds between side chains in a triad of residues known as the “molecular brake” [26, 28]. Genomic point mutations in the activation loop, the αC helix, or the molecular brake in utero frequently lead to developmental disorders [26, 29–31], while somatic mutations may lead to cancer [29, 30, 32]. Surprisingly, in contrast to most RTKs, there is little apparent motion of the αC helix in the FGFR kinases upon activation, with crystal structures showing that the helix moves together with the rest of the N-lobe. This suggests that the bulk of structural change in the activated kinase is concentrated in the activation loop structure. Thus it is especially crucial to investigate the details of activation loop rearrangement in order to understand FGFR2 kinase function.
Despite the many crystal structures of FGFR2 [33], a mechanism to explain how phosphorylation of the activation loop residues leads to catalytic activity has not yet surfaced. In FGFR1, the activation loop residues Arg661 and Pro663 block the active site in the inactive structure, but in the active structure change conformation to allow a substrate to bind [11]. This led to the hypothesis that phosphorylation of the activation loop alters its structure to move these two residues away from the active site, allowing substrate phosphorylation. However, neither the inactive nor the active crystal structure of FGFR2 (PDBs 2PSQ and 2PVF, respectively [26]) shows any activation loop residues in the active site. In this study, we use molecular dynamics (MD) simulation to probe the dynamics of FGFR2 kinase and propose a mechanism to explain the regulatory role of activation loop phosphorylation.
In order to visualize the process by which the inactive structure of the activation loop undergoes conformational transition(s) into the active structure, we used the string method in collective variables [34]. The string method finds the minimum free energy path (MFEP) connecting two states at the end points, in this case the inactive and active structures of the kinase. The MFEP is the most likely pathway that the system will use to transition from the inactive structure to the active structure [35]. As collective variables, we used the Cα atoms of the activation loop residues and the αC helix, as well as important activation loop side chain atoms, as described in Methods. In addition, we included the Nδ2 atom of Asn549 and the Cδ atom of Glu565, as these two atoms are part of the network of hydrogen bonds termed the “molecular brake,” which has been proposed to play a regulatory role in FGFR2 kinase activation [26].
The resultant MFEP demonstrates that the activation loop backbone structure changes in four steps (Fig 1). In step (1), residues 660 through 663 move closer to the αC helix and the kinase’s N-lobe. Concurrently, the αC helix moves closer to the C-lobe. This apparently facilitates the formation of hydrogen bonds between the Nζ atom of Lys526 in the αC helix and the hydroxyl groups of Thr660 and Thr661. In step (2), the Ile654 side chain moves away from Arg649, clearing space for the side chain of pTyr657. This motion is accompanied by the sliding of the pTyr656 and pTyr657 backbone along the loop connecting the αEF and αF helices. In step (3), the pTyr656-pTyr657 backbone rotates to form the short antiparallel β-hairpin with Val679 and Tyr680 seen in the active crystal structure. This rotation accommodates two important sidechain motions, namely the inward migration of pTyr657 and Lys659, which allow for the formation of the network of hydrogen bonds in the active structure of FGFR2 kinase. Finally, in step (4), residues 660 through 663 move outward.
The major sidechain motion involved in the activation pathway is the inward motion of pTyr657 to make contact with Arg649, Arg625 and Lys659. However, we observed another important sidechain motion that occurs during the activation process, namely the motion of Arg664 toward ATP (Fig 2A). In the inactive conformation, and in the first 19 frames of the activation process, Arg664 points outward or makes contact with Glu527, enabled by proximity of the αC helix to the activation loop facilitated by the backbone motion of step (1). In the active conformation, however, Arg664 makes contact with the γ-phosphate of ATP stabilizing its position. We observed that the simulated motion of Arg664 toward ATP is synchronous with the motion of pTyr657 toward Arg649 (Fig 2B). Additionally, the dissolution of the hydrogen bond between Asn549 and Glu665, part of the regulatory “molecular brake” thought to prevent autoactivation of the kinase [26], occurs one frame after motion of pTyr657 (Fig 2C). This suggests that these two conformational changes might be structurally related as well, although a structural mechanism for this coupling is not readily apparent from this simulation study. In order to test our results, we performed the same algorithm but with an alternate set of CVs based on interatomic distances, discussed further in S1 Text.
We calculated the free energy as a function of the collective variables chosen in this string method study [34]. The plot of the potential of mean force (PMF) along the activation pathway indicates that there are two free energy wells corresponding to the inactive and active conformations (Fig 2D). The activation barrier occurs at frame 21, the same frame during which pTyr657 rotates inward and Arg664 approaches ATP, confirming that these two structural changes define the inactive and active states.
In order to pinpoint general features of inactive and active conformations of the activation loop, without reference to a particular pathway, we ran a metadynamics simulation [36]. This generated a large pool of conformations similar to the inactive and active crystal structures as well as intermediate or related conformations. This used two contact map collective variables as the basis of the metadynamics simulations. Essentially, each collective variable corresponds to the number of interatomic contacts in the activation loop that are similar to contacts in the inactive or active structures, respectively; more details are discussed in Methods. The metadynamics simulation trajectories confirm the presence of two large free energy wells, roughly corresponding to inactive and active conformations of the protein (Fig 3A). The simulation was run for a long enough time to generate a large number of stable conformations with significantly divergent activation loop structures (see S1 Fig), rather than until convergence of the free energy landscape, which would likely have required unrealistic amounts of simulation time.
Clustering of the resulting pool of conformations based on a hierarchical agglomerative clustering scheme produced a final set of clusters in which no two conformations in any cluster were more than 3.0 Å apart, measured by RMSD of the activation loop backbone Cα atoms. This resulted in a total of 56 clusters. We then connected clusters whose conformations were no more than 3.8 Å apart (Fig 3B). We observed that eight of the 56 clusters represented “active” conformations, in which pTyr657 was rotated inward and made contact with Arg649, Arg625 and Lys659. In each of these clusters, two features were notable at the kinase’s active site (Fig 3D). First, the sidechains of Leu665 and Pro666 were rotated away from the active site. It is reasonable to conclude that this orientation of these side chains is necessary to allow catalysis because it enables the substrate tyrosine to insert near ATP and the presumed catalytic base, Asp626. A similar observation was made with regard to FGFR1, for which rearrangement of the activation loop prevents Arg661 and Pro663 from blocking the active site [11]. Second, in all eight of these clusters, Arg664 was pointed inward and made contact with the γ-phosphate of ATP, confirming that formation of this contact links with the inward motion of pTyr657.
An additional 11 out of 56 clusters featured both of these structural changes at the active site, namely Arg664 pointing into the active site and Leu665 and Pro666 pointing out of the active site. In these clusters, however, pTyr657 does not point inward to make contact with Arg649. Despite this, the backbone conformations of these clusters strongly resemble those of the eight clusters in which pTyr657 points inward. The average graph distance between each of these 11 clusters (active site ready, pTyr657-out) and the nearest pTyr657-in cluster is 2.4 Å, compared to 2.9 Å for all pTyr657-out clusters. This suggests that the backbone conformation common to both groups of clusters enables both Arg664 to point inward, and Leu665 and Pro666 to point outward from the ATP site. In turn, inward rotation of pTyr657, and the subsequent formation of contacts between pTyr657 and Arg649, Arg625, and Lys659, stabilizes this backbone conformation in order to preserve the catalytically permissive conformation of the activation loop near the active site. We examined the collective variable values for conformations in each of the three groups of clusters—active conformations, conformations with the pre-catalytic active site and the active-like backbone, and inactive conformations (Fig 3C). Notably, conformations with the active-like backbone form one of the smaller free energy wells (inside the brown dotted outline in Fig 3A) comprising the large free energy well corresponding to the inactive state. This free energy well is adjacent in CV space to the free energy well corresponding to the active state, suggesting that the active-like backbone structures are in an intermediate state between the fully inactive state and the active state.
The phosphorylation of pTyr657 in the activation loop shifts the conformational dynamics of the loop to favor motion of Arg664 toward ATP in this simulation. We performed NMR experiments to examine the effect of activation phosphorylation on loop dynamics, by monitoring the chemical shift of the Arg664 HSQC cross-peak (Fig 4). When ATP was added to the unphosphorylated FGFR2 kinase, a chemical shift perturbation was seen for the Arg664 peak, indicating that the presence of ATP altered the chemical environment of Arg664. We attribute this to the Arg664 residue in fast exchange between the outward-pointing conformation and the inward-pointing conformation. As indicated by our simulations, the presence of ATP changes the distribution through electrostatic interactions with the Arg664 side chain. However, when ATP was added to the phosphorylated FGFR2 kinase, a larger chemical shift perturbation was seen for the Arg664 peak. HSQC peaks for R664 in spectra of the apo kinases (ATP-free phosphorylated and unphosphorylated) are similar, and the observed chemical shift perturbations in the ATP-bound samples lie along a straight line. These shift perturbations reflect change in the population between two endpoints [15]. The greater magnitude of the perturbation for the phosphorylated kinase is consistent with the simulation, with phosphorylation of the activation loop altering loop dynamics to enable greater interaction between the Arg664 guanidinium moiety and ATP.
Conformations featuring an inward-pointing pTyr657 also show Arg664 in the active site in simulation, raising the likelihood that Arg664 is involved in mediating FGFR2 kinase activity. To test this, we generated two mutants of FGFR2 kinase, R664A and R664W. Kinetic assays were performed on the wild-type kinase and on each of these mutants (Fig 5). These assays showed that mutation of Arg664 to alanine or tryptophan caused a 52% and 60% decrease in autophosphorylation activity, respectively. Neither mutation has been demonstrated to definitively cause any pathology, but the R664W mutation has been found in a human colorectal tumor specimen [37], and bioinformatics analysis suggests that the mutation is highly deleterious to the protein’s function [37, 38]. Both mutations abolish the interaction between Arg664 and ATP, thus apparently reducing with the kinase’s catalytic activity.
The remaining unanswered question is precisely what role Arg664 plays in FGFR2 kinase catalytic activity. A previous crystallographic study of an FGFR3 mutant [24] also showed the homologous residue Arg655 in a similar position near the kinase’s active site, interacting with the γ-phosphate. The authors of that study proposed that the arginine residue stabilizes the position of the substrate tyrosine residue via π-cation interactions. To investigate this possibility in FGFR2 kinase, we performed MD simulations of FGFR2 kinase with a substrate peptide bound at the active site. Two simulations were performed, one simulation in which Arg664 was kept in the active site using a one-sided harmonic restraint, and another in which Arg664 was kept out of the active site using a one-sided harmonic restraint. The simulations showed that the positioning of the substrate tyrosine was indeed stabilized by the presence of Arg664 (S1 Movie). The tyrosine residue is neatly sandwiched by two arginine residues, Arg664 and Arg630. We observed that the root-mean-square fluctuations of the substrate tyrosine as well as of the ATP residue are higher for the simulation in which Arg664 was outside the active site (Fig 6). We also performed a third simulation with a restraint confining Arg664 to interact with Asp530, thus being close to the active site but not fully inside, as seen in several crystal structures [15, 26, 27, 39]. This simulation demonstrated an intermediate degree of substrate tyrosine confinement, though the ATP thermal motion was similar to that seen in the Arg664-out simulation (Fig 6).
To investigate whether Arg664 plays a role in the phosphotransfer reaction, we performed a series of QM/MM calculations in which the system transitioned from the reactant to the product state. At each step, the ESP-derived partial charges of the QM region atoms were calculated. The charges of the migrating phosphate group, the two Mg2+ ions, and the atoms of the guanidinium moiety are shown in Fig 7. As the phosphate group migrates toward the substrate tyrosine, its total charge and the charges of each individual phosphate oxygen become more positive, while the charges of the Mg2+ ions and protons of the guanidinium moiety of Arg664 become more negative. This is in accordance with our hypothesis that Arg664, like Mg2+, enables the progress of the phosphotransfer reaction by stabilizing the electron density of the phosphate group. More extensive QM/MM calculations are still needed to further explore the role of Arg664 on the phosphotransfer reaction.
Our work demonstrates that phosphorylation of the activation loop tyrosine residues alters FGFR2 kinase dynamics, allowing both the entry of the substrate tyrosine, and repositioning of Arg664 in the active site, and leading to facilitation of the catalytic reaction. Although there are numerous crystal structures of FGFR2 kinase, it has not been possible to infer mechanistic detail thus far because the static structures captured by X-ray crystallography do not encompass dynamic motions of the activation loop. For example, in one structure, the Arg664 side chain was not fully resolved (PDB 3CLY) [25], while in others its position appears to be influenced by the presence of ammonium sulfate (PDB 2PZP [26] and 1GJO [39]), and/or seen to form contacts with Asp530 (PDB 4J95, 4J96, 4J97, 4J98, 2PVY, 2PWL, 2PY3, 2PZ5, 2PZP, 2PZR, 2Q0B, 3B2T, 1OEC) [15, 26, 27, 39], as we observed in intermediate structures in the string method trajectory. The position of Arg664 seen in our simulations was observed previously in a crystal structure of the highly homologous FGFR3 kinase (PDB 4k33) [24], but not previously in FGFR2 kinase. Quite recently, after our simulations were completed, a crystal structure of FGFR2 kinase in complex with PLC-γ has been published showing Arg664 in contact with ATP and Leu665 and Pro666 facing outward [40]. Using string method and metadynamics simulations, we have found that this conformation is stabilized specifically by phosphorylation of the tyrosine residues in the activation loop. Moreover, MD simulations and kinase activity assays illustrate the functional role of Arg664 in kinase activity. Our results illustrate the need to complement these valuable crystal structures with dynamic information gleaned from simulation studies.
The pathway generated by the string method algorithm may also reveal several important features of FGFR2 kinase activation. Notably, the first step in the MFEP is the motion of residues 660 to 663 and the αC helix toward one another. This proximity enables hydrogen bonds to form between Lys526 and the hydroxyl and backbone carbonyl groups of Thr660 and Thr661. The importance of this motion in the pathway and its resultant proximity between the activation loop and the αC helix may contribute to the stable positioning of the substrate tyrosine. The role of the Lys526 residue has been established in studies demonstrating the significant gain of function caused by the K526E mutation responsible for Crouzon syndrome [26]. In this mutant, the Glu526 residue would be unable to form hydrogen bonds with Thr660 and Thr661. However, the K526E mutant can enhance catalytic activity through a similar mechanism to the wild-type, by formation of hydrogen bonds between the αC helix and activation loop. Glu526 can form hydrogen bonds with Arg664, reminiscent of contacts in the wild-type structure between Arg664 and Asp530. As our simulations indicate, the presence of Arg664 near the αC helix contributes to stability of the substrate tyrosine’s position, so the K526E mutant is likely to support this activating mechanism as well. Previous studies indicated that the K526E mutant significantly increases catalytic activity in both the unphosphorylated and phosphorylated state [26]. Since we hypothesize that in the phosphorylated state, Arg664 favors interaction with ATP, we propose that the K526E mutant favors Arg664 interacting with Glu526 predominantly when the phosphotyrosine residues are pointed outward, not interacting with Arg649 and Arg625. Since previous studies [15] and the current simulations indicate that the tyrosine-out state is predominant, even in the phosphorylated state, it is reasonable that the K526E mutation will significantly increase catalytic activity regardless of whether the kinase is phosphorylated.
The major limitation of the string method algorithm in this system is that the activation loop visits a large range of conformations, which are not all on the MFEP. Our metadynamics simulation explored a wide range of conformations, including some that diverged significantly from crystallographically observed structures. In particular, 13 of the 56 clusters contained structures in which pTyr656, not pTyr657, was pointed inward, making contact with Arg649 and Arg625 (S2 Fig). This conformation is of particular interest because previous studies in FGFR1 kinase have shown that Tyr653 (homologous to Tyr656 in FGFR2) is phosphorylated before Tyr654 (homologous to Tyr657 in FGFR2) [41]. Additionally, the monophosphorylated kinase, in which only Tyr653 is phosphorylated, has only a 50–100 fold increase in catalytic activity for trans-autophosphorylation compared to a 500–1000 fold increase in catalytic activity in the bisphosphorylated kinase relative to the unphosphorylated kinase [41]. Thus, understanding the structure of this monophosphorylated kinase might hold the key to designing inhibitors that selectively bind to this structure, which exists for some time before the bisphosphorylated kinase, is generated. While there are no experimentally-derived structures of a monophosphorylated kinase, our simulation-derived ensemble of pTyr656-inward structures which resemble the active conformation of the monophosphorylated kinase provides a foundation for exploring the unique features of these conformations of intermediate catalytic activity. Surprisingly, the conformations from the ensemble either did not feature Arg664 in the active site or had Leu665 and Pro666 blocking the active site. Further studies are needed to examine the structural features of the monophosphorylated kinase that enable catalytic activity.
The activation loop represents a formidable challenge in understanding RTK structure and dynamics, as it adopts not one structure but a large range of conformational states. As a result, many of the current experimental techniques are unable to probe completely the conformational space accessible to the activation loop. Within the range of crystal structures of FGFR2 kinase, the activation loop adopts many different conformations (S1 Fig). FGFR2 kinase is an especially useful system for investigating activation loop dynamics because other domains do not move significantly upon activation. Whereas in many RTKs, the αC helix undergoes movement upon activation, emphasizing its role in creating the catalytically active conformation of the active site, the αC helix in FGFR2 kinase does not move independently of the N lobe. Moreover, the N lobe structural change illustrated in crystal structures is itself subtle. We did not observe in our simulations any definite increase in proximity between the lobes in the active conformations compared to the inactive conformations. This suggests that interpreting the subtle differences between the inactive and active crystal structures warrants considerable caution. Because global conformational motions of the kinase lobes or secondary structures are not apparent, any determinants of catalytic activity are likely to be concentrated in the activation loop, making it a good choice for understanding how activation loop dynamics lead to kinase activation. Elucidating the structural mechanism by which phosphorylation of the activation loop enables catalytic activity is an important step toward designing specific inhibitors of FGFR2 kinase and other RTKs. Our work offers a model for activation loop dynamics, supported by experimental data, which may prove useful in better analyzing the dynamic changes that activate RTKs.
In order to generate input structures for the string method, we used the crystal structures of the inactive and active conformations of the FGFR2 kinase domain (PDB entries 2PSQ and 2PVF [26], respectively), superimposing them so as to minimize the RMSD between the structures. We used UCSF Chimera [42] to add missing terminal residues in each structure so that the final structures included residues 458 through 768 (seen in Fig 1). We used MODELLER [43] to add missing, low electron-density non-terminal loops, generating five structures for each missing loop and choosing the structure with the lowest score. The carbon atom of AMP-PCP in the catalytically active structure was changed to oxygen so that the bound ligand was ATP. Since the crystal structure of the catalytically inactive kinase does not include ATP, it was added to the structure based on its position in the active kinase. All crystallographic water molecules and other precipitant molecules in the crystal structures were removed, as was the peptide substrate from the structure of the active kinase. The LEAP program [44] was used to generate the remainder of each structure. LEAP added phosphate groups to the tyrosine residues of the activation loop and the kinase hinge. Parameters for the phosphotyrosine residue and ATP were based on [45] and [46] respectively. In each of these structures, residue 491, which had been mutated in the crystal structure from Cys to Ala, is reverted to Cys. The correct number of Na+ counter ions were added to each structure, as well as enough TIP3P solvent to create a 10 Å buffer between the protein edge and the box wall. For all simulations, the AMBER99SB force field [47] was used. Hydrogen bonds were constrained using the SHAKE algorithm [48]. Long-range electrostatics were computed using the particle mesh Ewald algorithm [49]. All molecular dynamics simulations used a 2 fs time step.
The simulation boxes containing the inactive and active conformations were each minimized using NAMD [50] for 1000 steps of conjugate gradient minimization, keeping 500 kcal mol-1Å-2 restraints on the CA atoms of the protein, followed by 2500 steps of minimization without restraints. The initial path for the string method in collective variables is derived from the zero-temperature string (ZTS) method [51, 52], which produces the minimum energy path of the conformational transition between the inactive and active structures. It has been shown that the minimum energy path is a good choice for input to the string method in collective variables, as the minimum energy path (MEP) produced by the ZTS method is likely to be similar to the minimum free energy path (MFEP) produced by the string method in collective variables [35]. Before implementing the ZTS method, the water molecules from the inactive and active structures are removed. A pathway of four structures is created by inserting two linearly interpolated structures between the inactive and active structures. These four structures then undergo 100 iterations of the ZTS method. In each iteration, each structure is minimized in AMBER [44] with 20 steps of steepest descent minimization, using an infinite cutoff for short-range interactions, followed by reparametrization of the string of structures so that they are equally spaced from one another along the string in conformational space. After 100 iterations, the number of structures in the string is doubled by interpolating one structure between each pair of successive structures, and two structures between the middle pair of structures in the string, and the procedure is repeated. This continues until the ZTS method runs for a string of 256 structures. Thirty-two equally spaced structures are extracted from this path (starting with image 7 and ending with image 255, the last image in the path) for input into the string method in collective variables.
LEAP is run on each of these 32 structures to add back Na+ counter ions and water molecules as before, followed by rotation and translation of the box to align the protein molecules in each box to one another. Each structure is then minimized in NAMD for 2000 steps with 10 kcal mol-1Å-2 restraints on the protein atoms, followed by gradual heating to 300 K over 600 ps with the same restraints using a Langevin thermostat with a damping coefficient of 1 ps-1. Each structure is equilibrated in the NPT ensemble for 2 ns using a Berendsen barostat with a target of 1 bar and a compressibility of 4.57 × 10−5 bar-1. The resulting structures are used as input for the string method algorithm.
The collective variables (CVs) for the string method include the Cα atoms of the αC-helix (residues 526–541) and the activation loop (residues 644–683), as well as sidechain atoms of residues postulated to be important in the mechanism of activation (two sidechain atoms of the “molecular brake” [26], Asn549:Nδ2 and Glu565:Cδ; the phosphorus atoms of the phosphotyrosine residues of the activation loop; the Cζ atom of Arg649 in the activation loop which makes contact with pTyr657 in crystal structures of the active conformation; and the Nζ atoms of Lys658 and Lys659 which make contact with pTyr656 and pTyr657). During the simulations, these atoms are constrained to target values (denoted z1*, z2*, …, zn* for each of the n collective variables) with a 1.0 kcal mol-1Å-2 restraint, with the initial target values extracted from the equilibrated structures. MD simulations are run for each image in the string independently. After every 10 steps, the target values zi* are evolved according to the equation
zi*(t+dt)=zi*(t)−γ−1m−1∂F∂zidt
(1)
where γ is a friction coefficient given by 125 ps-1, m is the mass (taken for simplicity to be identical to the mass of a carbon atom) and F is the free energy. The derivative of free energy with respect to a given CV is approximated by the average value
∂F∂zi=k(〈zi*(t)−zi〉)
(2)
where zi is the instantaneous value of the CV and <…> denotes the ensemble average over 10 steps. The target values are thus updated for each CV for each image, which together form a string, a path in CV space. After the update step, the string is reparameterized such that the new target values are still on the same string in CV space, but equidistant from one another; this is easily performed by calculating the optimal distance between target values for each image and changing the target values, as noted in [35]. This process is continued until the set of target values does not change significantly over time (represented by asymptotic behavior of the RMSD of target values from their initial values; see S3 Fig), suggesting that the string now represents a minimum free energy path in collective variable space. Throughout the simulations, we prevented translation and rotation of the protein by adding 0.5 kcal mol-1Å-2 restraints on the backbone atoms of the protein except those in the αC-helix or the activation loop.
In order to visualize the final pathway, we ran simulations with 20 kcal mol-1 Å-2 constraints on the restrained atoms to guide them toward the final target values. The PMF is calculated by running simulations with 1 kcal mol-1 Å-2 constraints on the restrained atoms at the final values from the string method algorithm. This allows calculation of ∂F/∂zi for each CV, then enabling the generation of a PMF curve whose equation is given by
F(α)=∫0α∑i∂F∂zi∂zi∂αdα
(3)
We calculate the value of ∂zi/∂α using centered differences (or forward or backward differences for the starting and ending frames, respectively), and we calculate the integral using the trapezoidal method.
A second run of the string method in collective variables was run with a different set of CVs. For this second run, the CVs were interatomic distances between the centers of mass of residues in the αC-helix and activation loop, as well as sidechain atoms of the “molecular brake.” The overall conclusions from this run were similar and are summarized in the supporting information, and S1 Table, and S4 and S5 Figs.
The starting and ending target values of the CVs pertaining to the activation loop were used to run metadynamics simulations [36] using distances in contact map space from the active and inactive states as the CVs. We calculated interatomic distances between all atoms in the activation loop whose coordinates were CVs in the string method simulations. We then selected the subset of those distances which changed between the inactive state (frame 0) and the active state (frame 31) from being less than 8 Å to greater than 8 Å, or vice versa, and in which the greater distance was at least 1.5 times larger than the smaller distance. This subset, containing 31 interatomic distances, was used to define the contact map. Then we used distances in contact map space as the collective variables in a metadynamics simulation. For a given conformation,
d=(∑r(1−((r−d0)/r0)61−((r−d0)/r0)12))1/2
(4)
where r is one of the 31 interatomic distances, r0 = 8 Å, and d0 is the reference contact distance in the inactive or active state.
As the metadynamics simulation progressed, every 500 steps, a 2D Gaussian hill with height 0.7 kcal mol-1 and width 0.1 was added, centered at the current value of the contact map distance from the active and inactive conformations, respectively. We used well-tempered metadynamics [53], with a bias temperature of 4200 K, which determined the height of the Gaussian at each step. A grid with spacing 0.002 was used to store the Gaussian hills. Additionally, the metadynamics simulation was significantly accelerated by using 10 simultaneous walkers [54] which shared a collective set of Gaussian biases. The total simulation time for all walkers exceeded 3 μs. The metadynamics simulation was performed using NAMD 2.9 [50] with PLUMED 2.1 [55].
All calculations were performed on local workstations as well as TACC Stampede and Maverick [56].
Constructs of FGFR2 kinase for NMR and kinetics studies were based on [15]. Human full-length FGFR2 cDNA was purchased from Sino Biological Inc. (Beijing, China), and the kinase domain fragment was extracted by restriction digest with EcoRI and XhoI, generating residues 458 through 768, which were subsequently cloned into the pRSFDuet vector. Subsequent mutations were introduced into the construct via site-directed mutagenesis using a QuikChange II XL kit, and sequences were confirmed by DNA sequencing. All constructs used in this study included a C491A mutation that aided in protein expression [15].
Protein expression was carried out in BL21-DE3 RIPL cells. Cells were grown either in Terrific Broth (Sigma-Aldrich) or in M9 minimal media supplemented with 15N NH4Cl for NMR studies, and were induced at OD 0.6–0.8 with 1 mM IPTG overnight at 20°C. Cells were lysed and protein was purified using TALON metal affinity resin (Clontech). To generate non-phosphorylated sample, trace phosphorylation was removed by alkaline phosphatase (FastAP, Thermo Scientific), followed by purification by size-exclusion chromatography. To prepare phosphorylated sample, 10 mM ATP and 5 mM MgCl2 were added to FGFR2 kinase and autophosphorylation was allowed to occur overnight at 4° followed by exhaustive dialysis to remove excess ATP.
For NMR studies, we used constructs in which all tyrosine residues except Tyr657 in the activation loop, shown to be sufficient for full catalytic activity [15], were mutated to phenylalanine. Additionally, we used the A648T mutation to improve protein expression and stability for the NMR experiments. NMR HSQC spectra of FGFR2 kinase with the A648T mutation were obtained from a sample of 150 μM protein in a buffer consisting of 20 mM HEPES pH 7.4 and 150 mM KCl. Assignments for the A648T construct were transferred from published assignments [15]. The assignment of the HSQC cross peak corresponding to R664 was confirmed using selective un-labeling experiments by adding 14N-arginine to the growth media, as well as using selective labeling experiments by adding 13C-glycine and 15N NH4Cl to the growth media, generating a spectrum which contained cross peaks from residues immediately C-terminal to the glycine residues in an 1H-15N HNCO plane.
In kinetic assays, the tyrosine residues of the kinase domain were kept intact, with C491A as the only mutation retained to allow for protein expression. Kinetic assays of autophosphorylation were performed by coupling the hydrolysis of ATP to the oxidization of NADH through enzymes in the glycolytic pathway, as discussed in [57]. The assay contained 500 nM of unphosphorylated FGFR2 kinase, along with 1 mM ATP, 20 mM MgCl2, 1 mM phosphoenolpyruvate, 45–70 units LDH, 30–50 units PK, and 416 μM NADH. The kinase activity was monitored by measuring absorbance at 340 nm, which reflects the amount of NADH in the sample that has not yet been oxidized.
The starting structure for MD simulations of the kinase with the bound substrate peptide was taken from a structure of the kinase performing autophosphorylation (PDB 3CLY [25]). After making copies of the unit cell using UCSF Chimera [42], a portion of the substrate-acting kinase is retained, with the sequence TTNEEYLDL, while the remainder of that copy of the kinase is discarded. The ACP molecule was changed to ATP, and missing non-terminal segments of the enzyme-acting kinase were built using MODELLER as described above. Missing atoms were added with LEAP, followed by addition of Na+ counter ions and solvent to generate a box with an 8 Å margin surrounding the protein on all sides. The box was minimized in AMBER with 500 steps of steepest descent minimization followed by 1500 steps of conjugate gradient minimization, all with 500 kcal mol-1Å-2 restraints on the protein. This was followed by minimization with restraints on only the Cα atoms, using 500 steps of steepest descent and 500 steps of conjugate gradient minimization. The box was heated in AMBER to 300 K over 20 ps, with 10 kcal mol-1 Å-2 restraints on the Cα atoms, and NPT equilibration was performed for 2 ns, with 5 kcal mol-1 Å-2 restraints on the atoms of the active site (including Asp626, Asp644, Arg664, the substrate tyrosine, ATP, the two Mg2+ ions, and five water molecules which were all within 5 Å of both Mg2+ ions). These simulations used the same parameters used for the simulations described above.
To generate starting configurations for simulations for comparison of tyrosine stabilization by Arg664 positioning, we ran preparatory simulations to move Arg664 toward a given position while the active site atoms were kept fixed with a 5 kcal mol-1 Å-2 restraint. To study the effect of Arg664 near the active site, Arg664 was moved toward ATP by placing a harmonic restraint on the Arg664:Cζ—ATP:Pγ distance, with an equilibrium value of 4.5 Å, whose force constant increased linearly over 200 ps from 0 to 10 kcal mol-1 Å-2, followed by another 300 ps of simulation with the restraint constant. During the subsequent production simulations in which the root-mean-square fluctuations of ATP and the substrate tyrosine were measured, a half-harmonic potential was placed, with a force constant of 5 kcal mol-1 Å-2, whenever the distance increased beyond 6 Å. To study the effect of having Arg664 far away from the active site, a second simulation was run to move Arg664 away, using a similar harmonic restraint as above but with an equilibrium value of 25 Å; the production runs included a half-harmonic restraint activated when the distance went below 10 Å. To study the effect of having Arg664 bound to Asp530, a preparatory simulation included a harmonic restraint on the Arg664:Cζ—Asp530:Cγ distance with an equilibrium value of 4.5 Å, followed by production runs with a half-harmonic potential activated if the distance increased beyond 6 Å.
The starting structure for QM/MM studies was identical to that used in the substrate positioning simulation to study the effect of positioning Arg664 in the active site. The QM region included the three phosphate groups of ATP; two Mg2+ ions; the sidechains of Asp626, Asp644, Arg664 and the substrate tyrosine; and five water molecules that were within 5 Å of both Mg2+ ions. The rest of the system was treated at the MM level using the same AMBER force field. The QM region was studied using density functional theory with the B3LYP exchange correlation functional, using the cc-pVDZ basis set [58]. Link atoms were used to connect the two regions. All QM/MM calculations were performed using NWChem 6.3 [59]. First, the geometry was optimized using an alternating optimization scheme, in which the QM region undergoes 10 steps of optimization, followed by optimization of the MM solute region while the QM region is represented by ESP-derived charges, and then optimization of the solvent. This alternating optimization scheme continues until convergence. After optimization, we used the reaction coordinate driving method [60] to drive the system from the reactant state to the phosphorylated product state. We used the reaction coordinate RC = d1 − d2 − d3, where d1 is the ATP:Pβ–ATP:Pγ distance, d2 is the ATP:Pγ—Tyr:OH distance, and d3 is the Tyr:HH—Asp626:Oδ2 distance. The system was optimized at each step with harmonic restraints on the reaction coordinate that successively increased its value until the reaction completed.
|
10.1371/journal.pbio.2005839 | Long-term all-optical interrogation of cortical neurons in awake-behaving nonhuman primates | Whereas optogenetic techniques have proven successful in their ability to manipulate neuronal populations—with high spatial and temporal fidelity—in species ranging from insects to rodents, significant obstacles remain in their application to nonhuman primates (NHPs). Robust optogenetics-activated behavior and long-term monitoring of target neurons have been challenging in NHPs. Here, we present a method for all-optical interrogation (AOI), integrating optical stimulation and simultaneous two-photon (2P) imaging of neuronal populations in the primary visual cortex (V1) of awake rhesus macaques. A red-shifted channel-rhodopsin transgene (ChR1/VChR1 [C1V1]) and genetically encoded calcium indicators (genetically encoded calmodulin protein [GCaMP]5 or GCaMP6s) were delivered by adeno-associated viruses (AAVs) and subsequently expressed in V1 neuronal populations for months. We achieved optogenetic stimulation using both single-photon (1P) activation of neuronal populations and 2P activation of single cells, while simultaneously recording 2P calcium imaging in awake NHPs. Optogenetic manipulations of V1 neuronal populations produced reliable artificial visual percepts. Together, our advances show the feasibility of precise and stable AOI of cortical neurons in awake NHPs, which may lead to broad applications in high-level cognition and preclinical testing studies.
| This report details the first successful application of long-term all-optical interrogation techniques in monkeys. We have overcome obstacles that prevented the combination of single- and two-photon (1P and 2P) optogenetic stimulation with 2P imaging in awake-behaving monkeys, retesting targeted individual cells and neuronal ensembles over periods that extended beyond 6 months. Our strategy results in repeatable primary visual cortex (V1) neuronal stimulation of the same neurons and produces reliable visual percepts, which monkeys report behaviorally in a visual–motor task. The animals’ behavioral responses to their optogenetic-induced perceptions are comparable to their responses to real visual stimulation. These technical advances establish the feasibility of combined long-term optogenetic manipulation and 2P imaging of neocortical neurons in awake-behaving monkeys. Our approach may be applied to investigate the molecular and circuit-level mechanistic pathways that are unique to primate neural function. These methods also provide a roadmap for preclinical testing of human optogenetic therapies and may serve as the basis for optogenetic studies involving sensorimotor functions relevant to human perception, cognition, behavior, and neurological/psychiatric disorders.
| Optogenetic techniques enable the functional characterization of neuronal populations and circuits with high spatial and temporal precision [1–7]. Though relatively understudied as compared to rodents, optogenetics techniques have been applied to the study of high-level cognition circuits in NHPs [8–13], including those underlying human neurological and psychiatric disorders [14–16], and they hold the potential to unveil the mechanistic pathways for visual processing circuits that are found only in humans and NHPs (as the only mammals with retinal foveas) [17,18]. NHP studies are moreover essential for preclinical testing of optogenetic therapies before they can be translated to human applications [11,19,20].
Previous research has recorded optogenetic activation using traditional electrophysiological techniques. This approach is limited, however, because repeated electrode recordings in the same neurons are difficult to achieve across recording sessions in NHPs. In addition, examining opsin expression patterns in vivo within the area targeted by viral vector infusions, while maintaining the health of the neurons, is not currently possible without 2P laser-scanning microscopy [14,21–24]. These combined hurdles call for an all-optical interrogation (AOI) approach to the application of optogenetic methods in NHPs.
AOI is achieved by the combination of optogenetics to perturb neuronal activity, while using calcium or voltage indicators—rather than electrode-based stimulation and recording—to minimize the invasiveness of the readout [25–30]. AOI’s implementation thus allows the monitoring of large neuronal populations—repeatedly and less invasively—with single-cell resolution [31–33], while enabling detailed mapping of neural circuits during behavior [34,35]. Pioneering efforts to apply AOI in NHPs combined optogenetics with both in vivo epifluorescence imaging and intrinsic signal optical imaging [19]. Whereas these techniques allowed for large-field viewing, the spatial resolution of the readout was limited, and specific neurons of interest could not be interrogated repeatedly across recording sessions.
Here, we combined wide-field single-photon (1P) and single-cell 2P optogenetic stimulation techniques with recently developed 2P imaging technique in awake macaques [4,36] to achieve AOI in NHPs. A red-shifted opsin ChR1/VChR1 (C1V1) and calcium indicators GCaMP5G/GCaMP6s were delivered into V1 with adeno-associated viruses (AAVs) and expressed in V1 neuronal populations. The labeled V1 neurons exhibited consistently robust responses, over several months, to either optogenetic or visual stimulation. The behavioral experiments confirmed that robust artificial visual perception could be induced by optogenetic stimulation of V1 neuronal populations.
We infected area V1 neurons in three monkeys with C1V1 (AAV9–CamKIIα–C1V1(T/T)–ts–EYFP)—a red-shifted channel-rhodopsin transgene—and GCaMP5G/GCaMP6s (AAV1–hSyn-GCamP5G/AAV1–Syn–GCamP6s)—calcium indicators of activity. Six weeks after virus injection, a 1-cm–diameter round optical window (glass coverslip attached to a titanium ring) was implanted onto the cortical surface using dental acrylic cement attached to the bone surrounding the craniotomy. To enhance the stability of 2P imaging, we used a three-point head-fixation design, with two head posts implanted on the forehead of the skull and one on the back. A T-shaped steel frame was connected to these head posts for head stabilization during subsequent imaging and stimulating sessions [36].
We imaged layer II/III neurons in the infected cortical area using 2P (Fig 1A). Dark cell bodies indicate that C1V1–ts–EYFP expression was localized to the membrane [37] (see also S1 Fig). Fluorescence of GCaMP6s was relatively weak in the absence of cellular responses to either visual or optogenetic stimulation.
NHPs maintained fixation while visual stimuli consisting of drifting gratings and color patches were presented sequentially on the neuronal receptive field for 1 second, with >2-second interstimuli intervals. We recorded robust neuronal calcium responses that showed normal orientation and color selectivity, as well as well-organized receptive-field spatial organization (Fig 1B, S3 and S4 Figs).
We then stimulated the neurons optogenetically. Using 1P stimulation (532-nm laser), we illuminated the entire imaging field (a 1-mm2 laser spot) while measuring neuronal activity simultaneously with 2P imaging. Simultaneous stimulation/imaging presented a significant challenge, because—although the stimulation and recording wavelengths were sufficiently separated and filtered optically—the optogenetic stimulation power was orders of magnitude higher than the fluorescence power emitted by the activated cells. Thus, stimulation light leaked through the filters and into the highly amplified photomultipliers (PMTs), with higher power than the relatively small GCaMP fluorescence signal. Therefore, the full-field optogenetic stimulation laser was powered down whenever each 2P imaging scan targeted the central 75% of the FOV (24 ms out of each 32-ms imaging scan frame). Thus, the entire field was stimulated for 8 ms out of every 32-ms scan (25% duty-cycle stimulation at 31.25 Hz). This allowed us to view the optogenetic activation responses artifact-free (S6 Fig).
The cells that were successfully stimulated optogenetically constituted a considerable fraction of the targeted population and responded vigorously (Fig 1C). By repeatedly stimulating—both optogenetically and visually—we made three observations: 1) responses from the two modes of stimulation were comparable to each other in both amplitude and dynamics (Fig 1D and 1E); 2) repeated stimulation resulted in similarly sized responses (Fig 1F); 3) optogenetic activation did not alter the receptive field properties of neurons that were subsequently stimulated with visual stimuli (S3A Fig). Notably, the dose-response curve revealed that the average laser-evoked responses were saturated at approximately 0.8 mW/mm2, indicating high sensitivity of the optical manipulation system (Fig 1G and S2 Fig).
Using AOI, we assessed the long-term stability of both transgene expression and the physiological response strength to visual and optogenetic stimulation in the behaving NHPs. Transgene expression level and pattern were maintained (Fig 2A), and neurons exhibited consistently robust responses and tuning to visual stimuli (Fig 2B, 2D and 2I, S3 Fig from monkey M1 and S4 Fig from monkey M3) over several months. The same neuronal population was also repeatedly and stably activated by optogenetic stimulation over a 4-month period (Fig 2C, 2G and 2H). We also evaluated the transgene expression at different cortical depths, from the surface to 500 μm. At 10 months post-infection, there was abundant expression between 150 to 300 μm (Fig 2E and S5 Fig), and neurons in this depth range responded robustly to optogenetic stimulation (Fig 2F). Thus, both expression level and optogenetic responses remained stable over long time periods (in our experience, 6 months or more) in NHP cortex.
A powerful way to assess neural circuit function is to photostimulate an individual neuron (minimizing stimulation of unwanted targets) while simultaneously monitoring the activity of the connected neurons in the network [5,16,35]. To perform simultaneous single-cell–resolution 2P optogenetic activation with 2P calcium imaging of the neuronal population, we added a second optical path to our microscope—driven by a mode-locked femtosecond laser (λ = 1070 nm, 50 fs)—and applied 2P stimulation with spiral galvanometer scanning targeted to the somas of the target cells [34].
To examine the spatial specificity of 2P activation, we measured the calcium response of the targeted neuron as a function of multiple stimulation sites (5 × 5 grid) (Fig 3A) [34,35]. We sequentially stimulated each of the sites using 2P spiral activation. Robust responses in the central neuron were evoked only when the target neuron was directly targeted (Fig 3B–3D), suggesting that spiral 2P stimulation has high spatial precision and must be focused on the neuron for strong optogenetic activation to occur. We then simultaneously monitored and sequentially manipulated several neurons in one imaging field (Fig 3E). Each of these neurons generated strong responses only when targeted by the 2P activation laser (Fig 3F and 3G).
To assess the monkeys’ perception from optogenetic stimulation of V1 neuronal populations, we designed a “GO”/“NO GO” visual object detection task, in which two monkeys were required to report the appearance of a visual cue using eye movements (Fig 4A).
Each trial began when the NHP fixated the central fixation point. Subsequently, a 0.5-degree Gaussian white dot was presented for 22 ms at an eccentricity of approximately 3 degrees as a visual cue for GO (an eye fixation break), and the NHP was rewarded for producing a saccade within 500 ms. On the NO GO trials (50%, no visual cue), the animal was rewarded for holding fixation for 2,000 ms for the entire trial. Training proceeded until the NHPs conducted this task with high accuracy (>80% correct rate; Visual Stim; Fig 4C). Notably, both monkeys tended to make eye movements towards the location of the visual cues (Fig 4D, green; SD of saccade endpoints from the target: 0.22 and 0.69 degrees for Monkey M1 and M2, respectively), though any saccade exceeding 1 degree in magnitude was sufficient to receive a reward.
We then examined the artificial visual perception generated by optogenetic stimulation—Opto Stim. The GO condition here had no visual cue. Instead, we conducted optogenetic stimulation (a 532-nm, 66-ms laser pulse, subtending 1 mm2 for Monkey M1, and a 15-Hz, 33% duty-cycle [22 ms on, 44 ms off], 0.8-mW laser pulse train for Monkey M2) at the position of the C1V1-expressing cortex (about 3 degrees eccentric from the fovea, in a different position from the stimulus in the Visual Stim block, so that saccadic targeting would indicate the monkey’s differential perceived stimulation within visual space). Similar to the Visual Stim condition, monkeys in Opto Stim received a juice reward if they produced a saccade (>2 degrees) after the optogenetic stimulation. Both monkeys performed this task well after 3–5 sessions as a result of Opto Stim, with 99% versus 96% accuracy for Monkeys M1 versus M2, respectively (Opto Stim; Fig 4C). The eye movements correctly targeted the stimulation locations within visual space, corresponding to the retinotopic C1V1-expressing loci (which were never otherwise targeted with Visual Stim cues; SD of saccade endpoints from the target: 0.33 and 0.51 degrees for Monkey M1 and M2, respectively). This further confirmed that optogenetic stimulation successfully induced artificial visual perception in the NHPs (Opto Stim; Fig 4E).
To rule out the possibility that any of the observed effects were due to artifacts resulting from the physical side effects of Opto Stim, we interleaved Mistargeted Stim trials (8.3%) in the GO condition: Here, we redirected the laser to a region of V1 cortex that did not express C1V1 (Fig 4B). This mistargeted laser should not have been capable of evoking either optogenetic activation of neurons or artificial visual perception. This control condition was treated as a GO task, and monkeys were again rewarded for saccades in any direction, launched immediately after laser onset (<500 ms). Despite this incentive, we observed significantly fewer saccades in the control condition (p < 10−20 for Monkey M1 and p < 10−10 for Monkey M2) (Opto Stim versus Mistargeted Stim; Fig 4C), indicating that the monkeys were truly not aware of the mistargeted laser stimulation. Note that we sometimes observed saccades in the “No Stim” period before the “Opto Stim” block that were biased slightly toward the optogenetic target area, perhaps because perception induced by our Opto Stim condition was weaker than from Visual Stim and thus the monkeys were more likely to guess. But in general, the percentage saccades launched was much lower in the “No Stim” condition.
We also studied saccadic latencies as a function of stimulus type and duration. For Visual Stim, saccadic responses were swift and robust (Fig 5A), and exhibited consistent latencies of approximately 119 ms, measured as the time between cue onset and the saccade crossing the 1-degree magnitude threshold (Fig 5D). During Opto Stim (2.4 mW/mm2), we found that laser pulses of 44-ms duration (or more) elicited robust responses (Fig 5B). For Monkey M2, similar results were found (Fig 5F–5K). Two pulse stimulation (44-ms total duration) induced responses in 92% trials (Fig 5K) in Monkey M2, similar to Monkey M1 under 44-ms duration single-pulse photostimulation. Saccadic latencies from optogenetic stimulation in both monkeys was 30–40 ms shorter than from visual stimulation, averaging about 90 ms after laser onset (Fig 5D and 5J). This 30–40 ms difference arose presumably because optogenetic stimulation bypassed the subcortical visual pathway. This observation is consistent with previous studies of visual signal propagation from the retina to V1 [38,39].
Optogenetic applications in NHPs have facilitated our understanding of sensory processing, decision making, and the bases of cognition [9,10,14,21,40] and will likely play a key role in future brain–computer interfaces, neural prosthetics, and methods to counteract cognitive decline in the aging human brain. As such, optogenetic techniques are undergoing rapid translation to human clinical use. A critical step in the approval, implementation, and efficacy of optogenetic therapies will be preclinical testing in NHPs, for which methods are currently lacking.
Here, we combined optogenetic stimulation with 2P calcium imaging of neuronal responses to achieve AOI in awake-behaving macaque monkeys, in which we co-infected V1 neurons with C1V1 and GCaMP6s and monitored calcium signals using 2P microscopy while stimulating optogenetically. Our experiments revealed consistently robust neuronal responses to both visual and optogenetic stimulation over many months (Figs 1 and 2). 2P optogenetic stimulation also evoked strong neuronal responses with targeted single-cell resolution (Fig 3). Optogenetic milliwatt-level stimulation in V1 cells produced strong and specific responses in functionally identified visual cells. Finally, we compared optogenetically derived to visually derived perception by assessing the dynamics of saccadic eye movements produced in response to both modes of stimulation (Figs 4 and 5). Together, the above results demonstrate the high sensitivity and stability of our AOI strategy.
Channelrhodopsin-2 (ChR2) is a commonly used optogenetic actuator for NHPs, though it often requires high laser power to evoke neuronal and behavioral responses [12–15,21]. The high conductance and red-shifted absorption spectrum of C1V1 makes it a preferable choice [10,37,41,42]. This is especially true for AOI experiments, since C1V1’s excitation spectrum is well separated from that of GCaMPs [34,35]. Expression of C1V1–ts–EYFP was robust, and we observed membrane-localized EYFP fluorescence, which has previously indicated membrane localization of C1V1 [41] (Fig 1 and S1 Fig). We visualized GCaMP6s fluorescence and filtered the widefield 1P stimulation pulses with a 500 ± 12.5-nm filter to block most of the EYFP fluorescence. Although the imaging quality was somewhat reduced due to the filter, we nevertheless identified robust responses derived from both visual and optogenetic stimulation (Figs 1 and 2). Note that we did not achieve high efficiency of co-expression of C1V1–ts–EYFP and GCaMP6s in single neurons, and we found that many neurons could be activated by our wide-field illumination but not by our single-cell photostimulation.
Precise quantification of the single-cell expression levels was not possible with our methods because the bright background fluorescence in our approach was likely contributed to by fluorescence of other neurons due to the membrane-bound targeting of the specific indicator we chose. Dendrites from other neurons—and even the soma membranes of directly abutting neurons—could not be perfectly isolated from any given target neuron. We expect that this issue of precise single-cell quantification would be ameliorated by using an indicator that expresses within the cytosol (labeling the cell body only) rather that the membrane. This is why we also tested C1V1–porcine teschovirus-1 2A (P2A)–mCherry, with the hope that the mCherry would express inside the cytosol of cell bodies, thus allowing direct quantification of single-cell fluorescence. Alas, the efficiency of photoactivation of this construct was much lower than the C1V1–ts–EYFP, for unknown reasons. We are thus currently working to develop soma-targeted C1V1-EYFP for both high efficiency photoactivation and quantification of expression [5]. We conclude that more powerful molecular tools and gene delivery techniques further advance their utility in NHPs.
AOI using C1V1 results in much lower tissue damage than what would be caused by repeated probe penetration, or from the photodamage expected with ChR2 constructs [16,20]. A primary limitation of our method arose from the 2P imaging-depth limit. This confined our AOI to superficial cortical circuits, lying within 500 μm of the surface [43]. New multiphoton microscopy methods will improve and extend the depth limit of AOI to as deep as 1 mm [44]. Cellular-resolution imaging of subcortical structures is currently achievable with fiber-optic confocal laser endomicroscopy (CLE) techniques [45,46].
Because the co-expression level of C1V1 and GCaMP was low in our experiments, it is unlikely that we accidentally stimulated unseen dendrites of untargeted neurons while stimulating our target neurons. But this problem—unwittingly stimulating unwanted hidden dendrites that drive neurons other than the targeted neurons—will rise in significance as expression density improves. Soma-targeted opsins serve to minimize this concern [5], which is why we are currently working to develop soma-targeted C1V1–EYFPs that could improve specificity of 2P stimulation.
Electrical microstimulation of the visual cortex evokes phosphene perception in humans, as well as saccadic eye movements in NHPs [39,47–50]. Similarly, optical stimulation of monkey V1 has been reported to induce saccades [13], which we also observed. One refinement of our current design over prior work was to include a control condition in which we targeted an unlabeled region of cortex to rule out potential non-optogenetic artifacts related to laser activation. Interestingly, the animals did not immediately respond to Opto Stim when switching from the Visual Stim block. Though both monkeys required fewer than 30 trials to first detect the Opto Stim, this could indicate that the percept derived by the Opto Stim was not identical to that derived from the Visual Stim. If so, the monkeys might have generalized their initial responses to novel stimuli (triggered by the Opto Stim), in much the same way as they might do during operant conditioning of an unfamiliar visual stimulus. Moreover, we discovered that saccadic responses were faster when elicited by optogenetic stimulation of visual cortex than by real visual stimuli, which follows from the known latencies of transmission from the retina and subcortical visual pathway.
Because of the tight homology between the human brain and the NHP brain, the functional characterization of neurons and neural circuits underlying high-level cognition—and cognitive decline—as well as neurological and psychiatric disorders remains heavily dependent on NHP research. Primates, moreover, are the only foveate mammals; thus, they are the only animal model with human-equivalent visual capabilities and oculomotor behaviors [17,18], which makes NHPs a critical animal model for human visual perception, as well as the development and testing of clinical therapies and neural prosthetics. By integrating optogenetics and calcium imaging, AOI offers the ability to precisely determine and manipulate fine functional maps in real time during NHP behavior. One of AOI’s main functions is the precise manipulation of single neurons and simultaneous monitoring of connected neuronal activity to determine the strength of connectivity within neural circuits without unwanted activation of nearby targets.
All procedures involving animals were in accordance with the Guide of Institutional Animal Care and Use Committee (IACUC) of Peking University Animals, and approved by the Peking University Animal Care and Use Committee (LSC-TangSM-5).
Rhesus monkeys (Macaca mulatta) were purchased from Beijing Prima Biotech, Inc. and housed at Peking University Laboratory Animal Center. The study used three healthy adult male monkeys 4–6 years of age and weighing 5–7 kg.
Two sequential sterile surgeries were performed on each animal under general anesthesia. In the first surgery, a 16-mm–diameter craniotomy was created in the skull over V1. We opened the dura and injected 200 nl of a 1:1 mixture of AAV1.Syn.GCaMP6s.WPRE.SV40 (CS0564, titer 2.2e13 [GC/ml], Penn Vector Core) or AAV1.hSyn.GCaMP5G.WPRE.SV40 (V4102MI-R, titer 2.37e13 [GC/ml], Penn Vector Core) and AAV9.CamKIIa.C1V1.TS.eYFP.WPRE.hGH (V4545MI-R, titer 1.6e13 [GC/ml], Penn Vector Core) at a depth of approximately 350 μm.
Injection and surgical protocols for each NHP followed from a previous study [36]. Briefly, a small cover glass (6 mm in diameter) with a single pore (0.3 mm in diameter) was used to target the injection pipette and stabilize the cortical surface during each injection. The quartz pipette (QF100-70-7.5, Sutter Instrument, USA) was pulled with a 15–20 μm tip using a laser-based pipette puller (P-2000, Sutter Instrument, USA) and used for virus injections. After injections, we sutured the dura, replaced the skull cap with titanium screws, and closed the scalp. The animal then returned to its cage for recovery and received Ceftriaxone sodium antibiotic (Youcare Pharmaceutical Group Co. Ltd., China) for one week. A second surgery was performed 45 days later to implant the head posts and imaging window. We used a three-point head-fixation design, with two head posts implanted on the forehead of the skull and one on the back. A T-shaped steel frame was connected to these head posts for head stabilization during subsequent imaging and stimulating sessions.
We trained each monkey to sit in a primate chair with its head restrained while performing visual fixation and behavioral choice tasks. Eye position was monitored with an infrared eye-tracking system (ISCAN, Inc.) at 120 Hz. Each trial started with the eye fixated on a white 0.1-degree point within a window of 1 degree. Visual stimuli were generated using a ViSaGe system (Cambridge Research Systems) and displayed on a 17-inch LCD monitor (Acer V173, 80Hz refresh rate) positioned 45 cm from the animal’s eyes. Receptive fields of C1V1- and GCaMP-expressing sites were initially localized with small patches of drifting oriented gratings.
We designed a two-block “GO”/“NO GO” detection task in which NHPs made targeted saccades as a means to report perceptually detected Visual Stim or Opto Stim cues (Fig 4A).
In the first block (Visual Stim), a real visual object was presented on the monitor as a GO cue. This visual object was flashed for 22 ms approximately 3 degrees peripheral to the fixation point. The NHPs were trained to generate a saccade within 500 ms of cue onset to obtain a juice reward. The central fixation point remained unchanged for the duration of the trial.
In the Opto Stim block, a 1P laser pulse (with a wavelength of 532 nm, a 1.0-mm diameter, 0.2–2.4 mW/mm2, and a duration of 22, 44, or 66 ms) was projected onto the C1V1-expressing cortical site in each monkey as a GO cue instead of a real visual object. We interleaved trials with either mistargeted laser stimulation of the cortex (to an area without C1V1 expression; 8.3% trials; Mistargeted Stim in Fig 4C and 4D) or without laser stimulation (66.7% trials; No Stim in Fig 4C and 4D) as control trials, allowing us to rule out artifacts related to laser operation.
We used a ratio of 1:1 (No Stim:Visual Stim) in visual stimulation sessions, whereas we used a ratio of 8:3:1 (No Stim:Opto Stim:Mis Stim) in optogenetic stimulation sessions. By using fewer Opto Stim trials, we sought to increase the confidence level of the response data by increasing the NHP decision criteria.
After a 10-day recovery period following the second surgery, the animals were trained to fixate their gaze on a fixation point. Imaging was performed using a Prairie Ultima IV 2P microscope (Bruker Nano, Inc., FMBU, formerly Prairie Technologies) and a Ti: Sapphire laser (Mai Tai eHP, Spectra Physics) with a 16× objective (0.8-N.A., Nikon). Whereas 920 nm is a commonly used wavelength for 2P imaging in rodents, we used 1,000 nm for our 2P imaging because we found that it achieved higher quality images (and at deeper depths) in our NHP experiments [36]. Fast resonant scanning (up to 32 frames per second) was used to obtain images of neuronal activity (8 fps by averaging every 4 frames). To discriminate GCaMP5G from C1V1–ts–EYFP, GCaMP5G fluorescence was acquired with a 920-nm excitation laser using a 500- ± 12.5-nm filter, whereas EYFP fluorescence was acquired with a 1,040-nm excitation laser using a 525- ± 35-nm filter.
To achieve 2P imaging with a 1,000-nm excitation source, our power density was approximately 7e-5 mW/um2 (<50 mW scanning over an 850-μm × 850-μm area), which was approximately 10,000 times less power than the stimulation power level (approximately 0.3 mW/um2, 30 mW, 1,070 nm focused on a diameter of 10 μm). Our imaging laser power could therefore not have caused significant photostimulation of C1V1. Even if it did, it follows that its effects must have been approximately 10,000 times smaller than the effects of our intended photostimulation [34].
A 532-nm laser was used for 1P optical stimulation. The laser was directly pointed at the target cortical area through the imaging window. Due to the brightness of the stimulation laser and the high sensitivity of the PMTs, a 500-nm band pass (25-nm width) filter was inserted before PMT of green channel during simultaneous 2P imaging. Nevertheless, simultaneous stimulation light could have potentially leaked through the filtering system to cause recording artifacts. We addressed this potential confound by blocking the 532-nm laser light during the scanning of the central image during 2P recordings, using an electronic circuit that powered down the full-field stimulation pulse whenever the imaging scan was within the central 75% of the FOV (S6 Fig).
A secondary femtosecond laser with 1,070-nm wavelength (maximal power, 2.3 watts; pulse width, 50 fs; Fidelity, Coherent, USA) was used on a secondary galvanometer path in the 2P microscope (Ultima IV, Prairie, Bruker, USA) to perform 2P optogenetic activation targeting single cells, while simultaneously recording calcium activity. Spiral regions (5 rotations, 1.2 expansion rate, 0.01 pixel/μs, and 30 repetitions) were defined to point target photo-activation areas (S7 Fig). The laser power was adjusted to 30 mW at the end of the objective with a polarization beam splitter into 1,070-nm femtosecond laser light pathway.
Customized Matlab software (The MathWorks, Natick, MA) was used to do data analyzing. To correct the image shifts caused by the movement between the objective and the cortex, we first obtained a template image by averaging 1,000 frames in the middle of an imaging session and then realigned images from each session to the template image using a normalized cross-correlation–based translation algorithm.
The visual stimuli were randomly interleaved during experiments.
No data were excluded during analysis.
Customized Matlab software was used to perform statistical analysis. As demonstrated in the figure legends, data were presented as individual data points or as mean ± SEM. Number of repetitions for each experiment was also noted within the figure legends.
The genetic constructs used in this work are available via Addgene.
|
10.1371/journal.ppat.1005382 | Regulation of Rac1 and Reactive Oxygen Species Production in Response to Infection of Gastrointestinal Epithelia | Generation of reactive oxygen species (ROS) during infection is an immediate host defense leading to microbial killing. APE1 is a multifunctional protein induced by ROS and after induction, protects against ROS-mediated DNA damage. Rac1 and NAPDH oxidase (Nox1) are important contributors of ROS generation following infection and associated with gastrointestinal epithelial injury. The purpose of this study was to determine if APE1 regulates the function of Rac1 and Nox1 during oxidative stress. Gastric or colonic epithelial cells (wild-type or with suppressed APE1) were infected with Helicobacter pylori or Salmonella enterica and assessed for Rac1 and NADPH oxidase-dependent superoxide production. Rac1 and APE1 interactions were measured by co-immunoprecipitation, confocal microscopy and proximity ligation assay (PLA) in cell lines or in biopsy specimens. Significantly greater levels of ROS were produced by APE1-deficient human gastric and colonic cell lines and primary gastric epithelial cells compared to control cells after infection with either gastric or enteric pathogens. H. pylori activated Rac1 and Nox1 in all cell types, but activation was higher in APE1 suppressed cells. APE1 overexpression decreased H. pylori-induced ROS generation, Rac1 activation, and Nox1 expression. We determined that the effects of APE1 were mediated through its N-terminal lysine residues interacting with Rac1, leading to inhibition of Nox1 expression and ROS generation. APE1 is a negative regulator of oxidative stress in the gastrointestinal epithelium during bacterial infection by modulating Rac1 and Nox1. Our results implicate APE1 in novel molecular interactions that regulate early stress responses elicited by microbial infections.
| Helicobacter pylori infection of the gastric mucosa is largely lifelong leading to continued stimulation of immune cells. This results in the generation of reactive oxygen species (ROS) which are produced to kill bacteria, but at the same time ROS regulate cellular events in the host. However, prolonged generation of ROS has been implicated in damage of DNA, which ultimately could lead to the development of cancer. We studied a molecule known as APE-1 in gastric and intestinal cells, which is activated upon encounter of ROS. Our results show that APE1 limits the production of ROS in cells that form the lining of the gastrointestinal tract. APE1 regulates ROS production by inhibiting activation of the molecule Rac1. Inhibition of ROS production by APE1 occurred after infection of gastric cells with Helicobacter pylori and after Salmonella infection of intestinal cells. These data demonstrate that APE1 inhibits production of ROS in cells that line the inside of the digestive tract.
| The gastrointestinal epithelium serves as an initial interface between the host and luminal microbiota [1] and initiates innate immune responses to infection. Gastric and intestinal epithelial cells infected by microbial pathogens or commensal microbiota typically activate Rho GTPases leading, amongst other effects, to the production of reactive oxygen species (ROS) [2,3] that arise from the activation of the NADPH oxidase complex (Nox1) [4]. Nox1 family proteins are the catalytic, electron transporting subunits of Nox1 in non-phagocytic cells that produce superoxide [5,6]. While production of microbicidal levels of ROS in professional phagocytes via Nox2 is well-studied, information on ROS generation by gastric and intestinal epithelial cells in response to microbial signals via epithelial Nox1 is limited. The levels of ROS produced by epithelial cells are much lower than in phagocytes, and are more important in redox-sensitive signaling than direct antimicrobial killing. Nox1 is associated with the membrane-integrated protein p22phox, NOXA1 and NOXO1 to form superoxide [5]. Nox1 is expressed in gastric tissues [4] and is thought to play a role in ROS production in H. pylori-infected human gastric epithelial cells. While NADPH oxidase can be activated in epithelial cells throughout the gut, little is known about its responses to enteric infection.
Helicobacter pylori causes a lifelong infection that can lead to gastric and duodenal ulceration and gastric cancer, one of the major causes of cancer mortality worldwide [7,8,9]. Following H. pylori infection of guinea pigs [10], humans [11] and cultured gastric epithelial cells [12], an increase in oxidative stress occurs. H. pylori lipopolysaccharide (LPS) activates the small GTPase, Rac1, leading to Nox1 activation and production of superoxide [10,13,14,15]. Since H. pylori is a persistent infection, chronic ROS exposure eventually leads to oxidative DNA damage [4,16,17] and activation of signaling pathways implicated in the pathogenesis of cancer [18,19].
Accumulation of ROS increases APE1 activation [20] which in turn, mediates vital functions designed to protect the host [18]. APE1 is a multifunctional protein that is widely express in epithelial cells and that regulates multiple responses to bacterial infections, including chemokine production, apoptosis, cell proliferation and responses to hypoxia. The carboxy-terminus of APE1 is responsible for repairing DNA damage induced by ROS, while its N-terminal region regulates transcription [18]. Another distinct transcriptional regulatory role of APE1 is mediated by the N-terminal Lys6/Lys7 acetylation, which modulates certain promoter activities [21,22,23]. We have shown that APE1 is upregulated in gastric epithelial cells in the context of H. pylori infection [20] and contributes to the activation of AP-1 and NF-κB that regulate cell responses, including IL-8 production [24,25] and inhibition of cell death during H. pylori infection [26]. Interestingly, in a model of mouse hepatic ischemia/reperfusion, overexpression of APE1 resulted in suppression of reperfusion-stimulated oxidative stress [27]. While infection of gastric epithelial cells with H. pylori is a suitable model system to study the mechanisms of APE1-mediated regulation of ROS, Salmonella enterica serovar Typhimurium can be used as model to study the mechanisms of ROS production by intestinal epithelial cells (IEC). The pathogenicity of Salmonella is in part dependent on the presence of the Salmonella pathogenicity island 2 (SPI2) that interferes with ROS production by Nox2 in macrophages [28,29]. As many of the established infection-induced effects on gastrointestinal physiology are mediated by ROS-dependent mechanisms, we sought to compare the role of APE1 in ROS generation following infection with gastric or enteric pathogens.
In the current study, we provide evidence that H. pylori- and Salmonella-induced ROS is inhibited by APE1 in gastric and intestinal epithelial cells respectively. We also demonstrate that the Lys residues at the N-terminus of APE1 at positions 6 and 7, are required for Rac1 binding. This interaction inhibits Rac1 activation and Nox1 expression, decreasing ROS generation that results from infection. Together, our findings show a novel role of APE1 in regulating ROS levels in gastrointestinal epithelial cells following infection.
Empty retroQ vector (pSIREN), APE1 shRNA expressing (shRNA) cells, or non-transfected AGS (AGS) cells obtained from American Type Culture Collection were harvested and cultured in Ham’s F/12 medium (Hyclone) supplemented with 10% heat-inactivated FBS (Hyclone) [21]. NCI-N87 cells obtained from ATCC were maintained in RPMI supplemented with 10% FBS. T84 and HT-29 cells (a kind gift from Dr. K. Barrett, University of California San Diego) were maintained respectively in L-Glutamine containing F12/DMEM supplemented with 5% FBS and in McCoy’s 5A medium supplemented with 10% FBS. H. pylori 26695, a cag PAI+ strain (ATCC) and its isogenic mutants, cag PAI− strain 8–1 and VacA (kind gift from Dr R.M. Peek, VanderBilt University, Tennessee, USA [30]), were maintained as previously described [21]; a MOI of 100 was used for all the experiments in this study as this was the highest dose with minimal necrotic cell death [26]. Previously, we reported that infection of gastric epithelial cells with H. pylori longer than 6h cause cell death and therefore, longer infection times do not result in reliable ROS data [26].
Gastric antrum-derived primary epithelial cells were isolated and maintained in culture according to the procedures developed by Dr. Stappenbeck [31]. Briefly, biopsy samples were obtained from consenting adult patients undergoing esophagogastroduodenoscopy (IRB UCSD HRPP 150476) were minced in small pieces and treated with collagenase at 37°C for approximately 1h. Then cells were washed and filtered. Cultures were maintained in matrigel and medium containing Wnt3a, R-spondin and Noggin, which was refreshed or passaged every other day. For luminol experiments, wells were coated with 1/30 matrigel for 30 min, which was removed immediately before cells were added and for imaging, glass slides were coated with 10 μg/cm2 with Collagen IV for 1.5h at 37°C, and washed with warm PBS prior to the addition of cells.
Salmonella enterica serovar Typhimurium strain SL1344 and a ΔSPI2 mutant (kind gifts from Drs. Olivia Steel Mortimer NIAID, Rocky Mountain Laboratory, Montana, USA and Brett Finlay, University of British Columbia, Canada), were used at MOI 30 in cultures of T84 cells and HT-29 cells. Salmonella cultures were grown as described previously [32]. Briefly, a single colony was inoculated into LB broth and grown for 8h under aerobic conditions and then under oxygen-limiting conditions overnight.
Wild type APE1, an N-terminal acetylation mutant of APE1 (N-K6R/K7R), and a C-terminal DNA repair mutant of APE1 (C-H309N) constructs were used as previously reported [26]. Active Rac1 V12 and dominant negative Rac1 N17 plasmids were kind gifts from Dr. Jim Casanova University of Virginia, Charlottesville, Virginia, USA. All epithelial cells were seeded in six-well plates 18–24h before transfection. For overexpression studies, cells were transfected using 2 μg of plasmid DNA with Lipofectamine 2000 reagent (Invitrogen) as per the manufacturer’s protocol. In keeping with the manufacturer’s recommendation cells were used for infected experiments 40h post-transfection.
Nox1 expression was suppressed with human NOX1 siRNA ON-TARGETplus SMARTpool (Dharmacon RNAi technologies, L-010193-00-0005). AGS cells in 6 well plates were transfected using Lipofectamine RNAiMAX transfection reagent according to the protocol and luminol oxidation was measured after 48h.
Antibodies used include the following: anti-APE1, mouse monoclonal anti-Nox1 (Novus Biologicals), rabbit polyclonal anti-APE1, mouse monoclonal anti-Rac1 clone 28A (Millipore) followed by incubation with anti-rabbit or anti-mouse HRP-conjugated IgG (Cell Signaling Technology). NADPH oxidase inhibitor diphenyleneiodonium (DPI) and Rac inhibitor NSC23766 were purchased from Calbiochem.
ROS in AGS and T84 cells were measured according to the protocol described in Lumimax Superoxide Anion Detection Kit (Stratagene). See S1 Supplementary Materials and Methods for details. Measurements of ROS in NCI-N87, HT-29 and primary gastric epithelial cells were performed using 1 mM luminol (Sigma A8511, without additional enhancers) dissolved in borax buffer (pH 9) and the Spectramax L (Molecular Devices) reader for detection. For microscopic detection of ROS, cells were loaded with 5 μM CM-H2DCFDA (Invitrogen) for 30 min in an incubator (5% CO2 37°C). Following loading with CM-H2DCFDA cells were washed and infected.
Protein expression of APE1, Rac1 and Nox1 was assessed by western blot. Co-immunoprecipitation experiments were performed using anti-FLAG M2 agarose beads (Sigma) to analyze components that bind to FLAG-APE1 or FLAG-Rac1. See S1 Supplementary Materials and Methods for details. Densitometry was performed using ImageJ (National Institutes of Health). The levels of the protein of interest were corrected for the levels of the loading control (e.g. α-Tubulin).
Rac1 activity was measured as described previously [32] (see S1 Supplementary Materials and Methods). Densitometry was performed using ImageJ. The levels of active Rac1 were normalized for levels of total Rac1.
cDNAs obtained from antral gastric biopsies of H. pylori infected and uninfected patients were kindly provided by Richard Peek, Vanderbilt University (Tennessee, USA). Additionally, antral gastric mucosa biopsy specimens were collected from H. pylori-infected and uninfected individuals during diagnostic esophagogastroduodenoscopy following a University of Virginia Human Investigation Committee (HIC) (IRB number 9686) approved protocol into HBSS with 5% FBS [21]. All patient samples were de-identified apart from being known to be H. pylori infected or uninfected. The samples were analyzed at the University of Virginia, Virginia USA. See S1 Supplementary Materials and Methods for details.
APE1-Rac1 interactions were detected with Duolink PLA Kit (Olink Bioscience, Uppsala, Sweden: PLA probe anti-rabbit plus; PLA probe anti-mouse minus; Detection Kit orange) according to the manufacturer's protocol. See S1 Supplementary Materials and Methods for details. Biopsy specimens for immunohistochemistry were obtained with Institutional Review Board approval of the Pontifical Catholic University, Santiago, Chile (IRB number 12–236) from adult subjects with abdominal symptoms in Santiago, Chile. Samples were collected and H. pylori status was determined by rapid urease test and microscopic evaluation, and a study subject was judged colonized with H. pylori if one or both tests were positive for the bacteria. In collaboration with Dr. Harris, these snap frozen samples were shipped to UCSD where PLA was performed. Quantification of co-localization was performed using the colocalization plugin (JACoP) for ImageJ which calculates Pearson’s coefficient. ImageJ was used to quantify the amount of PLA signal, which was corrected for the number of cells present in each field of view.
Results are expressed as mean ± SEM. Statistical differences were calculated using ANOVA for multiple comparisons and Bonferroni post-hoc testing in Graphpad Prism. Levels of significance are indicated as follows: * p<0.05, ** p<0.01 and *** p< 0.001.
Proteins studied in this manuscript are given below with a reference to the SwissProt database:
APE1 (gene name APEX1), P27695
Rac1 (gene name RAC1), P63000
Nox1 (gene name NOX1), Q9Y5S8
We observed a rapid increase in superoxide production in the human gastric adenocarcinoma-derived cell line AGS following infection with H. pylori (Fig 1A). To determine whether the production of ROS observed was not unique to AGS cells additional experiments were performed in an alternative cancer-derived cell line NCI-N87 and non-transformed antral-derived primary epithelial cells. Induction of ROS production was also observed in NCI-N87 cells and primary human gastric epithelial cells isolated from the antrum (Fig 1B and 1C respectively) following infection with wild type H. pylori strain 26695 although the kinetics where somewhat different from AGS cells. Superoxide generation by luminol oxidation was independent of the vacA and cagA pathogenicity island (PAI) status of H. pylori since no significant differences were seen when AGS cells were infected with wild type H. pylori or the vacA or cagA PAI mutant strain, 8–1 (S1 Fig). Prolonged infection studies showed that ROS is generated early following infection and not observed at 4h of infection or later (S2 Fig).
It is known that H. pylori activates Rac1, and another report shows that Rac1 activation initiates ROS production in guinea pig gastric cells [13,33]. To determine if Rac1 regulates H. pylori-mediated ROS generation in human gastric epithelial cells, a constitutively active Rac1 plasmid (V12) was overexpressed in AGS cells or cells were treated with the Rac1-specific inhibitor NSC23766, before H. pylori infection. Overexpression of active Rac1 resulted in increased ROS generation, while the Rac1 inhibitor reduced ROS generation compared to vector-transfected cells (Fig 1D). To confirm Rac1 activation during H. pylori infection, active Rac1 was assessed using a pulldown assay. As shown in Fig 1E and 1F, H. pylori 26695 infection increased Rac1 activation in AGS cells in 30 min and in NCI-N87 cells at 60 min after infection. Since there was no difference in ROS generation or Rac1 activation by H. pylori strains 26695 and 8–1 (S3 Fig), subsequent experiments were performed with H. pylori 26695 only. Although activation of Rac1 by H. pylori has been previously reported, here we expand this finding by showing that Rac1 is involved in the production of ROS by gastric epithelial cells following infection with H. pylori.
After establishing that H. pylori induce ROS production by gastric epithelial cells through activation of Rac1, we investigated whether the ROS were generated by Nox1 as a major NADPH oxidase expressed in gastric epithelial cells [4]. Our results demonstrate that AGS cells infected in the presence of the general ROS inhibitor N-acetyl-L-cysteine (NAC), showed significant inhibition of superoxide production (Fig 2A). Also, infection in the presence of the NADPH oxidase inhibitor diphenyleneiodonium (DPI) resulted in inhibition of superoxide production, suggesting that NADPH oxidase is involved in H. pylori-induced ROS generation. As DPI is not a specific inhibitor of Nox1, we used siRNA-mediated suppression of Nox1 to show a comparable decrease in luminol oxidation following infection with H. pylori ROS (Fig 2B). To evaluate the relative contributions of Nox1 and Rac1 in H. pylori-induced ROS generation, luminol oxidation was measured in AGS cells in the presence of DPI, the Rac1 inhibitor NSC23766 or overexpression of active Rac1. Comparable inhibition of ROS generation was observed when NSC23766 or DPI was used alone or in combination, indicating that Rac1 and NADPH oxidase share the same pathway to generate ROS. The increase in ROS in the presence of V12 was abrogated by DPI suggesting that Rac1 activation alone is not sufficient to generate ROS when NADPH oxidase activity is inhibited (Fig 2C). In a parallel experiment, Nox1 protein expression was increased in AGS cells within 1h of H. pylori infection. This induction was further enhanced in the presence of active Rac1 but decreased in the presence of the Rac1 inhibitor NSC23766 (Fig 2D). Together, our data demonstrate that NOX1 is the major source of ROS in gastric epithelial cells infected with H. pylori.
It is known that ROS induces APE1, but whether APE1 modulates ROS generation has not been previously examined. As illustrated in Fig 3A, luminol oxidation was increased in APE1 suppressed cells indicating regulation of ROS by APE1. Corroborating the findings with luminol, immunofluorescence with CM-H2DCFDA demonstrated increased ROS generation in APE1 suppressed cells following infection (Fig 3E). The additional increase of ROS in APE1 suppressed cells was absent in the presence of NSC23766 or DPI suggesting that both Rac1 and NADPH oxidase act downstream of APE1 in the pathway of ROS generation (Fig 3B). Furthermore, overexpression of exogenous APE1 in cells with suppressed endogenous APE1 significantly reduced H. pylori-induced ROS generation (Fig 3C). APE1 overexpression was also sufficient to inhibit ROS generation in the presence of V12 overexpression implicating APE1 as a major regulator of Rac1-mediated oxidative stress (Fig 3D).
To address if APE1 directly regulates Rac1, Rac1 activity was compared in vector control and APE1 suppressed cells. Fig 4A and 4B demonstrate a significant increase in active Rac1 in APE1 suppressed AGS or NCI-N87 cells within 60 min of infection. Overexpression of exogenous APE1 in APE1 suppressed AGS cells resulted in a decrease in Rac1 activity (Fig 4C). To establish whether APE1 binds to Rac1 to inhibit its activity, we immunoprecipitated APE1 and demonstrated that Rac1 interacted with APE1. This interaction was augmented within 30 min of H. pylori infection (Fig 4D). The enhanced association between APE1 and Rac1 after H. pylori infection was further confirmed by confocal microscopy showing co-localization of Rac1 and APE1 staining in the cytosol as indicated in the merged image (Fig 4E). Using in situ proximity ligation assay (PLA) we confirmed cytosolic co-localization of APE1 and Rac1 following H. pylori infection in AGS, NCI-N87 and antral-derived primary gastric cells (Fig 4F). To demonstrate that the findings in cell lines also occur in native human gastric epithelial cells we performed PLA in primary gastric epithelial cells from gastric mucosal biopsy samples (Fig 4G). Our experiments showed that the APE1-Rac1 interaction was greater in biopsy samples from patients infected with H. pylori compared to those from uninfected control subjects (Fig 4F right panel). Moreover, by performing Co-IP experiments we observed that APE1 interacted with the constitutively active form of Rac1 (V12) but not with the dominant negative form (N17) (S4 Fig). From these observations we conclude that APE1 negatively regulates activation of Rac1.
To examine the effect of the level of APE1 on the previously reported increase of Nox1 after H. pylori infection [13], levels of Nox1 were assessed by western blot in AGS cells with varying APE1 levels after infection at various times. Increased levels of Nox1 were observed in the APE1 suppressed cells compared to the vector control cells within 1h of H. pylori infection (Fig 5A). This was confirmed by immunofluorescence staining that showed increased Nox1 after infection in APE1 suppressed cells compared to controls (Fig 5B). To determine if the observations found in cell lines could be translated to native human gastric epithelial cells, real time RT-PCR for Nox1 and APE1 were performed with the total RNA isolated from gastric antral biopsies from uninfected or H. pylori infected patients. The expression of Nox1 and APE1 was significantly increased in tissue from infected patients (Fig 5C). These in vivo data suggest a role for Nox1 and APE1 in the response to infection of human stomach with H. pylori.
Earlier we established that various regulatory functions of APE1 are largely regulated by its N-terminal lysines (K6K7) and C-terminal histidine (H309) [26]. Therefore, co-immunoprecipitation was performed in AGS cells to determine the binding of Rac1 with the acetylation mutant (N-K6R/K7R) and the DNA repair mutant (C-H309N) of APE1. Our results showed that the N-terminal acetylation mutant of APE1 had minimal binding with Rac1 whereas the binding of the DNA repair mutant was comparable to that of WT APE1 (Fig 6A). To establish if this interaction between APE1 and Rac1 is essential in regulating H. pylori-induced ROS generation, ROS were measured in APE1 suppressed AGS cells that were transfected with WT APE1, N-terminal mutant or C-terminal mutant and then infected with H. pylori. We observed a greater than 2-fold ROS increase in the N-terminal mutant overexpressing cells compared to WT APE1. Although overexpression of the C-terminal mutant also showed increased ROS generation compared to WT APE1, this was significantly less than the N-terminal mutant (Fig 6B). To determine if Rac1 activity is modulated by the non-acetylatable mutant of APE1, APE1 suppressed AGS cells were similarly transfected as described in Fig 6B, and Rac1 activation was measured after 30 min of H. pylori infection. Analogous to the findings of ROS generation, we observed that the non-acetylatable mutant was unable to inhibit Rac1 activation compared to WT APE1 or the DNA repair mutant of APE1 (Fig 6C).
To determine if the suppression of ROS production by APE1 occurs in other epithelial cells within the gastrointestinal tract and with other infections, we generated stable APE1 suppressed human colonic epithelial T84 cells and compared responses to wild type Salmonella SL1344 and the Salmonella ΔSPI2 mutant. The ΔSPI2 mutant of Salmonella was used for the ability of the pathogenicity island 2 of Salmonella to inhibit ROS production in phagocytes [34]. For both HT-29 and T84 colonic epithelial cells, we found that infection with ΔSPI2 mutant of Salmonella generated ROS that was further increased in corresponding APE1 suppressed cells (Fig 7A and 7B). Compared to the ΔSPI2 mutant, limited amounts of ROS were induced by wild type Salmonella. Also, immunofluorescence with CM-H2DCFDA demonstrated increased ROS generation in APE1 suppressed T84 cells following ΔSPI2 mutant infection (Fig 7C). This finding suggests that in addition to interfering with Nox2 in macrophages, Salmonella may also interfere with the Nox1 complex in intestinal epithelial cells.
In this study we show that APE1 regulates the induction of reactive oxygen species (ROS) by gastroenteric pathogens in a panel of relevant human gastrointestinal epithelial cells. Multifunctional APE1 was demonstrated to inhibit Nox1-mediated ROS production through its direct interactions with Rac1. In addition to preventing formation of the functional NADPH complex, APE1 limits ROS production by decreasing Nox1 expression. Together, these data support the concept that through its molecular interactions with Rac1, APE1 provides negative feedback on Nox1 and oxidative responses in the gastrointestinal epithelium during bacterial infection. These data implicate APE1 in novel molecular interactions that regulate early stress responses elicited by microbial infections.
Microbial pathogens affect host cells through the generation of various radicals [3,35,36]. For example, we and others have demonstrated that H. pylori infection stimulates the accumulation of intracellular ROS in human gastric epithelial cell lines and freshly isolated native human gastric epithelial cells [37,38]. The potential roles of VacA and CagA in regulating ROS production in cells are also illustrated by other reports showing VacA-dependent regulation of autophagy and associated ROS production [39,40]. In our studies H. pylori lacking VacA had no significant effect on ROS production as assessed by luminol oxidation. Although CagA has been implicated in increased levels of ROS, the 8–1 mutant lacking CagA did not significantly alter ROS production in our assays [41]. Since dyes that detect ROS species have varying sensitivities and detect ROS in intracellular or extracellular compartments the role of VacA or CagA in the generation of ROS was not conclusively demonstrated in our studies [38,41]. Commensal bacteria that reside in the gut are reported to induce ROS generation from intestinal epithelial cells [42]. High levels of ROS are associated with molecular damage to cellular components and consequent tissue injury but APE1 may represent an important host factor to limit this damage. The differences in the kinetics of ROS generation in the various cell lines employed in this study could be resolved in future studies in animal models. This is particularly relevant to model the persistent infection of humans with H. pylori.
Advancing our prior observations showing that H. pylori-induced apoptosis is inhibited by APE1 [26], the present work establishes a novel role of APE1, mediating the inhibition of oxidative stress. This function of APE1 may contribute to its ability to inhibit oxidative stress-induced cell death as well as a fine-tuning of the redox-sensitive responses induced during infection [43]. Although APE1 is referred to as a stress response molecule [44], concordant with a recent report showing the regulation of stress by APE1 in the mitochondria of neuronal cells [45] our work demonstrates its role in regulating stress generation in gastrointestinal epithelial cells.
To understand the mechanism of APE1 as a determinant of ROS regulation, we focused on Rac1 and Nox1, two major contributors of ROS generation in non-phagocytic cells. The small GTPases, Rac1 and Rac2, are common mediators of NADPH-dependent ROS production in diverse signaling pathways that lead to mitogenesis, gene expression and stress responses [18,46]. Our findings corroborate the dependence of Rac1 on ROS production as we show that H. pylori-induced ROS generation is downregulated by the APE1-Rac1 interaction that subsequently inhibits Nox1. Further characterization of the molecular association between active Rac1, cellular ROS levels and APE1 provides new mechanistic insight into the control of redox-sensitive host responses with potential relevance to the development of novel therapies for gastrointestinal infections and associated inflammation.
As Rac1 is an integral part of the functional NADPH oxidase complex [14], inhibition of Rac1 activity by APE1 is expected to interfere with this assembly, thereby providing negative feedback on ROS generation. Regulation of Rac1 by APE1 was observed in AGS and NCI-N87 cells, however, the kinetics of the regulation of Rac1 and APE1 were different. Although those kinetics varied somewhat, intracellular co-localization of APE1 and Rac1 following infection assessed by using the proximity ligation assay showed a significant increase in both AGS and NCI-N87 at 1h after infection. This co-localization of APE1 and Rac1 was also observed in antrum-derived primary gastric epithelial cells. Overexpression of APE1 decreased ROS comparable with the effect of DPI or of NSC23766 (Fig 3C and 3D), underlining that APE1 is a major regulator of the Rac1-NADPH oxidase axis of ROS production. In addition to Rac1 inhibition, we identified another level of inhibition by APE1 when APE1 suppressed cells were found to express significantly more Nox1 compared to the vector control cells. The observation of an augmentation of H. pylori-induced ROS generation in two different APE1 suppressed gastric epithelial cells supports a broadly relevant role for APE1 in regulating ROS. Interestingly, APE1 and the phytochemical Ginko biloba both regulated mitochondrial oxidative stress in neuronal cells [45]. APE1 and phytochemical-mediated regulation of mitochondrial oxidative stress could also be of relevance in Helicobacter-induced ROS generation in gastric epithelial cells.
Given APE1’s multiple functions, it is not surprising that interacting molecular partners of APE1 have already been identified. It appears likely that acetylation-mediated conformational changes in APE1's N-terminal domain modulate its interaction with partner proteins, including Rac1 [47]. We have not manipulated the various redox-responsive cysteine residues of APE1 in our study. As various reports show a role for the redox function of APE1 in regulating responses to cell stress, the redox function of APE1 may also be involved in cellular responses to oxidative stress. Unlike the stable interaction between APE1 and Rac1, the minimal association between Rac1 and the N-terminal acetylation mutant of APE1 underscores the necessity of the Lys residues for the interaction. Our data indicate that this interaction is essential for the ability of APE1 to inhibit the production of ROS since significantly increased ROS generation was found with the non-interacting acetylation mutant compared to WT APE1 (Fig 6B). Taken together with our previous observation that H. pylori induced APE1 acetylation [21], this finding highlights a previously unrecognized modification of regulatory molecules during infection. We speculate that the role of APE1 could be similar to the Rho-GDP dissociation inhibitors (Rho-GDI), which translocates Rac1 from the membrane to the cytoplasm, effectively deactivating NADPH oxidase [48,49].
Our observation of the inhibition of Rac1 by APE1 in intestinal cell lines indicates that APE1-regulated ROS generation is conserved between gastric and intestinal epithelial cells. These data suggest a common role for APE1 in the pathogenesis of various prolonged gastrointestinal bacterial infections. Unlike the robust ROS generation typically induced by acute infection, lower levels of ROS produced by host epithelial cells are increasingly recognized to play a critical physiological role [18] including regulation of the molecular machinery of epithelial secretory lineages and autophagy [50]. As such, redox signaling through Nox1 represents a unique intracellular regulator of diverse signaling pathways involved in normal cell physiology, inflammation and carcinogenesis. Due to the nature of in vitro infection models, including uncontrolled bacterial growth and related cell stress-induced mitochondrial ROS production in cell models, future experiments in vivo are needed to determine the physiological importance of acute versus chronic infections with H. pylori in relation to regulation of oxidative stress by APE1.
In summary, we have shown that APE1 controls the regulation of epithelial responses to gastroenteric infections and the subsequent generation of oxidative stress. Our findings provide new insights into APE1’s role as a host molecule that modulates ROS generation via negative regulation of Rac1 and Nox1. Our future studies will aim to examine models of prolonged infection and the physiological responses to infections.
|
10.1371/journal.pmed.1002482 | Brain and blood metabolite signatures of pathology and progression in Alzheimer disease: A targeted metabolomics study | The metabolic basis of Alzheimer disease (AD) is poorly understood, and the relationships between systemic abnormalities in metabolism and AD pathogenesis are unclear. Understanding how global perturbations in metabolism are related to severity of AD neuropathology and the eventual expression of AD symptoms in at-risk individuals is critical to developing effective disease-modifying treatments. In this study, we undertook parallel metabolomics analyses in both the brain and blood to identify systemic correlates of neuropathology and their associations with prodromal and preclinical measures of AD progression.
Quantitative and targeted metabolomics (Biocrates AbsoluteIDQ [identification and quantification] p180) assays were performed on brain tissue samples from the autopsy cohort of the Baltimore Longitudinal Study of Aging (BLSA) (N = 44, mean age = 81.33, % female = 36.36) from AD (N = 15), control (CN; N = 14), and “asymptomatic Alzheimer’s disease” (ASYMAD, i.e., individuals with significant AD pathology but no cognitive impairment during life; N = 15) participants. Using machine-learning methods, we identified a panel of 26 metabolites from two main classes—sphingolipids and glycerophospholipids—that discriminated AD and CN samples with accuracy, sensitivity, and specificity of 83.33%, 86.67%, and 80%, respectively. We then assayed these 26 metabolites in serum samples from two well-characterized longitudinal cohorts representing prodromal (Alzheimer’s Disease Neuroimaging Initiative [ADNI], N = 767, mean age = 75.19, % female = 42.63) and preclinical (BLSA) (N = 207, mean age = 78.68, % female = 42.63) AD, in which we tested their associations with magnetic resonance imaging (MRI) measures of AD-related brain atrophy, cerebrospinal fluid (CSF) biomarkers of AD pathology, risk of conversion to incident AD, and trajectories of cognitive performance. We developed an integrated blood and brain endophenotype score that summarized the relative importance of each metabolite to severity of AD pathology and disease progression (Endophenotype Association Score in Early Alzheimer’s Disease [EASE-AD]). Finally, we mapped the main metabolite classes emerging from our analyses to key biological pathways implicated in AD pathogenesis. We found that distinct sphingolipid species including sphingomyelin (SM) with acyl residue sums C16:0, C18:1, and C16:1 (SM C16:0, SM C18:1, SM C16:1) and hydroxysphingomyelin with acyl residue sum C14:1 (SM (OH) C14:1) were consistently associated with severity of AD pathology at autopsy and AD progression across prodromal and preclinical stages. Higher log-transformed blood concentrations of all four sphingolipids in cognitively normal individuals were significantly associated with increased risk of future conversion to incident AD: SM C16:0 (hazard ratio [HR] = 4.430, 95% confidence interval [CI] = 1.703–11.520, p = 0.002), SM C16:1 (HR = 3.455, 95% CI = 1.516–7.873, p = 0.003), SM (OH) C14:1 (HR = 3.539, 95% CI = 1.373–9.122, p = 0.009), and SM C18:1 (HR = 2.255, 95% CI = 1.047–4.855, p = 0.038). The sphingolipid species identified map to several biologically relevant pathways implicated in AD, including tau phosphorylation, amyloid-β (Aβ) metabolism, calcium homeostasis, acetylcholine biosynthesis, and apoptosis. Our study has limitations: the relatively small number of brain tissue samples may have limited our power to detect significant associations, control for heterogeneity between groups, and replicate our findings in independent, autopsy-derived brain samples.
We present a novel framework to identify biologically relevant brain and blood metabolites associated with disease pathology and progression during the prodromal and preclinical stages of AD. Our results show that perturbations in sphingolipid metabolism are consistently associated with endophenotypes across preclinical and prodromal AD, as well as with AD pathology at autopsy. Sphingolipids may be biologically relevant biomarkers for the early detection of AD, and correcting perturbations in sphingolipid metabolism may be a plausible and novel therapeutic strategy in AD.
| Metabolomics, which measures the biochemical products of cell processes, can be used to measure alterations in biochemical pathways related to AD.
Several recent studies have applied metabolomics to explore potential blood biomarkers for Alzheimer disease (AD).
Prior blood biomarker studies have not linked signals in the blood to those in the brain and have relied mainly on discriminating between AD/mild cognitive impairment (MCI) and control samples.
These study designs ignore the long preclinical prodrome of AD and do not provide biological insights into the evolution of AD pathology in the brain and eventual development of clinical symptoms.
Our study was designed to link alterations in metabolite signals in the brain to those in the blood, explore how those alterations were associated with distinct endophenotypes of AD, and identify the key biological pathways implicated.
We used quantitative and targeted metabolomics assays on brain tissue samples (N = 44) and machine-learning methods to identify a brain metabolite signature of AD, i.e., a 26-metabolite panel that discriminated AD and control samples with accuracy, sensitivity, and specificity of 83.33%, 86.67%, and 80%, respectively.
We then assayed the same 26 metabolites in blood from two longitudinal cohorts that represent prodromal (Alzheimer’s Disease Neuroimaging Initiative [ADNI], N = 767) and preclinical (Baltimore Longitudinal Study of Aging [BLSA], N = 207) AD and tested their associations with MRI measures, CSF biomarkers, risk of conversion to incident AD, and cognitive performance.
We found that higher blood concentrations of sphingolipid species were consistently associated with severity of AD pathology at autopsy and AD progression across prodromal and preclinical stages.
These metabolites map to several biologically relevant pathways in AD, including tau phosphorylation, Aβ metabolism, calcium homeostasis, acetylcholine biosynthesis, and apoptosis.
Our study design represents a novel approach for identifying markers of disease progression in AD and potential avenues for therapeutic intervention.
Perturbations in sphingolipid metabolism are consistently associated with preclinical and prodromal AD, as well as with AD pathology at autopsy, providing compelling evidence for their significant role in AD pathogenesis.
| The relationships between systemic abnormalities in metabolism and the pathogenesis of Alzheimer disease (AD) are poorly understood. It is unclear how global perturbations in metabolism are related to severity of AD pathology and the eventual expression of AD symptoms in at-risk individuals. Understanding the metabolic basis of AD and its impact on disease progression during the early, preclinical, and prodromal stages is likely to provide insights into novel disease-modifying treatments for this irreversible, progressive neurodegenerative disorder.
Metabolomics, which measures the biochemical products of cell processes downstream of genomic, transcriptomic, and proteomic systems, has generated excitement because of its potential to capture snapshots of the complex and multifactorial biochemical pathways that may be altered in AD [1,2]. These include changes across multiple physiological pathways driven by the complex interactions between behavioral, genetic, and environmental risk factors. Recent studies have applied metabolomics to examine alterations in blood metabolite profiles in AD; such studies have the potential to both discover peripheral biomarkers as well as identify key metabolic pathways intrinsic to AD pathogenesis [3–7]. However, one of the key challenges in these metabolomics studies is the inability to link alterations in metabolite signals in the blood to those in the brain. It is therefore difficult to assess whether a peripheral signal associated with disease status is also reflected in the brain, where accumulation of distinct pathological features in specific regions is believed to trigger symptom onset. As is common with late-onset and gradually progressive diseases, there are many alterations in cell processes due to chronic comorbid medical conditions that may be reflected in peripheral blood metabolite concentrations. Additionally, traditional blood biomarker studies have relied mainly on the binary discrimination of established AD/mild cognitive impairment (MCI) from control (CN) samples. This study design ignores the long preclinical prodrome of AD, when brain pathology is accumulating but has not yet triggered the onset of cognitive impairment and functional decline in individuals eventually diagnosed with AD. As we have proposed previously [8], alternative study designs in biomarker analyses, in which the primary end points are well-established endophenotypes of AD pathology rather than binary discrimination of case versus control, offer the potential to identify biologically relevant blood biomarkers for AD.
Here, we describe a four-step approach to the discovery of brain and blood metabolites associated with pathology and progression of AD (Fig 1). (1) Identifying a brain metabolite signature of AD: in this phase of the study, we first used quantitative and targeted metabolomics to identify a panel of metabolites that accurately differentiated brain tissue samples from neuropathologically confirmed AD and CN subjects. (2) Testing blood metabolite associations with AD endophenotypes: we then tested whether serum concentrations of the same metabolites in two independent samples representing preclinical AD and prodromal AD were associated with distinct clinical, cognitive, neuroimaging, and cerebrospinal fluid (CSF) endophenotypes of AD. (3) Summarizing results: we developed an integrated blood and brain endophenotype score (Endophenotype Association Score in Early Alzheimer’s Disease [EASE-AD]) summarizing the relative importance of specific brain and blood metabolites to severity of AD pathology and disease progression. (4) Mapping biological pathways: we finally mapped the main metabolite classes emerging from these analyses to key biological pathways implicated in AD pathogenesis to understand the potential roles of these molecules and their interactions in triggering symptom onset and progression of AD.
The Baltimore Longitudinal Study of Aging (BLSA) is a prospective cohort study of community-dwelling participants that began in 1958 [9,10]. Detailed clinical and cognitive evaluations, including neurological, laboratory, and radiological evaluations, were conducted every 2 years. Since 2003, participants older than 80 years received yearly assessments. The autopsy subsample used in Step 1 (i.e., Identifying a brain metabolite signature) to generate the brain metabolite signature of AD included 44 participants (N = 15 AD; N = 14 CN; N = 15 asymptomatic Alzheimer’s disease [ASYMAD], described below). For Step 2 (i.e., Testing blood metabolite associations with AD endophenotypes), metabolomic analyses in serum samples were performed on 207 BLSA (exclusion criteria described below) participants divided into “converters” and “non-converters.” Converters were defined as participants who were cognitively normal at the initial blood draw and developed incident AD based on consensus clinical diagnosis (described below) during follow-up approximately 5 years later. These participants were age and sex matched to non-converters and defined as participants who remained cognitively normal over a similar follow-up interval. Initial serum samples were collected while both groups were cognitively normal; we therefore characterize the converters in this sample as representing “preclinical AD.” Demographic characteristics of the autopsy sample and blood study sample in the BLSA are included in Table 1. Written informed consent was obtained at each visit; the BLSA study protocol has ongoing approval with the institutional review board of the National Institute of Environmental Health Science (NIEHS), National Institutes of Health.
As described below, in addition to identifying a brain metabolite signature of AD in Step 1, the BLSA sample was used in Step 2 to test associations between blood metabolite concentrations and the following AD endophenotypes: (1) differences by diagnoses (i.e., converters versus non-converters), (2) risk of conversion to incident AD, and (3) trajectories of cognitive performance prior to onset of AD symptoms.
The Alzheimer’s Disease Neuroimaging Initiative (ADNI) sample was also used in Step 2 (i.e., Testing blood metabolite associations with AD endophenotypes). ADNI is an ongoing, longitudinal study launched in 2003 as a public–private partnership, led by principal investigator Michael W. Weiner, MD. The primary goal of ADNI has been to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessments can be combined to measure the progression of mild cognitive impairment (MCI) and early AD. Details on study design, participant recruitment, study approval, and informed consent procedures have been published previously [11]. The study was approved by the institutional review boards of all of the participating institutions/study sites. The full list of participating institutions is included in S1 Table. Informed written consent was obtained from all participants at each site. Metabolomics data for ADNI samples were generated by the Alzheimer Disease Metabolomics Consortium and (ADMC) deposited to LONI. The mission of the ADMC is to create a comprehensive metabolomics database for AD. ADNI data used in the preparation of this article were also obtained from the ADNI-1 database (adni.loni.usc.edu) and included baseline blood serum metabolite concentrations (with concurrent structural MRI data) on 767 participants and concurrent CSF AD biomarker data on 403 participants. ADNI was enriched with participants with MCI and therefore represents “prodromal AD” (participants with MCI at baseline who subsequently converted back to normal cognition were excluded). Demographic characteristics of the ADNI sample are included in Table 1.
As described below, the ADNI sample was used in Step 2 to test associations between blood metabolite concentrations and the following AD endophenotypes: (1) MRI measures of AD-related brain atrophy, (2) CSF measures of AD pathology, and (3) risk of conversion to incident AD.
The autopsy program of the BLSA was initiated in 1986 and has been described previously [12]. The autopsy subsample is not significantly different from the BLSA cohort as a whole in terms of the rates of dementia and clinical stroke [13]. Postmortem brain examinations were performed by an experienced neuropathologist. Assessment of neuritic plaques and neurofibrillary tangles using Consortium to Establish a Registry for Alzheimer's Disease (CERAD) [14] and Braak criteria [15], respectively, have been described previously [16]. We have previously described the clinico-pathological features of BLSA participants categorized as ASYMAD after neuropathological assessment at death [17]. Briefly, these individuals had significant AD neuropathology at autopsy but were found to be cognitively intact, as assessed by longitudinal neuropsychological assessments, within 1 year prior to death.
In BLSA, cognitive status was considered at consensus diagnosis conferences after each assessment/visit, using established procedures described previously [18]. The consensus conferences included neurologists, neuropsychologists, and neuroimaging scientists. At each assessment, participants underwent a battery of neuropsychological testing. Clinical and neuropsychological data were reviewed at multidisciplinary consensus case conferences if they made four or more errors on the Blessed Information, Memory, and Concentration (BIMC) test, if their Clinical Dementia Rating (CDR) score was equal to or greater than 0.5, or if concerns were raised about their cognitive status by a reliable informant. In addition, all participants were evaluated by case conference on death or withdrawal. It is also important to note that longitudinal data reviewed during consensus case conferences include (besides detailed cognitive assessments) medication history, self-reported diagnoses of comorbid medical conditions, neuroimaging data, as well as laboratory evaluation for reversible causes of cognitive impairment such as serum TSH and B12 levels. The diagnoses of dementia and AD were based on the Diagnostic and Statistical Manual (DSM)-III-R [19] and the National Institute of Neurological and Communication Disorders and Stroke–Alzheimer’s Disease and Related Disorders Association (NINCDS-ADRDA) criteria, respectively [20].
For individuals diagnosed with AD, age at onset of initial symptoms of AD was estimated at consensus case conferences using longitudinal cognitive performance data as well as informant-based history.
In ADNI, dementia diagnosis was determined based on NINCDS-ADRDA criteria for probable AD. MCI participants met criteria for amnestic MCI [21], and CN participants were cognitively normal. Additional details on the ADNI protocol are available at http://www.adni-info.org.
Blood serum samples were collected from BLSA participants at the NIA Clinical Research Unit in Harbor Hospital, Baltimore. Details on collection and processing have been published previously [7]. Briefly, venous blood samples were collected between 6 and 7 AM following an overnight fast. Serum samples were aliquoted into 0.5-mL volume in Nunc cryogenic tubes and stored at −80°C until further use. Samples were not subject to any freeze–thaw cycles prior to metabolomic assays. Additional details on sample selection have been published previously [7]. The average storage time of serum samples in BLSA participants was 17.84 years (SD: 6.45) in converters and 13.28 years (SD: 5.98) in non-converters (Table 1). In order to minimize potential effects of long storage times on serum sample stability and metabolite concentrations, we excluded all samples (N = 9 non-converters; N = 34 converters) with methionine sulfoxide (Met-So) concentrations greater than 5 μM (3 SD above average) [22,23]. The original sample included 250 participants; after excluding samples with high Met-So concentration, the final sample included 207 participants (N = 115 non-converters; N = 92 converters).
Details on collection and processing of ADNI blood serum samples have been published previously (http://adni.loni.usc.edu/wp-content/uploads/2010/11/BC_Plasma_Proteomics_Data_Primer.pdf). Briefly, blood was collected at 8 AM prior to CSF collection after an overnight fast, immediately placed on dry ice, and shipped on the same day to the ADNI Biomarker Core at the University of Pennsylvania for processing. The final sample included 767 participants (N = 216 normal; N = 366 MCI; N = 185 AD). All samples had Met-So concentrations below 5 μM and no samples were excluded.
Quantitative metabolomics was performed on BLSA brain and BLSA and ADNI blood samples on the Biocrates AbsoluteIDQ p180 platform. This commercially available platform allows for the quantification of amino acids, acylcarnitines, sphingomyelins (SMs), phosphatidylcholines (PCs), hexoses (h1s), and biogenic amines. Details on the assays have been published previously [24]. Briefly, the validated assay uses two different mass spectrometric methods with isotope labeled and other internal standards for absolute quantification of metabolites. The acylcarnitines, lipids, and h1s are analyzed by flow injection analysis-tandem mass spectrometry (FIA-MS/MS). The amino acids and biogenic amines are derivatized using phenylisothiocyanate and analyzed by liquid chromatography tandem-mass spectrometry (HPLC-MS/MS) using an AB SCIEX 4000 QTrap mass spectrometer (AB SCIEX, Darmstadt, Germany) with electrospray ionization. Concentration of each metabolite was measured in μM.
For brain tissue metabolomics, regions were selected a priori in the middle frontal gyrus (MFG), inferior temporal gyrus (ITG), and cerebellum (CBL). The MFG and ITG were sampled to represent brain regions vulnerable to amyloid β and tau deposition, respectively; the CBL was sampled to represent a brain region resistant to classical AD pathology. A sterile 4-mm-diameter tissue punch was extracted from the cortical surface of the brain tissue regions, which were stored at −80°C. To extract metabolites, samples were homogenized using Precellys with ethanol phosphate buffer; samples were then centrifuged, and the supernatant was used for analysis. Metabolite concentrations in brain tissue samples indicated as less than the limit of detection (LOD) were imputed as the highest value below the LOD. This method removed all differences below the LOD but still allowed machine-learning classifiers to pick up any differences in metabolite concentrations between those above the LOD and those below.
For blood metabolomics, in BLSA, converter and non-converter samples were randomly divided into 6 batches. Each batch was processed in separate runs with technicians blinded to diagnostic status. Additional data processing and checking steps, including reproducibility and testing for equality of coefficient of variance across metabolites, has been described in detail previously [25]. BLSA serum samples indicated as less than LOD were not imputed due to minimal missingness; 25/26 metabolites had 0 < LOD values. ADNI data processing has been described in detail previously and included imputing values indicated as less than the LOD as the metabolite LOD/2, as determined by the ADMC [25]. Metabolite concentrations from participants with duplicate measurements were averaged in all analyses.
Batch effects were controlled for using a set of CN samples. Standardized quality control (QC) material, i.e., commercially available pooled human plasma spiked with a defined set of metabolites, was used across all batches to control and adjust for batch effects by applying MetIDQ software-implemented normalization procedures.
Cognitive performance was analyzed from assessments administered to BLSA participants every two years. Memory was assessed using the California Verbal Learning Test (CVLT), including learning (total recall over 5 learning trials), immediate free recall, and long delay free recall. Attention was assessed using the Trails Making Test Part A and the WAIS-R Digits Forward test. Executive function was measured using the Trails Making Test Part B and the WAIS-R Digit Backward test. Language was measured using letter fluency and semantic fluency tests. Visuo-spatial ability was measured using the Clock Drawing Test and the Card Rotation Test.
MRI protocol, including scanner specifications, image acquisition, and image processing, are described in detail at www.adni-info.org. Briefly, protocol specifications included T1 weighted MR images, including sagittal volumetric 3D MPRAGE with 1.25 × 1.25-mm in-plane spatial resolution, 1.2-mm thick sagittal slices, 8° flip angle, and target TR of about 8.9 mm and TE of about 3.9 ms [26]. We utilized the Spatial Patterns of Abnormality for Recognition of Early Alzheimer’s disease (SPARE-AD) index [27] as a neuroimaging measure of “AD-like” brain atrophy patterns [28]; this measure was calculated for baseline visits of ADNI-1 participants.
Participants underwent lumbar puncture in the morning following overnight fasting and blood draws. Samples were immediately placed on dry ice and shipped to the ADNI Biomarker Core for processing. Total tau (t-tau), phosphorylated tau (p-tau), and amyloid beta 1–42 (Aβ1–42) were measured using the multiplex xMAP Luminex platform (Luminex Corp, Austin, TX) with Innogenetics (INNO-BIA AlzBio3, Ghent, Belgium) immunoassay kit-based reagents. See http://adni.loni.usc.edu/wp-content/uploads/2012/01/2011Dec28-Biomarkers-Consortium-Data-Primer-FINAL1.pdf for additional details on sample collection and processing, including reproducibility and data quality checks. CSF samples were available in 395 participants (N = 109 normal; N = 186 MCI; N = 100 AD).
The first two stages of the analytic plan used for BLSA, including Step 1, Identifying a brain metabolite signature of AD, and Step 2, Testing blood metabolite associations with AD endophenotypes, were developed conceptually in May 2016 prior to any data analyses. There were no subsequent data-driven alterations to this conceptual analytic plan; final data visualization for Step 3, Summarizing results, was based on various iterations during analyses. The inclusion of ADNI data occurred in fall 2016 after our data request was approved by the ADMC. Step 4, Mapping biological pathways, occurred after we identified the principal classes of metabolites emerging from Steps 1–3. Sensitivity analyses testing blood metabolite associations with AD endophenotypes in BLSA (indicated below in Step 2) were conducted at the request of reviewers.
The demographic characteristics of BLSA participants in the autopsy cohort whose brain tissue samples were used in the metabolomics assays are included in Table 1. The mean age at death in the autopsy sample was 81.33 years (SD: 10.19), and the mean interval between last evaluation and death (postmortem interval) was 14.93 h (SD: 6.86). Participants in the three groups—CN, ASYMAD, and AD—did not significantly vary by age at death, sex, or postmortem interval. The demographic characteristics of BLSA participants who provided blood data are included in Table 1. Participants were aged 78.47 years (SD: 6.96) at initial blood draw and 51.69% were female. Converter and non-converter groups did not vary by age or sex. Serum samples in the converter group were, on average, stored for four years longer than non-converter samples (17.84 years [SD: 6.45] versus 13.28 years [SD: 5.98]; p < 0.05).
The demographic characteristics of ADNI participants are included in Table 1. Participants were aged 75.19 years (SD: 6.82) at baseline, and 42.63% were female. MCI participants were significantly younger (74.69 years [SD: 7.35]) and had fewer females (36.07%). Samples did not vary by storage time.
Higher blood concentrations of sphingolipids were broadly associated with greater AD-like brain atrophy patterns and more AD-like CSF levels of pathology. Specifically, two sphingolipids, SM C16:0 (β = 0.593, 95% CI = 0.147–1.040, p = 0.009) and SM C18:1 (β = 0.466, 95% CI = 0.0687–0.863, p = 0.022), were associated with more AD-like patterns of brain atrophy on MRI scans measured by the SPARE-AD index. Lower blood concentration of the glycerophospholipid, PC aa C40:6 (β = −0.323, 95% CI = −0.617–−0.029, p = 0.032), was associated with a more “AD-like” pattern of brain atrophy. Fig 3A shows all significant cross-sectional associations between blood concentrations of metabolites described above and the SPARE-AD index. A summary of results across all metabolites is included in S9 Table.
Higher blood concentrations of eight metabolites were associated with greater CSF levels of t-tau, while higher concentrations of 10 metabolites were associated with greater CSF levels of p-tau. All significant associations were among either sphingolipids (t-tau: 6 out of 8; p-tau: 8 out of 10) or glycerophospholipids (t-tau: 2 out of 8; p-tau: 2 out of 10). Higher blood concentrations of two of these sphingolipids (SM C16:0 and SM [OH] C14:1) were also associated with lower CSF levels of Aβ1–42. Lower blood concentrations of C3 and serotonin were also associated with lower CSF levels of Aβ1–42. Fig 3B shows associations between blood concentrations of SM C16:0 (t-tau: β = 0.347, 95% CI = 0.103–0.592, p = 0.006; p-tau: β = 0.331, 95% CI = 0.086–0.575, p = 0.008; Aβ1–42: β = −0.169, 95% CI = −0.328–−0.011, p = 0.036) and SM [OH] C14:1 (t-tau: β = 0.346, 95% CI = 0.109–0.583, p = 0.004; p-tau: β = 0.416, 95% CI = 0.179–0.653, p = 0.001; Aβ1–42: β = −0.179, 95% CI = −0.333–−0.025, p = 0.023), with all three CSF biomarkers. We additionally show associations between blood concentrations of PC aa C38:4 (t-tau: β = 0.279, 95% CI = 0.035–0.522, p = 0.025; p-tau: β = 0.251, 95% CI = 0.008–0.494, p = 0.043) and PC ae C34:0 (t-tau: β = 0.358, 95% CI = 0.053–0.663, p = 0.022; p-tau: β = 0.423, 95% CI = 0.118–0.728, p = 0.007) and CSF t-tau and p-tau. A summary of significant results across all metabolites is included in S10 Table.
The mean interval between baseline blood sampling to the onset of AD or follow-up (for individuals who remained MCI) was 2.97 years (SD = 2.33 years). Higher blood concentration of one sphingolipid, SM C18:1 (HR = 2.351, 95% CI = 1.268–4.360, p = 0.007), was associated with a significantly greater risk of conversion to incident AD among individuals with MCI. This sphingolipid was also associated with greater risk of conversion to incident AD among cognitively normal individuals (described above in BLSA results). Higher blood concentration of one glycerophospholipid, PC aa 38:4 (HR = 2.375, 95% CI = 1.189–4.745, p = 0.014), was also associated with a significantly greater risk of conversion to incident AD. A summary of results across all metabolites is included in S11 Table.
Fig 4 shows the heat map summarizing statistically significant associations between the metabolites and brain and blood-specific AD endophenotypes. Metabolites are ranked (in decreasing order) based on their EASE-AD score. P-values from each cell in Fig 4 are included in S12 Table.
Fig 5 summarizes the main biosynthetic and catabolic reactions (“Metabolic pathway”) of the major metabolite classes and their interactions as well as their roles in signaling cascades (“signaling pathway”) relevant to AD pathogenesis and evolution of the principal pathological features of the disease.
To the best of our knowledge, this is the first study to apply quantitative and targeted metabolomic analyses of both brain and blood tissue to identify metabolites associated with the severity of AD pathology as well as measures of AD progression. Our results indicate that distinct metabolites belonging to the sphingolipid and glycerophospholipid classes are related to the severity of AD pathology in the brain and that their concentrations in blood are associated with preclinical disease progression. Furthermore, we were able to identify these specific metabolites through a data-driven process that first used machine-learning methods to generate an AD-specific brain metabolite signature, and then clustered these metabolites based on the EASE-AD summary score representing cumulative associations of each metabolite, with outcome measures related to AD pathology and progression.
This process identified sphingolipids as a class of metabolites that are consistently associated with preclinical and prodromal AD, as well as with AD pathology at autopsy. Additionally, for all sphingolipid species—across all endophenotypes in brain, prodromal, and preclinical blood samples—increased concentration was associated with a more “AD-like” phenotype. Our results add substantially to a growing body of literature suggesting that perturbations in sphingolipid metabolism are related to key aspects of AD pathogenesis [43,44]. SMs are a subclass of sphingolipids that are enriched in the central nervous system as important constituents of lipid rafts [38] and play a critical role in neuronal cell signaling [45,46]. In the brain, sphingolipids mediate a diverse array of biological functions that are relevant to critical molecular mechanisms in AD, including amyloidogeneic processing of the amyloid precursor protein (APP) within SM-rich lipid rafts [47] and regulation of hippocampal neuronal excitability [48]. While previous studies in postmortem human brain tissue have demonstrated altered levels of total SM content in AD relative to CN [49,50], few have quantified absolute concentrations of distinct SM species within brain regions differentially vulnerable to AD pathology. Our findings are broadly consistent with those of Chan and colleagues, who demonstrated higher levels of the SM species SM d18:1/22:1 and d18:1/26:1 in the prefrontal and entorhinal cortices of AD patients, relative to CN [51].
Most previous studies reporting on altered blood sphingolipid levels in AD have used an untargeted lipidomics approach (e.g., [52,53]). Some recent studies have used the p180 targeted metabolomics platform to assay absolute concentrations of metabolites associated with AD. An important distinction in the design of these previous studies and our current report is in our use of a brain-derived AD metabolite signature to guide focused analyses of these metabolites in blood as well as a comprehensive exploration of their associations within both preclinical and prodromal AD samples. Two studies [25,54] have recently reported on p180 metabolite data within blood samples in the ADNI and the Atherosclerosis Risk in Communities (ARIC) cohorts. While there is minimal overlap between these results and our current report, it is striking to note that two sphingolipids we observe to be increased in the temporal cortex of AD patients and identified in our brain metabolite signature of AD (SM C16:0 and SM [OH] C14:1) were associated with brain atrophy, cognitive decline, and risk of conversion from MCI to AD in ADNI [25]. Similarly, blood concentrations of SM C16:0 and SM C26:1 were also associated with a diagnosis of MCI and dementia, respectively, in the predominantly African-American ARIC cohort [54].
Our findings that blood concentration of sphingolipids represented in the brain metabolite signature of AD are also associated with progression during preclinical and prodromal AD suggest that these are biologically relevant, early signals of disease progression. Equally importantly, correcting perturbations in sphingolipid metabolism may represent a plausible novel strategy for therapeutic intervention in AD. In this context, the emerging roles of sphingosine 1-phosphate (S1P)-metabolizing enzymes and S1P analogs in ameliorating Aβ-induced neuroinflammation in AD [55,56] are especially promising.
The second major class of metabolites we observed to be related to measures of AD pathology were glycerophospholipids (i.e., PCs and lysophosphatidylcholines [LysoPCs]). The majority of associations between these metabolites were in the brain tissue samples: generally, lower concentrations of glycerophospholipids were associated with greater severity of both amyloid and neurofibrillary pathology; associations between glycerophospholipids and preclinical and prodromal AD endophenotypes were sparse. In previous studies using untargeted and semiquantitative metabolomics, we demonstrated that AD patients show lower plasma concentrations of distinct phosphatidylcholines (PC aa C36:5, PC aa C38:6, and PC aa C40:6), relative to CN [57]. We recently extended these findings to show that reduced plasma concentration of these phosphatidylcholines is also related to lower levels of cognitive performance in non-demented older individuals and reflects resting state cerebral blood flow (rCBF), a marker of neuronal activity, in several brain regions related to higher order cognitive processing [58]. Taken together, these prior findings and our current results add further evidence for a role of altered phosphatidylcholine metabolism in AD pathogenesis.
In order to develop an integrated understanding of central–peripheral lipid metabolite fluxes as well as interactions between the major metabolite classes observed in this study, we applied a network biology approach. Fig 5 summarizes these networks, based on prior knowledge of transport mechanisms related to these metabolites and their precursors as well as their known biosynthetic pathways and catabolic fates. Long-chain fatty acid (LCFA) precursors for glycerophospholipid and sphingolipid biosynthesis are transported both across the blood-brain barrier (BBB) and through plasma membranes within the brain through protein-mediated active transport by fatty acid transport proteins (FATPs), long-chain acyl-CoA synthetases (ACSLs), fatty acid binding proteins (FABPs), and the fatty acid transporter (FAT)/CD36 [35,36]. In the context of neurodegenerative diseases in general and AD in particular, transport of the omega-3 (ω-3) polyunsaturated fatty acid (PUFA), docosahexaenoic acid (DHA; 22:6n-3), into the brain is especially important [59,60]. In a recent untargeted lipidomic analysis in brain tissue samples from the BLSA, we showed that dysregulation of fatty acid metabolism is associated with severity of AD pathology [24]. Fig 5 also shows key enzymatically regulated steps in the biosynthesis of phosphatidylcholines through the Kennedy pathway [37] and their reversible conversion to LysoPCs through Land’s cycle [41]. The transfer of phosphocholine headgroups to ceramides by the enzyme phosphatidylcholine transferase (sphingomyelin synthase [SGMS]) is a key intermediary step in sphingolipid biosynthesis [43] and is a potentially critical link between glycerophospholipid and sphingolipid metabolism observed in our current report.
By performing our initial discovery analyses in brain tissue samples at autopsy and subsequent validation in preclinical (i.e., BLSA) and prodromal (i.e., ADNI) serum samples, we were able to ask whether metabolic changes associated with markers of AD neuropathology in established disease are similar to blood metabolite changes in early AD pathogenesis. Broadly, our results indicate that there are shared pathways between metabolite changes in brain and blood, with the prodromal serum samples (i.e., ADNI) sharing more metabolites with brain samples than the preclinical (i.e., BLSA) serum samples (see Fig 4). A plausible explanation for these findings is that blood metabolite changes associated with later stages of AD progression prior to symptom onset are more similar to metabolic correlates of AD pathology in established disease. In independent analyses, we have also used the BLSA and ADNI serum samples as discovery datasets to ask whether principal metabolites associated with preclinical and prodromal AD-related endophenotypes in blood are also represented among the brain metabolites (i.e., in “established disease”) reported in the current study. Among serum metabolites previously shown to be associated with AD progression in BLSA, we find that propionylcarnitine (C3) concentration in serum discriminates between converter and non-converter samples [7], and its concentration in the ITG is related to the severity of neuritic plaque pathology in our current study (Fig 4). Among serum metabolites previously shown to be associated with AD endophenotypes in ADNI, we find that SMC (OH) C14:1 and SM 16:0 concentrations in serum are associated with CSF Aβ, concentration, brain atrophy, cognitive decline, and risk of MCI progression [25], and their concentrations in the ITG (i.e., our current report) are both related to the severity of neuritic plaque pathology and differ across the three groups studied (Fig 4). Taken together, while these findings suggest that there are metabolic pathways common to both AD-related neuropathology and blood-related disease progression, there are also those that are specific to disease stage and tissue compartment. Establishing the relative importance of common and distinct metabolic pathways across tissue types and disease stages will require subsequent studies in larger datasets.
Our study has limitations. First, the relatively small number of brain tissue samples in our primary analyses may have limited our power to detect significant associations with other metabolites assayed and precluded the use of a discovery and validation dataset. The small number reflects the challenges of assembling brain tissue samples from well-characterized, longitudinally followed participants who also undergo detailed neuropathological assessment at death; future studies in larger brain samples are needed to validate our findings. Second, while the Biocrates AbsoluteIDQ platform is a standardized platform for multiplexed quantitative analysis of 187 different metabolites, these metabolites represent only a small proportion of the brain and blood metabolomes. Future analyses will expand our study framework across additional classes of metabolites. Third, it must be noted that we based our primary analyses on metabolites associated with AD pathology in brain tissue samples. In future studies, it would be important to perform similar analyses in cognitively normal individuals using primary outcomes derived from neuroimaging/CSF-based measures of early AD pathology in prodromal/preclinical AD. Fourth, testing of pre-analytical variables in the BLSA serum samples indicated a potential selection bias: converter samples were subject to longer storage time at −80°C, compared to non-converter samples (approximately 17 years versus 13 years, respectively; Table 1). Additionally, the converter group compared to the non-converter group had more samples above the cutoff values for Met-So concentration used as an indicator of sample quality. Therefore, we performed sensitivity analyses within a subsample of converters and non-converters matched on storage time. In these sensitivity analyses, we confirmed that 10 of the 12 metabolites associated with AD-related outcomes in the BLSA serum samples (Fig 4) remained significant. We therefore interpret these sensitivity analyses to suggest that our observed results on serum metabolite concentrations in BLSA are not driven primarily by group differences in sample storage time or quality. Finally, it is important to note that the BLSA is a predominantly Caucasian sample of highly educated and relatively healthy older individuals. Our findings therefore merit confirmation in other cohorts with higher prevalence of cardiovascular and cerebrovascular disease.
In summary, we have applied quantitative and targeted metabolomics to identify a panel of sphingolipids, the concentrations of which, in brain tissue, are associated with severity of AD neuropathology and, in blood, with measures of progression during preclinical and prodromal AD. We propose that perturbations in sphingolipid metabolism may be integral to the evolution of AD neuropathology as well as to the eventual expression of AD symptoms in cognitively normal older individuals. Our study design, which takes a machine-learning and data-driven approach to identify blood metabolites associated with AD progression and explores how those metabolites are integrated within biologically relevant pathways, suggests a novel framework for identifying markers for early detection and potential avenues for effective therapeutic intervention in AD.
|
10.1371/journal.ppat.1005103 | Intrahepatic Transcriptional Signature Associated with Response to Interferon-α Treatment in the Woodchuck Model of Chronic Hepatitis B | Recombinant interferon-alpha (IFN-α) is an approved therapy for chronic hepatitis B (CHB), but the molecular basis of treatment response remains to be determined. The woodchuck model of chronic hepatitis B virus (HBV) infection displays many characteristics of human disease and has been extensively used to evaluate antiviral therapeutics. In this study, woodchucks with chronic woodchuck hepatitis virus (WHV) infection were treated with recombinant woodchuck IFN-α (wIFN-α) or placebo (n = 12/group) for 15 weeks. Treatment with wIFN-α strongly reduced viral markers in the serum and liver in a subset of animals, with viral rebound typically being observed following cessation of treatment. To define the intrahepatic cellular and molecular characteristics of the antiviral response to wIFN-α, we characterized the transcriptional profiles of liver biopsies taken from animals (n = 8–12/group) at various times during the study. Unexpectedly, this revealed that the antiviral response to treatment did not correlate with intrahepatic induction of the majority of IFN-stimulated genes (ISGs) by wIFN-α. Instead, treatment response was associated with the induction of an NK/T cell signature in the liver, as well as an intrahepatic IFN-γ transcriptional response and elevation of liver injury biomarkers. Collectively, these data suggest that NK/T cell cytolytic and non-cytolytic mechanisms mediate the antiviral response to wIFN-α treatment. In summary, by studying recombinant IFN-α in a fully immunocompetent animal model of CHB, we determined that the immunomodulatory effects, but not the direct antiviral activity, of this pleiotropic cytokine are most closely correlated with treatment response. This has important implications for the rational design of new therapeutics for the treatment of CHB.
| Approximately 250 million people are chronically infected with HBV, and over 500,000 people die every year because of associated liver diseases. IFN-α has been used to treat patients with chronic HBV infection for over 20 years, but it is not well understood why some patients respond to treatment and others do not. In large part, this is because it is not practicable to obtain liver samples to characterize the intrahepatic response to IFN-α in patients with different treatment outcomes. In this study we used the woodchuck model of chronic HBV infection to study how IFN-α changes gene expression patterns in the liver during treatment. Surprisingly, we found that the treatment response did not correlate with the expression of antiviral effector genes that have previously been shown to mediate the direct antiviral effects of IFN-α in vitro. Instead, we found that the response to IFN-α treatment was associated with the presence of select immune cells (natural killer cells and T cells) in the liver. Our work also indicates that these immune cells inhibit the virus by killing infected cells, as well as in ways that do not require killing of liver cells. Altogether, our study suggests that new therapies that stimulate these immune cells in the liver may hold promise for the treatment of chronic HBV infection.
| Approximately 250 million individuals live with chronic hepatitis B (CHB), and over half a million people are estimated to die each year due to CHB-associated liver diseases, such as cirrhosis and hepatocellular carcinoma (HCC) [1]. End-points of therapies for CHB are virological response (durable reduction in serum HBV DNA levels to a degree which varies by therapy), serological response (HBV e antigen (HBeAg) loss and seroconversion to anti-HBe in HBeAg-positive patients) and biochemical response (normalization of ALT levels). However, sustained loss of HBV surface antigen (HBsAg) off therapy is currently considered the ideal end-point. Recombinant interferon-α (IFN-α) is licensed for the treatment of CHB, but in contrast to potent nucleos(t)ides, virologic response is limited to a subset of patients [2]. Conversely, the rate of durable HBsAg loss is higher with IFN-α than with nucleos(t)ides, although still only occurs in <10% patients [2]. Despite more than two decades of clinical use, the mechanisms by which IFN-α controls HBV in responders are not well understood [3]. Defining the molecular basis for response remains an important goal, since mechanistic understanding of IFN-α activity could drive rational design of novel immunotherapeutic strategies and may lead to the identification of novel biomarkers of treatment response and/or patient stratification.
IFN-α is a pleiotropic cytokine that has both direct antiviral and immunomodulatory properties [4,5]. With regard to the former, IFN-α induces the expression of hundreds of interferon-stimulated genes (ISGs), many of which have antiviral effector functions [4]. Although the identification of key restriction factors has been challenging, various studies have indicated that IFN-α induces antiviral effectors of HBV. Most notably, the direct antiviral response to IFN-α has been demonstrated to inhibit the formation or accelerate the decay of replication-competent HBV capsids [6–9], inhibit virion secretion [10], reduce transcription from the viral genome (cccDNA; covalently closed circular DNA) [11,12], and to induce non-cytolytic degradation of cccDNA [13]. The direct antiviral activity of IFN-α is consistent with the reduction in viral antigen levels by high dose pegylated IFN-α in HBV-infected humanized mice that lack immune cells [14]. The immunomodulatory properties of IFN-α include activation of NK cells and B cells, as well as both direct and indirect activation of CD8+ T cell function [5,15]. Despite this potential to activate both innate and adaptive immunity, recent studies have revealed that IFN-α treatment boosts the number and function of NK cells in the periphery, but does not improve peripheral HBV-specific CD8+ T cells responses [16–19].
Antiviral and mechanistic studies of IFN-α treatment of HBV infection have been performed in vitro, in transgenic and immunodeficient mouse models, and in peripheral blood from CHB patients, but there is very little data regarding the intrahepatic response to IFN-α treatment in an immunocompetent host. A baseline (i.e. pre-treatment) intrahepatic transcriptional signature of response to treatment with pegylated IFN-α and adefovir (response defined as HBeAg loss, HBV DNA <2,000 IU/mL and ALT normalization) has recently been described [20]. However, due to the difficulty in obtaining multiple liver biopsy specimens from chronically infected HBV patients, longitudinal evaluation of the intrahepatic response to IFN-α treatment is only possible with an animal model. Since ethical and cost considerations limit the use of chimpanzees for biomedical research and there is no small animal model of natural HBV infection, we selected the woodchuck model for this purpose.
The Eastern woodchuck (Marmota monax) is naturally infected with WHV, a hepadnavirus which is genetically closely related to human HBV [21]. WHV infection displays a disease course similar to that in HBV-infected persons [21]. Although the woodchuck model has been used in a number of studies to characterize antiviral response to IFN-α treatment [22,23] these studies relied on adenovirus delivery of woodchuck IFN-α or utilized a recombinant human hybrid (B/D) IFN-α. Furthermore, these studies did not define the molecular basis of antiviral response. We recently described the sequencing, assembly and annotation of the woodchuck transcriptome, together with the generation of custom woodchuck microarrays. Using this new platform, we established the translational value of the woodchuck model and characterized the immune determinants of WHV clearance during self-limiting infection [24,25]. Since these studies yielded important insights into immune responses in the liver during hepadnavirus infection, in the current study we used a similar approach to characterize the intrahepatic transcriptional signature associated with antiviral response to recombinant woodchuck IFN-α treatment.
The amino acid sequence and in vitro antiviral activity of woodchuck IFN-α5 (wIFN-α) have previously been described [26,27]. wIFN-α was expressed, purified and biological activity confirmed as described in the Methods. The tolerability and pharmacodynamic activity of wIFN-α were then evaluated in a single dose study in WHV-negative woodchucks. Subcutaneous administration of a single dose of 2, 20 or 200 μg wIFN-α per animal (n = 3/group), induced dose-dependent increases in ISG and cytokine mRNA expression in the blood relative to the placebo group (S1 Fig). Pharmacokinetic (PK) analysis of serum wIFN-α levels was not performed due to the lack of a sufficiently sensitive quantitative method (see Methods). There was a trend towards changes in several hematological and clinical chemistry parameters at the higher doses, although these were likely due to the drawing of large blood volumes over a short time period.
The antiviral efficacy of wIFN-α was then evaluated in a repeat-dose study in adult woodchucks chronically infected with WHV. To model vertical transmission in humans, chronic infection in these animals was established by neonatal WHV infection. The study design is described in Fig 1. To match the frequency of non-pegylated IFN-α dosing in CHB patients, animals (n = 12/group) were dosed subcutaneously three times per week (TIW) on Monday, Wednesday and Friday with either placebo (vehicle control) or wIFN-α for a total of 15 weeks. Based on activity and safety considerations from the single dose study in WHV-negative woodchucks, the 20 μg dose was selected as the starting dose for the efficacy study. Initially wIFN-α was given for 7 weeks at a low dose of 20 μg/animal TIW. However, since an interim analysis indicated that this dose did not induce a significant decline in serum WHsAg or WHV DNA (Figs 2 and 3A), at the start of week 7 the wIFN-α dose was increased to 100 μg/animal TIW. Thus, in the wIFN-α treatment group, animals received a low dose of wIFN-α for 7 weeks (21 doses total), followed by a high dose of 100 μg/animal for another 8 weeks (24 doses total). Note that one animal in this group (M1004) was excluded from the analyses described below since it was likely naturally clearing WHV as the study initiated (Table 1).
In contrast to low dose (20 μg) wIFN-α, high dose (100 μg) wIFN-α treatment induced a rapid decline in serum WHsAg and WHV DNA (Figs 2 and S2), which was statistically significant relative to the placebo group (Fig 3A). The maximum reduction of serum WHsAg and WHV DNA was at week 16 in most animals, with a mean maximal reduction of 2.0 log10 for WHsAg and 3.0 log10 for viral load. Notably, wIFN-α treatment induced the complete loss of detectable (<20 ng/mL) WHsAg in one animal (F1022), although WHV DNA was still detectable (>1,000 genome equivalents (ge)/mL) at all time-points (S2 Fig). After completion of treatment there was WHsAg and WHV DNA rebound in most woodchucks, albeit not always to pre-treatment levels (Figs 2 and S2). There was a high degree of variability in the antiviral response of individual woodchucks in regard to the kinetics and magnitude of serum WHsAg and WHV DNA decline, as well to the time interval between cessation of treatment and return of these viral parameters to pre-treatment levels (S2 Fig). For correlative analyses with treatment response (see below), response groups were defined as the following: R, responder ≥1 log10 reduction in WHsAg at week 15 (end-of-treatment) and week 23 (end-of study) (n = 3 animals); PR, partial responder ≥1 log10 reduction in WHsAg at week 15 but not week 23 (n = 2 animals); NR, non-responder <1 log10 reduction in WHsAg at week 15 and week 23 (n = 2 animals) (Table 1). Notably, baseline (pre-treatment) levels of serum WHsAg and WHV DNA were comparable in these different treatment response groups (Table 1). The four animals in the wIFN-treatment group that did not survive until end-of study (see below), together with animal M1004 which was likely naturally clearing infection, were excluded from treatment response analyses (Table 1).
High dose wIFN-α treatment significantly reduced intrahepatic cccDNA, WHV DNA replicative intermediate (RI) and WHV RNA levels (Figs 3B and S3). Reductions in these intrahepatic parameters typically correlated with reductions in serum WHsAg and viral load (Table 1). Only two woodchucks (M1004 and F1022) with sustained WHsAg reduction developed consistently detectable anti-WHs antibodies (S1 Table), one of which (M1004) was likely naturally clearing WHV as the study initiated (Table 1). The overall seroconversion rate was therefore 0/9 (placebo group) and 1/7 (wIFN-α group) for animals that survived until end-of-study (excluding M1004).
wIFN-α treatment was well-tolerated, and there were no signs of overt toxicity based on gross observations, body weights, hematology or clinical chemistry. Although several animals died during treatment, the causes of death (e.g. HCC-related conditions, biopsy complications) were likely not treatment related (Table 1). There was a trend towards elevated serum ALT and AST levels during high dose treatment, but on a group level these overall differences were not statistically significant (Fig 3C). This is reflected in a poor temporal association between peak antiviral response and elevation of ALT, AST and SDH in some animals (Fig 4). Similarly, even though there was considerable fluctuation in liver histology scores in both placebo and wIFN-α groups (S1 Table), antiviral response was correlated temporally with an increase in liver inflammation in some (although not all) wIFN-treated animals (S4 Fig). Conversely, baseline liver enzyme levels and pre-treatment histology scores were comparable in the different treatment response groups (Figs 4 and S4).
wIFN-α treatment induced dose-dependent increases in blood ISG mRNA expression. There was significant induction at both low and high dose levels, with a larger increase observed for the higher dose (Fig 5A). In contrast, only high dose treatment significantly induced the expression of various T helper cell type 1 (TH1)-type cytokines (Fig 5B). Given that only high dose treatment was associated with a significant antiviral response, this suggests cellular immunity (and associated cytokines) may play a role in and/or be a useful biomarker of treatment response. Although comparative analysis is limited by small animal numbers in each response group, a role for cellular immunity in antiviral response is also suggested by the significant difference in IFN-γ expression in animals with an on-treatment response (R and PR) relative to those with no treatment response (NR) (S5 Fig).
As outlined in Fig 1, intrahepatic transcriptional profiles of placebo-treated and wIFN-treated animals were determined by RNA-Seq at various times during the study. RNA-Seq was performed rather than using the microarray platform from previous studies [24,25] because this method has superior concordance with qRT-PCR data [28] and also enabled generation of a more complete (version 2) woodchuck transcriptome assembly (S2 Table). Principal Component Analysis (PCA) demonstrated that wIFN-α treatment substantially altered gene expression within the liver of chronic carrier animals (S6 Fig). In contrast to the significant difference in antiviral response, there were only relatively modest differences (restricted to PC#2) between intrahepatic transcriptional changes induced by low dose (20 μg) and high dose (100 μg) wIFN-α treatment. A gene module approach [29] confirmed that there was substantial modulation of intrahepatic gene expression by wIFN-α overall, with only moderate differences between low and high dose treatment (Fig 6). The modular signature for wIFN-α treatment revealed an increase (>10% of the transcripts in each module significantly up-regulated) in the number of differentially expressed genes in the IFN response (Module, M3.1), cytotoxic cell (NK cell/CD8+ T cell) (M2.1), plasma cell (M1.1), B cell (M1.3), myeloid cell lineage (M1.5 and M2.6) and inflammation (M3.2) modules (Fig 6). Consistent with an increase in liver inflammation in many wIFN-treated animals (S4 Fig), the transcriptional data suggest that wIFN-α induced migration of immune cells into the liver and/or proliferation of intrahepatic immune cells.
In contrast to the differential antiviral response (Fig 3A and 3B) and dose-dependent ISG induction in the periphery (Fig 5A), module analysis revealed a striking increase (>80% of the transcripts significantly up-regulated) in the intrahepatic IFN response module (M3.1) at all on-treatment time-points, regardless of wIFN-α dose (Fig 6). Consistent with the modular analysis, low dose and high dose wIFN-α treatment were both associated with strong induction of a large number of intrahepatic ISGs, including many antiviral effector genes (Fig 7A, cluster 3). Furthermore, there was no apparent difference between the intrahepatic expression of these ISGs in animals with a treatment response (R and PR) and those with no treatment response (NR). Comparable induction of select ISGs in the liver by low and high dose wIFN-α treatment (regardless of treatment response) was confirmed by qRT-PCR (Fig 7B, S4 Table). Taken together, these data indicate that the antiviral response to wIFN-α does not correlate with the intrahepatic expression of the majority of ISGs, suggesting they do not play a key role in the antiviral response to treatment (see Discussion). Furthermore, pre-treatment (week -3) ISG levels were comparable in the different response groups (Fig 7A), indicating that baseline ISG expression was not an important determinant of treatment response.
In the context of defining the molecular basis of IFN-α treatment response, the APOBEC proteins are ISGs of particular interest since various family members have been reported to be restriction factors for HBV [13]. It is therefore notable that the intrahepatic expression profile of APOBEC3H (A3H) was unlike the majority of antiviral ISGs, in that it was selectively induced by high dose wIFN-α treatment (Table 2). However, the degree of A3H induction was modest (maximum 3.6-fold) relative to many other ISGs, consistent with low A3H induction by IFN-α in purified primary human hepatocytes [13]. Furthermore, intrahepatic induction of A3H was only statistically significant at end-of-treatment (week 15), suggesting that it is not likely to be a main mediator of the wIFN-α antiviral response. In contrast to A3H, A3D and A3F were not significantly modulated (FDR<0.05, FC>2) by wIFN-α treatment. Other APOBEC3 family members (including A3A) were not available in the woodchuck transcriptome assembly.
Since there was a strong association between wIFN-α dose and antiviral response (Fig 3A and 3B), we reasoned that determining which genes were selectively induced by high dose wIFN-α would enable the identification of genes and/or pathways closely associated with treatment response. This approach identified genes that were selectively modulated during high dose wIFN-α treatment (S7 Fig, high dose n = 468), as well as genes induced only by low dose treatment (low dose n = 29) or by both low and high dose wIFN-α (low & high dose n = 775). The full gene list from each set is displayed in S8 Table. Consistent with the previous analyses, module analysis (M3.1) and Ingenuity Pathway Analysis (IPA) confirmed significant induction of an IFN-α response at both low dose and high dose wIFN-α treatment (S8 Fig). In contrast, module analysis revealed that cytotoxic cell (NK cell/CD8+ T cell) responses were selectively induced by high dose wIFN-α treatment, and hence were temporally associated with treatment response (Fig 8A). Significant enrichment of NK and T cell signatures with high dose wIFN-α treatment was confirmed by IPA (Fig 8B). To complement the approach focused on identifying genes selectively induced by high dose wIFN-α, Weighted gene coexpression network analysis (WGCNA) was used to identify modules of co-regulated treatment-induced genes that correlated most closely with antiviral response (S5 Table, Modules 1 and 2). These modules were also significantly enriched for NK and T cell associated genes (S9 Fig), consistent with the trend for induction of an NK/T cell signature in animals that had an antiviral response to treatment (M2.1, S10 Fig). Notably, these diverse analytical approaches identified common intrahepatic transcriptional signatures associated with treatment response, suggesting that NK/T cells play an important role in the antiviral response to wIFN-α treatment.
On the individual gene level, induction of T cell associated genes (CD3D, CD8A) suggests that there is migration of T cells into the liver and/or proliferation of intrahepatic T cells during high dose wIFN-α treatment (Table 2). Expression of the T cell TH1-type transcription factor T-bet (TBX21) was also significantly induced during high dose treatment (Table 2). Strikingly, qRT-PCR analysis revealed that T-bet expression was strongly induced by high dose treatment in animals with treatment response but not in animals without an antiviral response (Fig 8C and S6 Table). This is notable since it may indicate improved functionality (antigen-specific proliferation and IFN-γ production) of intrahepatic HBV-specific CD8+ T cells, particularly since high dose wIFN-α also induced IL-12 expression (Fig 5B) [30]. However, it is important to note that this transcriptional analysis cannot determine whether T-bet is expressed by virus-specific or virus non-specific CD8+ T cells, or potentially other cell types [31]. Induction of NKG2D (KLRK1; activating receptor) expression, but not NKG2A (KLRC1; inhibitory receptor), CD16 (FCGR3A) or CD56 (NCAM1) (Tables 2 and 3), is consistent with activation, but not migration or proliferation of intrahepatic NK cells.
As discussed previously, the peak antiviral response to treatment and elevation of liver injury biomarkers were temporally correlated in some animals (Fig 4), indicating that wIFN-α induced killing of WHV-infected hepatocytes. This biochemical evidence of liver damage is consistent with intrahepatic induction of the receptor-mediated cell death genes TRAIL (TNFSF10), Fas (FAS) and Fas ligand (FASLG) and the cytotoxic effector gene perforin (PRF1) during high dose treatment (Table 3). These genes, as well as a death receptor signaling pathway (S8 Fig), were also significantly induced (on a group level) by low dose wIFN-α, consistent with liver enzyme elevations in some animals during this treatment period (Fig 4). Notably, although there was substantial induction of TRAIL expression (>17-fold) by high dose wIFN-α in two animals with a treatment response, one responder animal had only modest intrahepatic TRAIL induction (animal F1013; maximal 6-fold induction), and an animal with no treatment response had the greatest TRAIL induction (animal F1014; 118-fold) (S6 Table). This overall poor correlation of intrahepatic TRAIL with treatment response suggests that additional antiviral mechanisms may be required to control infection.
CD8+ T cells and NK cells have the potential to inhibit HBV infection by non-cytolytic mechanisms mediated by IFN-γ and TNF-α, as well as by killing infected cells via cytotoxic effector molecules. It is therefore notable that the two genes induced to the greatest degree in the liver by high dose wIFN-α treatment, PLA2G2A and CXCL9, are IFN-γ responsive genes [24] (Table 2). Furthermore, PLA2G2A, CXCL9 and other IFN-γ inducible genes (as well as IFN-γ itself) are members of a subset of intrahepatic ISGs that correlated with wIFN-α dose (Figs 7A, Cluster 2, and S11). Strikingly, a large number of IFN-γ-regulated genes (e.g. MHC class I and II (HLA) genes, CXCL9) were also induced in the liver of chimpanzees during clearance of acute HBV infection [32]. In addition, although the on-treatment profile was not determined, MHC class I and II genes as well as CXCL9 were also up-regulated prior to treatment in the liver of CHB patients that subsequently responded to pegylated IFN-α and adefovir treatment compared to non-responder patients [20]. Consistent with the association between blood IFN-γ expression and antiviral response (S5 Fig), high dose wIFN-α significantly induced intrahepatic CXCL9 expression in animals with a treatment response, but not in those without an antiviral response (Fig 8C and S6 Table). These data indicate that IFN-γ-mediated, non-cytolytic mechanisms may play a role in the antiviral response to wIFN-α treatment. This is supported by the observation that the initial reduction in WHsAg and WHV DNA by high dose treatment in two responder animals (F1013 and F1022, weeks 7–15 and 7–11, respectively) occurred in the absence of substantial liver enzyme elevations (Fig 4). In both animals, there were subsequently modest increases in liver enzyme levels together with a further decrease in viral levels, suggesting that the antiviral response induced by high dose wIFN-α treatment is mediated by both cytolytic and non-cytolytic NK/T cell responses.
In addition to positive effects on antiviral immunity, wIFN-α also induced various counter-regulatory mechanisms that may have limited the antiviral response to treatment. Notably, intrahepatic mRNA levels of the inhibitory T cell receptor PD-1 (PDCD1) and its ligand PD-L1 (CD274) were significantly increased during wIFN-α treatment (Tables 2 and 3). Intrahepatic expression of indoleamine 2,3-dioxygenase 1 (IDO1), which limits the availability of the essential amino acid tryptophan and produces immunosuppressive kynurenine to locally suppress T cells [33], was also significantly increased by wIFN-α treatment (Table 3). Furthermore, high dose wIFN-α modestly elevated intrahepatic FOXP3 mRNA levels (Table 2), which suggests treatment-associated migration and/or proliferation of T regulatory cells (Tregs) that may negatively regulate CD8+ T cell and NK cell function. In contrast, expression of IL-10 (IL10) and TGF-β (TGFB1), immunosuppressive cytokines produced by Tregs and various other cells, was not significantly modulated by treatment (Table 3).
Recombinant IFN-α has been used to treat CHB for over 20 years, but the molecular basis of treatment response remains poorly understood [3]. Previous transcriptome analyses have shown there are important parallels between the immune response to WHV in woodchucks and HBV in man [24], and that self-limiting hepadnavirus infection in woodchucks and chimpanzees share key immunological features [25]. Together these studies suggest that the woodchuck is a relevant model to study the mechanisms that govern antiviral response to IFN-α. Consequently, we characterized the intrahepatic transcriptional profile of WHV chronic carrier woodchucks during treatment with recombinant woodchuck IFN-α. Treatment with wIFN-α produced variable antiviral effects, inducing multi-log reduction in serum WHV DNA and WHsAg in a subset of animals, and sustained WHsAg loss and seroconversion to anti-WHsAb in one animal, while not exerting antiviral effects in other animals. Importantly, the variability and degree of antiviral response in these animals are comparable to those observed with CHB patients treated with pegylated IFN-α [2,34]. Furthermore, viral rebound in the WHV-infected woodchucks was typically observed following cessation of wIFN-α treatment, consistent with the low rate of durable HBsAg loss in patients treated with IFN-α [2]. Together, these data reveal important parallels between the IFN-α treatment response of chronic hepadnavirus infection in woodchucks and man, establishing the translational value of the woodchuck model for characterizing the immune correlates of IFN-α treatment response.
Since various studies have demonstrated that IFN-α can directly inhibit HBV [6,11,13], a striking finding of this study was that the antiviral response to wIFN-α did not correlate with the intrahepatic induction of the majority of antiviral ISGs. Since WHV is sensitive to the direct antiviral effects of wIFN-α in vitro [23,35], our data suggest that IFN-induced antiviral effectors of WHV do not play a key role in the antiviral response to treatment in vivo. However, there are several important caveats to consider. Firstly, since there are a large number of antiviral ISGs, and not all were available in the woodchuck transcriptome (e.g. APOBEC3A), intrahepatic expression of antiviral ISGs that were not evaluated in this study may correlate with treatment response. In addition, since intrahepatic transcriptional analysis was restricted to 6 hours post-dose, it is possible that the expression of certain ISGs with slower induction kinetics may be associated with the antiviral response to IFN-α. Secondly, although low dose wIFN-α induced intrahepatic ISG expression but not a significant antiviral response, it is conceivable that prolonged ISG expression (7 weeks) by low dose treatment played an important role in the antiviral response subsequently induced by higher dose wIFN-α. Finally, transcriptional analysis of whole biopsy tissue cannot define cell-specific ISG expression, which may be important in treatment response [36]. This is also an important caveat if hepatocytes and non-parenchymal cells (e.g. Kupffer cells) display markedly different sensitivity to ISG induction, since it may preclude accurate correlation of treatment response with induction of antiviral ISGs in infected cells using whole biopsy tissue.
The significant difference in ISG induction by low and high dose wIFN-α in the blood but not the liver of woodchucks chronically infected with WHV is noteworthy considering a recent study demonstrating that HBV can inhibit IFN-α signaling in human hepatocytes [37]. This suggests that WHV may limit (although not abrogate) wIFN-α signaling in woodchuck hepatocytes. Alternatively, induction of USP18, SOCS1 and SOCS3 (Table 3) and/or other inhibitors of IFN-α/β receptor signaling may limit the intrahepatic ISG response to wIFN-α treatment. Since liver biopsies were not taken after wIFN-α treatment of WHV-negative animals, and there is currently no sensitive, quantitative wIFN-α ELISA (see Methods), additional studies will be required to determine whether there are significant differences in PK-PD responses to wIFN-α treatment in WHV-negative and WHV-infected animals.
In contrast to intrahepatic ISG expression, the expression of other gene sets showed a correlation with antiviral response. Both NK/T cell and IFN-γ transcriptional signatures in the liver were increased in animals with antiviral response to wIFN-α treatment. The peak antiviral response was also associated with liver enzyme elevations in some (although not all) animals. Collectively these data suggest that the antiviral response induced by wIFN-α treatment was mediated by both cytolytic and non-cytolytic NK/T cell responses. The correlation of liver injury biomarkers with antiviral response is notable because host-induced ALT flares are associated with IFN-α treatment response in CHB patients [38]. The association of intrahepatic NK cell and IFN-γ transcriptional signatures with antiviral response to treatment is also striking because NK cells in CHB patients have a markedly impaired capacity to produce IFN-γ [39,40]. This dysfunctional phenotype can be reversed (at least in NK cells in the periphery) by treatment with IFN-α [17], which suggests that NK cell IFN-γ production may represent a common mechanism of IFN-α antiviral response to chronic hepadnavirus infection in woodchucks and man. Clearance of acute HBV infection in chimpanzees is also characterized by an intrahepatic IFN-γ transcriptional signature [32], suggesting that there are important parallels between the immunological mechanisms of natural clearance of HBV and those induced by IFN-α treatment.
Recent studies have revealed that IFN-α treatment does not improve peripheral HBV-specific CD8+ T cell responses [16–18]. In view of the aforementioned NK cell activation by IFN-α, this failure to augment virus-specific CD8+ T cell responses may be explained, at least in part, by the observation that NK cells can directly kill HBV-specific CD8+ T cells via TRAIL and other mechanisms [41]. The induction of an intrahepatic NK signature as well as TRAIL expression suggests that WHV-specific CD8+ T cell responses may be inhibited by similar mechanisms during wIFN-α treatment. Conversely, IFN-induced protection of antiviral CD8+ T cells might limit NK regulation of T cell immunity in this setting [42,43], consistent with the induction of an intrahepatic T cell transcriptional signature coupled with significant elevation of T-bet (TBX21) mRNA during wIFN-α treatment. In addition to potentially inducing NK cell killing of virus-specific CD8+ T cells, wIFN-α treatment induced various counter-regulatory mechanisms, including intrahepatic PD-1 (PDCD1) and PD-L1 (CD274) expression, which may also have limited antiviral CD8+ T cell function in the liver. However, it is important to the note that a limitation of the woodchuck model is that it is challenging to confirm that changes in gene expression are associated with corresponding changes in protein levels and/or cellular function. This is particularly important for characterization of CD8+ T cell specificity in the context of wIFN-α treatment, since studies in HBV transgenic mice as well as CHB patients indicate that antigen-nonspecific inflammatory cells (including nonvirus-specific CD8+ T cells) can accumulate to high levels in the liver under inflammatory conditions [44,45]. Unfortunately, blood volume and biopsy material limitations precluded functional analysis of WHV-specific CD8+ T cells in the current study. In addition, the lack of woodchuck-specific immunological reagents prevented immunophenotyping of WHV-specific CD8+ T cells by flow cytometry. Attempts to develop high-quality monoclonal antibodies against woodchuck CD56 and CD8a to enable detection of NK and CD8+ T cells, respectively, by immunohistochemistry were also not successful. Therefore, additional studies in immunocompetent models of natural infection and/or CHB patient biopsies will be required in order to define the relative contribution of intrahepatic NK and virus-specific CD8+ T cells to IFN-α treatment response.
In summary, by studying recombinant IFN-α in an immunocompetent animal model of CHB, this study provided new insights into the immune mechanisms that mediate the antiviral response to treatment. In addition, various immune pathways were identified that may act to limit treatment response. These findings have important implications for the design of new therapeutics for CHB, and also provide rationale for evaluating combinations of immunotherapeutic agents currently in development.
The sequence of woodchuck IFN-α5 (wIFN-α) has previously been described [26]. Recombinant wIFN-α was expressed by transient transfection of human embryonic kidney (HEK) 293F cells using the FreeStyleTM 293 expression system according to the manufacturer’s instructions (Invitrogen, Inc., Carlsbad, CA). Culture supernatant was filtered and then purified by two chromatographic steps. Firstly, after adjusting to pH 6.0 with 50 mM KH2PO4, pH 5.0, the sample was loaded on a 5 mL SP HP Hi Trap (GE Healthcare, Little Chalfont, Buckinghamshire, UK) that had been pre-equilibrated with 50 mM KH2PO4, pH 6.0. The wIFN-α was then eluted with a 17 column-volume salt gradient from 0–500 mM NaCl. Fractions were analyzed via SDS-PAGE and wIFN-containing fractions were pooled. Secondly, size exclusion chromatography on Superdex 75 (GE Healthcare, Little Chalfont, Buckinghamshire, UK) was performed in 20 mM His/HCl, 140 mM NaCl pH 6.0. The eluted wIFN-α was filtrated with a 0.22 μM syringe filter and stored at -80°C. The wIFN-α concentration was determined by measuring optical density (OD) at 280 nm. Purity and monomer content were confirmed by SDS-PAGE and SE-HPLC, respectively, and the integrity of the wIFN-α amino acid backbone was verified by Nano Electrospray QTOF mass spectrometry. The protein was kept in a storage buffer (20 mM His/HCl, 140 mM NaCl pH 6.0) prior to dosing. The endotoxin level of the wIFN-α preparation was <0.454 EU/mL. The in vitro biological activity of wIFN-α was confirmed by dose-dependent induction of mRNA levels of the interferon-stimulated genes (ISGs) Mx1 and OAS1 in woodchuck PBMCs (n = 2 animals) treated with 0.1, 1 and 10 μg/mL wIFN-α.
The animal protocol and all procedures involving woodchucks were approved by the Georgetown University IACUC (Protocol Number: 11–006) and adhered to the national guidelines of the Animal Welfare Act, the Guide for the Care and Use of Laboratory Animals, and the American Veterinary Medical Association.
All woodchucks used in this study were obtained from Northeastern Wildlife. Prior to the study, male woodchucks were confirmed negative for WHV surface antigen (WHsAg) and for antibodies against WHsAg (anti-WHsAb) and WHV core antigen (anti-WHc). Animals were assigned to four groups (n = 3/group) using stratification based on body weight, clinical biochemistry and hematology. Animals received a single subcutaneous dose of 2, 20 or 200 μg wIFN-α, or a placebo control (all n = 3/group). Various measurements (body weight, body temperature, clinical serum chemistries, and CBCs) were obtained to monitor drug safety.
All woodchucks used in this study were obtained from Northeastern Wildlife. These woodchucks were born in captivity and were infected at 3 days of age with the cWHV7P2a inoculum containing WHV strain WHV7-11. cWHV7P2a has the same biological and virological characteristics as the cWHV7P2 inoculum as both were derived from cWHV7P1 [46]. Chronically infected animals were all anti-WHs negative, with detectable serum WHV DNA, WHsAg and anti-WHc at approximately 1 year post-infection. Absence of liver tumors in woodchucks with low GGT was confirmed by ultrasonography. Chronic WHV carrier woodchucks were assigned and stratified by gender, body weight, and by pretreatment serum markers (WHsAg and WHV DNA concentrations, serum GGT and SDH activities) into treatment and placebo groups (n = 12/group). The study design and sampling scheme are summarized in Fig 1.
The PK of wIFN-α was not measured due to the lack of a suitable analytical method. Although a wIFN-α ELISA has previously been described [27], it was discovered during method development that one of the antibodies likely recognized the 6xHis tag of the antigen used for immunization, which was not present in our preparation of wIFN-α. Despite extensive screening of available anti-human, anti-macaque, anti-mouse and anti-pig IFN-α antibodies (PBL, Piscataway, NJ), as well as additional anti-woodchuck IFN-α antibodies (Digna Biotech, Pamplona, Spain), none were identified that robustly detected wIFN-α in an ELISA format.
Serum WHV DNA was quantified by two different methods depending on concentration: dot blot hybridization or real time PCR assay on a 7500 Real Time PCR System instrument (Applied Biosystems, Foster City, CA) as described previously [47]. Serum WHsAg and anti-WHsAb were measured by WHV-specific enzyme immunoassays as described [48]. Liver WHV RNA was measured quantitatively by Northern blot hybridization as previously described [49]. Liver WHV DNA replicative intermediates (RI) and WHV cccDNA were quantitatively determined by Southern blot as previously described [50].
The revised woodchuck transcriptome assembly (version 2) consists of a previous assembly (version 1), generated with Roche-454 sequencing data [24], that was merged with newly assembled contiguous transcripts (contigs) from Illumina sequencing data of the 24 animals from the current study (n = 12 placebo, n = 12 wIFN-α treated). The main improvement of version 2 over version 1 is that the sequencing depths of the Illumina data is significantly higher than that of 454 and therefore resulted in a higher dynamic range and increased number of genes as compared to assembly version 1 (S2 Table). The assembly method of transcriptome version 2 consisted of three stages: 1) initial contig assembly, 2) contig annotation and 3) contig refinement. First, Illumina RNA-Seq paired-end reads from liver samples were assembled using Trinity [51] (release 2011-08-20). The obtained contigs were further refined and merged by applying the sequence assembly algorithm PHRAP [52]. As a result, the number of contigs was reduced by about 25% and the contig lengths were increased. Second, all contigs were subjected to an in-house developed gene annotation pipeline which performs sequence homology searches within reference transcript databases from other species. First, woodchuck contigs were mapped to transcripts from RefSeq reference database containing human, mouse, and rat transcripts using BLAST [53], with a 1.e-5 E-value cutoff. Matches with the highest BLAST scores were further pair-wise aligned by applying the Needleman-Wunsch algorithm [54] in order to obtain more accurate alignments and to calculate the sequence identities (i.e. number of identical nucleotides in percentage of alignment length) between RefSeq transcripts and woodchuck contigs. If the identity difference between the two best hits exceeded 25%, then the top gene was used for contig annotation. Only contigs that could be mapped to known mouse, rat or human genes were used for further data processing. Because the assembly often contained more than one contig per gene, a final sequence refinement was then performed to remove redundancies. Contigs annotated with identical genes were subjected to the CAP3 assembler [55], and as a result, the number of contigs was further reduced and the sequence lengths of numerous contigs were increased.
Sequencing libraries were created using Illumina’s TruSeq RNA sample preparation kit (San Diego, CA) according to manufacturer’s protocol. Total RNA was purified using oligo(dT) magnetic beads, fragmented, and reverse-transcribed using SuperScript II (Invitrogen, Inc., Carlsbad, CA) to synthesize first strand cDNA. After second strand synthesis, Illumina specific adapters containing unique barcodes were ligated to the ends of the double-stranded cDNA. Fragments containing adapters on both ends were then enriched and amplified with PCR, quantified with qPCR, and run on the Agilent Bioanalyzer DNA-1000 chip to estimate fragment size. Samples were then multiplexed and sequenced on the Illumina 2500. The data was demultiplexed using CASAVA and run through FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/) to assess sequencing data quality. Paired-end 50 nucleotide read data from mRNA-Seq were mapped against the revised woodchuck transcriptome with Bowtie2 [56] and prioritized for concordant paired alignments with unique hits. The resulting SAM/BAM files were processed with SAMtools [57] to yield count data that was normalized and processed by DESeq [58] for differential expression analysis and subsequent pattern recognition and pathway analysis. Multiple testing correction was performed using the method of Benjamini and Hochberg [59]. Principal component analysis was performed with Partek Genomics version 6.6beta (Partek, St. Louis, MO). Heatmaps of the expression data were generated by unsupervised hierarchical clustering of least square means expression values, after z-score normalization across samples. The enrichment of differential genes relative to the gene modules described previously [29] was calculated with R version 2.13.2 (http://www.r-project.org) using the humanized gene symbols for the woodchuck genes. Gene Set Enrichment Analysis (GSEA) was performed as previously described [60], with ranks determined by the multiplicative product of the fold-change and–log(FDR) values for each gene. Weighted gene coexpression network analysis (WGCNA) [61] was performed within the R statistics environment. Pathway analysis was performed using Ingenuity Pathway Analysis (Ingenuity Systems, Redwood City, CA).
Total RNA was isolated using the RNeasy Mini Kit (Qiagen Inc., Redwood City, CA) with on-column DNase digestion using the RNase-Free DNase Set (Qiagen Inc., Redwood City, CA). Following reverse transcription into cDNA with the Transcriptor First Strand cDNA Synthesis Kit (Roche Applied Sciences, Indianapolis, IN), samples were analyzed by real time PCR on a 7500 Real Time PCR System instrument (Applied Biosystems, Inc., Foster City, CA) using EagleTaq Universal Master Mix (Roche Applied Sciences, Indianapolis, IN). Target gene expression was normalized to 18S rRNA expression. The primers and probes used in this study are displayed in S7 Table, or have previously been described [24,62–65]. Note that, although there was insufficient mRNA from biopsy samples to perform extensive qRT-PCR validation of gene expression, in contrast to microarray, RNA-Seq has high concordance with qRT-PCR data [28].
|
10.1371/journal.ppat.1007865 | Unraveling the role of the secretor antigen in human rotavirus attachment to histo-blood group antigens | Rotavirus is the leading agent causing acute gastroenteritis in young children, with the P[8] genotype accounting for more than 80% of infections in humans. The molecular bases for binding of the VP8* domain from P[8] VP4 spike protein to its cellular receptor, the secretory H type-1 antigen (Fuc-α1,2-Gal-β1,3-GlcNAc; H1), and to its precursor lacto-N-biose (Gal-β1,3-GlcNAc; LNB) have been determined. The resolution of P[8] VP8* crystal structures in complex with H1 antigen and LNB and site-directed mutagenesis experiments revealed that both glycans bind to the P[8] VP8* protein through a binding pocket shared with other members of the P[II] genogroup (i.e.: P[4], P[6] and P[19]). Our results show that the L-fucose moiety from H1 only displays indirect contacts with P[8] VP8*. However, the induced conformational changes in the LNB moiety increase the ligand affinity by two-fold, as measured by surface plasmon resonance (SPR), providing a molecular explanation for the different susceptibility to rotavirus infection between secretor and non-secretor individuals. The unexpected interaction of P[8] VP8* with LNB, a building block of type-1 human milk oligosaccharides, resulted in inhibition of rotavirus infection, highlighting the role and possible application of this disaccharide as an antiviral. While key amino acids in the H1/LNB binding pocket were highly conserved in members of the P[II] genogroup, differences were found in ligand affinities among distinct P[8] genetic lineages. The variation in affinities were explained by subtle structural differences induced by amino acid changes in the vicinity of the binding pocket, providing a fine-tuning mechanism for glycan binding in P[8] rotavirus.
| The interaction of viruses with host glycans has become an important topic in the study of enteric virus infectivity. This interaction modulates several aspects of the viral cycle, including viral attachment, which in most cases depends on the host glycobiology dictated by the secretor and Lewis genotypes. The capacity to synthesize secretory type-I antigen H (fucose-α1,2-galactose-β1,3-N-acetylglucosamine; H1) at the mucosae, dictated by the presence of one or two functional copies of the fucosyl-transferase FUT2 gene (secretor status), has been clearly linked to infectivity in important enteric viruses such as the noroviruses. However, a big controversy existed about the contribution of H1 antigen to infection in the leading cause of viral gastroenteritis in young children (rotavirus). It has not been until recently that epidemiological data evidenced a diminished incidence of rotavirus in non-secretor individuals unable to produce H1. In the present manuscript we offer the evidence that P[8] RV bind H1 via a binding site common for the P[II] RV genogroup and that the H1 precursor lacto-N-biose (galactose-β1,3-N-acetylglucosamine; LNB) is also bound to this pocket with diminished affinity. The P[8] VP8* structures show a marginal role for the L-fucose moiety from H1 in protein interaction. However, its presence provides conformational changes in the LNB moiety that increase the affinity of VP8* for the H1 ligand and would account for a stronger RV binding to mucosa in individuals expressing H1 (secretors). We thus offer a mechanistic explanation for the different incidence of P[8] RV infection in different secretor phenotypes.
| Rotaviruses are the leading etiologic agent of viral gastroenteritis in infants and young children worldwide and are responsible for an estimated 140,000 deaths each year in developing countries [1]. The typical classification of rotaviruses was derived from their genome composition and the immunological reactivity of three of their structural proteins: VP6, VP7 and VP4. Rotaviruses are classified into at least 7 groups (A to G) according to the immunological reactivity of the VP6 middle layer protein, with group A rotavirus being the most commonly associated with infections in human. The two outer capsid proteins VP7 and VP4, elicit neutralizing antibodies that can induce viral protection. Using these two proteins, a traditional dual classification system of group A rotaviruses into G (depending on the VP7 glycoprotein) and P (depending on the protease-sensitive VP4) types was established [2]. At least 36 different G-serotypes and 51 P-types have been identified among human and animal rotaviruses [3]. Viruses carrying G1[P8], G2[P4], G3[P8] and G4[P8] represent over 90% of human rotaviruses strains co-circulating in most countries [2], with the P[8] genotype, which comprises four different genetic lineages (I to IV [4]), being particularly relevant [5].
Diverse interactions between histo-blood group antigens (HBGAs) and rotavirus have been described and it is believed that HBGAs expressed on the surface of target cells serve as viral receptors. The distal VP8* portion (~27 kDa, N-terminal) of the rotavirus spike protein VP4 from P[8], P[4], P[6] and P[19] genotypes recognize the secretor HBGAs. P[8] and P[4] are closely related genetically and both genotypes were reported to bind the Lewisb (Fuc-α1,2-Gal-β1,3-[Fuc-α1,4-]GlcNAc; Leb) and H type-1 (H1) antigens (Fuc-α1,2-Gal-β1,3-GlcNAc) by some authors [6], while there are controversial reports that show no Lewisb binding for these genotypes [7]. P[6], a slightly further related genotype, binds the H1 antigen but not Lewisb [6], whereas P[19] binds mucin core glycans with GlcNAc-β1,6-GalNAc motif and the type-1 HBGA precursor [8]. In addition, P[9], P[14] and P[25] genotypes bound specifically to the type A antigens (GalNAc-α1,3-[Fuc-α1,2-]Gal)[9, 10], whereas P[11] interacted with single and repeated N-acetyllactosamine (Gal-β1,4-GlcNAc; LacNAc), the type-2 precursor glycan [11]. Detailed evidences of VP8*-HBGAs interactions has been obtained by X-ray crystallography of P[14] VP8* and P[11] VP8* in complex with the type A oligosaccharide [9] and LacNAc [11], respectively. Recently, the structure of porcine P[19] VP8* complexed with lacto-N-fucopentaose I (Fuc-α1,2-Gal-β1,3-GlcNAc-β1,3-Gal-β1,4-Glc; LNFPI), and the mucin core-2 oligosaccharide (GlcNAc-β1,6-[Gal-β1,3] GalNAc) has been solved, showing a carbohydrate binding pocket alternative to the one used by P[11] and P[14] [7]. This binding pocket was first suggested by protein sequence analyses in other members belonging to the P[II] genogroup of rotaviruses (i.e. P[4], P[6], P[8] genotypes) [7] and recently confirmed for P[4] and P[6] VP8*s [12]. However, why non-secretors individuals (lacking α1,2 fucosylation in secretory HBGAs) have reduced rotavirus susceptibility [13–15] and what is the role of the secretory L-fucose in H1 ligand recognition for the most relevant human rotavirus remains unknown, as no structure was still available for the P[8] genotype in complex with ligand HBGAs. By using VP8* from a clinical isolate belonging to the lineage III of P[8], in the present work we show that P[8] VP8* binds H1 antigen at a similar site as their ligand HBGAs bind to P[19], P[4] and P[6] genotypes. Our structural and functional results also show that the H1 precursor lacto-N-biose (Gal-β1,3-GlcNAc; LNB), devoid of L-fucose, also interacts with VP8* and we discard Lewisa or Lewisb antigens as ligands of P[8] genotypes. We provide the molecular bases for the role of secretor antigen in rotavirus binding to its receptor. Our results show two-fold increase in the affinity for the H1 antigen compared to LNB. This increase is explained by reduced contacts of the L-fucose with solvent molecules and the structural stabilization of LNB moiety in the competent conformation for binding. Furthermore, we show how subtle differences at the H1/LNB binding pocket in different P[8] lineages influence antigen affinities, giving clues for the relevance of the host glycobiology in P[8] rotavirus impact in humans. The partial anti-adhesin effect of LNB against rotavirus reported in the present article and the acquired knowledge on rotavirus-host cells interaction during virus attachment might open new avenues for the treatment and prevention of rotavirus infections.
VP8* domains from P[4], P[6], P[9], P[11], P[14], P[25] genotypes and from different genetic lineages (I, III and IV) from P[8] genotype were produced (S1 Fig). In order to confirm their functionality, the different proteins were challenged by an ELISA-like binding assay against a panel of biotinylated histo-blood group antigens (HBGAs) (Fig 1 and S1 Table), corroborating the previously described interactions (S2 Fig). Genotypes P[4], P[6] and P[8] recognized the H type-1 antigen (Fuc-α1,2-Gal-β1,3-GlcNAc, H1), P[11] recognized the H type-2 antigen (Fuc-α1,2-Gal-β1,4-GlcNAc, H2) and P[9], P[14] and P[25] the blood group A antigen trisaccharide (GlcNAc-α1,3-(Fucα1,2)-Gal, Atri). The P[8] genotype additionally displayed low binding to this trisaccharide. However, and contrarily to previous reports [7], in our assays genotypes P[4] and P[8] exhibited very low or absence of binding to Lewisb (Fuc-α1,2-Gal-β1,3-[Fuc-α1,4-]GlcNAc). Remarkably, VP8* from P[8] genotype recognize the H1 antigen precursor lacto-N-biose (Gal-β1,3-GlcNAc, Lewisc, LNB; S2 Fig) but differences in the binding abilities were found among different genetic lineages and strains. Thus, VP8* from the cultivable human rotavirus Wa (P[8]Wa) and the Rotarix vaccine (P[8]Rotarix) strains (both lineage I) gave lower signals in the ELISA tests with H1 and LNB than the VP8* from the lineage IV strain (P[8]LIV) and from a clinical isolate belonging to lineage III (P[8]c) (S2 Fig). The newly discovered interaction with LNB was further characterized by testing different concentrations of VP8* from this isolate (P[8]c) and the cultivable P[8]Wa (S3 Fig). The results showed that both proteins were able to bind H1 and LNB, although binding to the first antigen was higher. We performed a more detailed characterization of the interaction of VP8* to H1 and LNB by determining the apparent affinity constants (Kda) for each interacting pair by SPR (Table 1). The Kda of the VP8* P[8]c recognition of the H1 antigen (Kda = 27.9 ± 0.71 μM; Fig 2A) was two-fold lower compared to the LNB precursor (Kda = 52.1 ± 4.26 μM; Fig 2B), this difference was significant (p = 0.0045) suggesting that the H1 L-fucose moiety contributes actively to the binding. Surprisingly, the affinity constant for the interaction of VP8* from P[8]Wa with H1 was three times higher (Kda = 80.2 ± 2.21 μM) than that of P[8]c (Fig 2C). Furthermore, the VP8* from P[8]Wa showed a similar apparent affinity for LNB than for H1 (Kda = 66.5 ± 6.47 μM; p > 0.05 Fig 2D). Interestingly, interactions for VP8* from Rotarix strain (lineage I) and the lineage IV strain with H1 and LNB were too low to be determined under our SPR conditions.
To further characterize the role of VP8* interaction with the H1 precursor LNB, complete virions of the Wa strain (triple-layered particles; TLP) and double-layered particles (DLP; obtained after VP4 and VP7 removal by EDTA treatment) were assayed by an ELISA-like binding assay. The Wa TLP, but not the Wa DLP, were able to bind H1 antigen (Fig 3A) and its precursor LNB (Fig 3B), in a concentration-dependent manner. This indicated that the observed interaction between P[8] VP8* and LNB was also relevant in a complete rotavirus context. These results also point to the fact that despite the low affinity of P[8] VP8* from lineage I to H1 and LNB, the high avidity of a multi-binder particle (virions contain 120 molecules of VP4) results in a measurable interaction.
To confirm the binding of VP8* from the P[8]c genotype to LNB we settled up a binding blocking assay where LNB (Gal-β1,3-GlcNAc) and its structurally-related disaccharide galacto-N-biose (Gal-β1,3-GalNAc; GNB), were tested as potential inhibitors of binding. The results showed a moderate but significant (p < 0.05) reduction in the binding to the H1 antigen by both disaccharides (24.2% reduction for LNB and 30.1% for GNB; Fig 4A). The monosaccharide constituents of LNB and GNB (D-galactose, N-acetyl-glucosamine and N-acetyl-galactosamine) and L-fucose were also tested (Fig 4). Among these sugars only D-galactose possessed a discrete but significant blocking capacity of VP8* binding to the H1 antigen (14.7% reduction; p = 0.032). Interestingly, soluble L-fucose significantly increased the binding of VP8* to the H1 antigen (Fig 4A). As expected, when the precursor LNB was used as the ligand, soluble LNB and GNB were also able to reduce the binding of VP8* P[8]c, by 52.2% and 44.1%, respectively (Fig 4B).
We next investigated the role of the H1 precursor antigen in rotavirus infection by incubating MA104 cells with rotavirus Wa strain that was preincubated with LNB, GNB, their monosaccharide constituents and L-fucose. Only LNB significantly blocked viral infection (33% reduction; Fig 4C), suggesting that this HBGA precursor interferes with the binding of the rotavirus Wa strain to its receptor in MA104 cells.
To understand the molecular basis of LNB and H1 recognition and binding to P[8] VP8* we determined the crystal structure of the clinical isolate P[8]c VP8* in its apo form and bound to LNB and H1 glycans. Two different crystalline forms of P[8]c VP8* in its apo form were obtained. The first form, VP8*-Apo1, diffracts at 1.35 Å resolution and presents a single copy of P[8]c VP8* in the crystal asymmetric unit (ASU) while the second form, VP8*-Apo2, diffracts to 1.5 Å and presents two copies in the ASU (Table 2). The three copies of P[8]c VP8* in these two crystalline forms present the galectin fold with two twisted β-sheets separated by a superficial cleft that conforms the glycan binding site in P[11] and P[14] genotypes [9, 16] (Fig 5). Superimposition of the individual VP8* protomers from these crystals showed that the three molecules of P[8]c VP8* are almost identical (RMSD 0.28 Å for the superposition of all the Cα; S4A Fig and S2 Table). Crystals of P[8]c VP8* bound to H1 or its precursor LNB were obtained in a third different crystalline form and the structures of P[8]c VP8*-H1 and P[8]c VP8*-LNB complexes were solved to 1.8 and 1.3 Å, respectively (Fig 6A, 6B and Table 2). Two protomers of VP8* were present in the ASU of each of these crystals and, remarkably, only one out of the two protomers showed a bound glycan molecule. Sugar binding induces negligible conformational changes in the VP8* (S4A Fig) since the structural comparison of the glycan-bounded and glycan-free P[8]c VP8* protomers showed minimal differences (RMSD 0.22–0.35 Å; S2 Table). Furthermore, P[8]c VP8* are also structurally identical (RMSDs 0.37–0.57 Å) to the VP8* apo forms of the linage I P[8]Wa and P[8]Rotarix (S4B Fig and S2 Table), supporting that the glycan binding site is preformed in the P[8]c VP8* protein. This characteristic also seems to be shared by other P genotypes belonging to the P[II] genogroup, since structural comparison with recently reported glycan-bound structures of VP8* from human P[4] and P[6], and porcine P[19] genotypes showed modest differences (RMSD 0.51–0.84 Å) (S4C Table and S2 Table).
P[8]c VP8* binds H1 and LNB in a pocket formed by one of the β-sheets and the C-terminal α helix (Fig 6A and 6B). The structures showed that the LNB moiety is embedded in the pocked interacting with the protein while the L-fucose is projected out with minimal, mainly mediated by solvent, contacts with the protein (Fig 6A and 6B). Only seven residues recognize the LNB, contacting the N-acetyl-glucosamine moiety via L167, W174, T185, R209, and E212 and the galactose moiety via T184, T185, E212 and N216 (Fig 6C, 6D and S3 Table). This pocket shared identical interacting residues to the novel VP8* binding pocket recently discovered in P[19] [7] and P[4]/P[6] [12] genotypes for their interaction with LNFPI and it differs from the previously defined carbohydrate binding site in VP8* from P[11] (LacNAc binding [11]) and P[14] (A-antigen binding [9]) genotypes, that is located in the cleft between the two β sheets (S4D Fig). The N-acetyl-glucosamine has a major contribution to the binding since it is inserted in the pocket while the galactose acquires a more superficial position (Fig 6C and 6D). In this way, the N-acetyl-glucosamine ring is packet between W174 and R209 that define two faces of the binding site and the O4 and O6 oxygens mediated hydrogen-bonds with E212 that it placed at the bottom of the pocket (Fig 6C, 6D and S2 Table). To confirm the role of these amino acids in H1 and LNB interaction, the P[8]c VP8* mutants W174A (M1, VP8W174A), R209A (M2, VP8R209A) and E212A (M3, VP8E212A) were obtained (S1 Fig). The three mutant VP8* lost their ability to interact with their receptor when they were assayed for binding to H1 and LNB by SPR (S5 Fig), supporting the structural data and suggesting that the identified pocket is the only binding site for H1 antigen and its precursor in P[8] VP8*.
Fucosylation of the H1 precursor is genetically determined by the FUT2 gene (α1,2 fucosylation of the terminal galactose), resulting in different secretor status and by the FUT3 gene (α1,4 fucosylation at the precursor N-acetyl-glucosamine), defining the Lewis status (Fig 1). This genetically determined glycan profile is a susceptibility factor in human rotavirus [17, 18]. Our P[8]c VP8* structure in complex with H1 antigen confirms that the secretory L-fucose portion has reduced contacts with VP8* that are mediated by solvent molecules. Similarly, minimal or null contacts of the L-fucose moiety were observed in the complexes of LNFPI with other VP8* from the P[II] genogroup [7, 12]. Therefore, the difference in the affinity for H1 and LNB in diverse VP8* from this genotype were difficult to explain. A close view of the relative disposition of the precursor LNB moiety in both P[8]c VP8* complexes after superimposition of VP8* proteins showed some differences but not so for the interacting residues (Fig 7A). If, alternatively, the N-acetyl-glucosamine moieties from both structures are superimposed it becomes visible that the galactose moiety occupies a more solvent-exposed position in the LNB structure compared to H1 (Fig 7B), explaining its weaker binding in comparison with this antigen. The atomic resolution of the diffraction data and the high quality of the density maps derived from these data (Fig 7C and 7D) allows modeling accurately the ligand structures to observe these differences. This observation indicates that the L-fucose moiety induces an LNB conformation more suitable for the union to VP8*. The P[8]c VP8* structure in complex with H1 shows that the L-fucose ring stacks over the acetamido group of N-acetyl-glucosamine stabilizing the glycan conformation (Figs 6C and 7B). Therefore, the structures suggest that the L-fucose moiety could favor the VP8* binding by a double mechanism: inducing a competent conformation that facilities the LNB module recognition and binding and mediating indirect interactions that stabilize the glycan-VP8* complex.
Oppositely, the fucosylation in the N-acetyl-glucosamine in the LNB precursor to produce the Lewisb antigen seems to be incompatible with the binding to P[8] VP8*. Docking of the Lewisb antigen (Fuc-α1,2-Gal-β1,3-[Fuc-α1,4-]GlcNAc) in the P[8]c VP8* structure taking H1 antigen as a reference showed that the α1,4-linked L-fucose points towards the inside of the glycan binding pocket, clashing with different residues (mainly T185 and E212) that generate this cavity (S6A Fig). Since glycan binding pocket is highly conserved among P[8] VP8* linages and P[II] genotypes [7, 12], it seems that this group of rotaviruses are non-competent to bind Lewis HBGAs. This observation is confirmed by our ELISA assays where any of the VP8* proteins from P[II] genotype showed binding capacity to these HBGAs, and it questions previous results were P[8] VP8* was shown to interact with Lewisb [6]. Following an identical approach the A-type I antigen can be docked in the P[8]c VP8* sugar-binding pocket. The docked sugar showed that the N-acetylgalactosamine (GalNAc) added to the non-reducing end of the H1 galactose is pointing towards the solvent without showing any interaction with the protein (S6B Fig), supporting that P[8] VP8* should be able to bind this sugar.
The structures of VP8* in complex with LNB and H1 obtained here confirm that P[4], P[6], P[19] and the prevalent human genotype P[8] of rotavirus share a common glycan binding site which is highly conserved in structure and sequence. Therefore, the differences in affinity and specificity for HBGAs observed in our ELISA and SPR assays as well as those reported by others are striking. Particularly, the differences in the capacity to interact with the H1 antigen observed among P[8] lineages are difficult to understand. Inspection of the amino acids defining the glycan binding site in the four P[8] lineages oriented by the comparison of the P[8]c, P[8]Wa and P[8]Rotarix structures revealed interesting differences. First, the VP8* protein from Rotarix strain (lineage I) differed from other members of the same lineage, including the Wa strain, by the replacement of leucine 167 by phenylalanine. Since L167 is placed at the bottom of the sugar-binding pocket (Fig 6C and 6D), the introduction of a bulky Phe residue could explain the lack of interaction of this vaccine strain to H1 [19] that was corroborated in our study. Second, a close inspection of the P[8] structures revealed small differences in the disposition of the β-hairpin connecting the strands β9 and β10 (S7 Fig). These two β-strands define the bottom of the glycan binding pocket where the LNB moiety is settled (Fig 6C and 6D). These subtle movements are allowed by the presence of two Gly residues (G170 and G171) conferring flexibility to the loop. Y169 and R172 at both sides of the Glys delimit the loop and play a pivotal role in the correct organization of the binding pocket since Y169 stacks over R209 and R172 interacts with W174, the main residues recognizing N-acetyl-glucosamine (Fig 6C, 6D and S7 Fig). Therefore, the disposition of this loop could modulate de glycan affinity and specificity. The sequence analysis of this region among P[8] lineages reveals that at position 173 lineage I presents a Val whereas an Ile is found in the rest of lineages (S4 Table). Position 173 is placed in the base of the β-hairpin facing to the hydrophobic core of VP8* protein (Fig 6C, 6D and S7 Fig). Therefore, we wondered if this subtle change could influence the glycan-binding site architecture accounting for the difference in the H1 affinities observed between lineages. To test this possibility we designed a new mutant in which I173 in P[8]c VP8* was replaced by Val (M4, VP8I173V), emulating P[8]Wa. This replacement resulted in a VP8* with a diminished interaction with the H1 antigen, with an affinity constant (Kda = 39.0 ± 1.25 μM, S5 Fig, Table 1) 1.4-fold higher compared to the wild-type protein (Kda = 27.9 ± 0.71 μM, Table 1; p = 0.0002). The VP8I173V variant retained the binding ability to LNB with an apparent affinity constant (Kda = 49.4 ± 1.74 μM, Table 1) that was still higher than that of H1 (p = 0.0033). These results support that subtle amino acid changes at the loop close to the binding pocket may have contributed to modulate the glycan affinity between P[8] VP8* from lineages I and III. Differences at this site are also evident in the structures of VP8* from the P[II] genogroup. While conservation between P[8] and P[4] is high, the sequence divergence between P[4] and P[6] (S4 Table) might explain the different affinities for LNFPI between these two genotypes [12].
Virus-HBGAs interaction has emerged as an important factor in viral infectivity. Contrarily to other enteric viruses (i.e.: norovirus), the relevance of HBGA interaction in rotaviruses was first neglected, and virus-host cell attachment studies were mainly focused on binding to sialic acid, until interactions with HBGA were suggested by VP8* structural analyses [20] and experimentally determined in sialidase-insensitive strains [6]. In norovirus many studies point to the human FUT2 polymorphism as a key feature affecting viral infectivity [17, 18]. Individuals carrying two null FUT2 alleles lack fucosyl transferase-2 activity, do not express H antigen structures at the intestinal mucosa and in secretions (non-secretors) and are less susceptible to norovirus. While previous studies showed no correlation between the secretor status and rotavirus infection [17], the most recent studies show that antibody titers to rotavirus [13], rotavirus gastroenteritis incidence [14] and vaccine take [15] correlate with the FUT2 phenotype. However, the molecular mechanisms of these correlations were unknown until now. The previously reported interaction of H1 antigen (Fuc-α1,2-Gal-β1,3-GlcNAc) with the most common human rotavirus P genotype P[8] that has been further characterized here at the structural level, highlights the importance of the secretor phenotype on the incidence of rotavirus diarrhea. We have determined the characteristics of this interaction, acknowledging a new binding site for H1 in VP8* common for all the members of P[II] genogroup. Our results show that physical interaction between the H1 antigen and P[8] rotavirus occurs through the precursor side of the molecule (LNB), reinforcing the idea that the main carbohydrate-protein contacts are made via the N-acetyl-glucosamine moiety [7]. NMR studies on A-antigen binding of P[9] and P[14] VP8*, demonstrated that the L-fucose moiety does not make contacts with VP8* and rather it remains exposed to the solvent with a high degree of flexibility [21]. However, in the same study VP8* from genotypes P[4] and P[6], that did not recognize A-antigen in our assays, were shown to bind this antigen and L-fucose-protein contacts were evidenced [21]. Structural data from P[4] and P[6] VP8* in complex with LNFPI also showed a limited but direct interaction of the α1,2-linked L-fucose with the protein, namely via the R209 residue, which is conserved in all proteins from genogroup P[II] [12]. Due to the minimal interaction of the secretory L-fucose to VP8*, the authors of this study hypothesize that this glycan moiety has a low contribution to binding affinity and that a strong interaction would be expected for the unfucosylated H1 precursor, explaining the epidemiological studies that do not correlate the FUT2 status to infection by P[4] and P[6] genotypes [22]. Contrarily to this, we show that although the L-fucose moiety of H1 makes indirect contacts with P[8] VP8*, it stabilizes the competent conformation of the LNB moiety to interact with the sugar binding residues, resulting in two-fold lower Kda for H1 compared to LNB. This small but significative difference may be of relevance in the viral susceptibility context between secretors and non-secretor (FUT2-/-) individuals. Furthermore, a weaker interaction to LNB may also explain why infection of P[8] rotaviruses can occur, at a lower level, in non-secretor individuals [23] and it also accounts for the inhibitory effect of LNB in in vitro rotavirus infection reported here.
The previously reported interaction of the P[8] genotype with Lewisb identified by ELISA assays [6], could not be reproduced in our experiments. The structural evidence obtained here and in the analyses of the P[4] and P[6] structures in complex with LNFPI [12] argues against interaction with Lewisb, which differs with H1 in the presence of an extra L-fucose α1,4-linked to N-acetyl-glucosamine that generates steric hindrances to the interaction. These discrepancies, together with the differences found between ELISA and SPR for H1 binding in the different P[8] lineages, suggest that simple qualitative ELISA tests do not always provide reliable results and that other techniques need to be implemented in order to assess VP8* affinities for HBGAs. However, structural data from P[4] and P[6] genotypes predict interaction with A- and B-types HBGAs, as the N-acetyl-galactosamine (A-type) and galactose (B-type) located at the non-reducing ends in these glycans do not make any steric hindrance [12]. This coincide with our observation that P[8] VP8* from different lineages interact with blood group A trisaccharide and it is also supported by modelling additional N-acetyl-galactosamine in H1 bound to P[8]c VP8* (S6 Fig).
We showed that variations in the binding domain in the P[8] lineages exist that justify the differences in the affinity for H1 and LNB as measured by SPR. Additionally, even if the architecture of the binding site is similar for most P[8] lineages and other genotypes belonging to the P[II] group, other protein residues outside this site may possess epistatic effects over the capacity of the binding pocket to accommodate H1 and LNB, explaining the diverse affinities among P[II] genotypes and in the different P[8] lineages. This is exemplified by the fact that we were able to associate a residue that does not participate in direct protein-ligand contacts (Valine 173) in P[8]wa VP8* to its lower affinity for H1 and LNB and that VP8* from the vaccine strain Rotateq (lineage II), although sharing identical key binding residues to lineage III, do not bind H1 [19]. It is postulated that subtle changes in residues within and outside the defined pocket leads to a fine tuning in HBGA affinities that may ultimately impact host infection capacity. Epidemiologic studies revealed the occurrence of P[8] lineages I, II and III as the major circulating rotavirus with a prevalence of the lineage III [4, 24–26], while lineage IV is rarely found [25]. It is worth mentioning that this lineage, which is phylogenetically distant [4], carries some differences in the H1 binding pocket and showed low affinity to H1, has been isolated from few countries but it seems to be rapidly expanding [27]. Many studies have focused on the rotavirus genotypes circulating before and after the introduction of rotavirus vaccination programs [28, 29], but no work addressed the question of whether a link exists between the secretor status and the incidence of different P[8] lineages. In an study on the effect of the FUT2 status on rotavirus gastroenteritis it was shown that 100% of the patients (n = 51) were secretor positive compared to a healthy control or a group of non-rotavirus gastroenteritis patients (14–19% of non-secretors) and that all rotavirus involved were P[8] from lineage III [30]. Studies on the dynamics of G1P[8] rotavirus in a western population showed that ancient strains were Wa-like (lineage-I) and that new lineages emerged since late nineties [4], although the three main lineages are co-circulating nowadays in most geographical locations. Some authors have argued that the lack of (or low) interaction of Wa-like strains to H1 may help these viruses to prevail, because they do not discriminate by the secretor status [21], but our results suggest that only affinities for the receptor may be varying within the different P[8] lineages and it is not known how this may impact viral fitness. Notably, the amino acids comprising the H1/LNB binding pocket fall outside the defined epitopes in VP8* that elicit protection (S8 Fig). Mutations in VP8* which result in antigenic variants that could escape neutralizing antibodies are frequently isolated [24–26], but they are not affecting the H1/LNB interacting residues defined here. The impact of the different rotavirus P[8] lineages in the population depending on the secretor status deserves further studies in order to ascertain if the prevalence or co-circulation of different P[8] lineages responds to an adaptation to the HBGA profiles of the different hosts.
Despite the need for more exhaustive research on the relevance of HBGA and host specificity/infectivity in P[8] rotavirus, surface glycans possess a clear application in the development of antiviral strategies. It is established that human milk, in addition to other antiviral components, carry a set of oligosaccharides (human milk oligosaccharides; HMO) that share structural similarities to HBGA [31] and could act as anti-adhesins by competition with pathogen ligands at the mucosa. This blocking ability by soluble carbohydrates resembling rotavirus ligands has been evidenced. HMO were shown to inhibit binding of VP8* from P[6] and P[11] genotypes [32]; P[8] and P[4] genotypes infection is inhibited by the HMO 2'-fucosyllactose, 3'-sialyllactose and 6'-sialyllactose [33]; LNFPI inhibited infection of P[19], P[4], P[6] and P[8] genotypes [7], while we showed that LNB inhibited infection of Wa strain in vitro. The anti-adhesin potential of this simple disaccharide (LNB) is susceptible for being exploited in antiviral strategies. LNB is present in human milk in its free form [34] but mainly as a building block of type I HMO, which are predominant in human milk over type II HMO (based on LacNAc), which are characteristic of other mammals and primates [31]. Thus, LNB has been considered the human milk ‘bifidus factor’, and many bifidobacterial species from the infant gastrointestinal tract have the enzymatic machinery for its metabolism [35]. LNB would not only act as a bifidobacteria-stimulating prebiotic but also as a viral anti-adhesin to counteract rotavirus infection. Furthermore, its relatively simple synthesis, which can be undertaken enzymatically and by metabolic engineering approaches [36], makes this disaccharide a candidate for the development of new functional foods (e.g. improved infant formula). In this respect, it is important to consider that high affinity constants (in the mM range) have been determined for free oligosaccharides binding to VP8* [12, 21] and, in order to obtain good competitors, conjugated multivalent oligosaccharides seem to be a better option.
Detailed determination of the interactions between viruses and their host is crucial to develop appropriate antiviral strategies. We have defined the molecular interactions of P[8] VP8* from human rotavirus with its ligand HBGA giving a physical explanation as to why the secretor status influences rotaviral infectivity. Notwithstanding, extra structural elements beyond the identified binding site in VP8* are probably responsible for modulating HBGA interactions within P[8] lineages. Dissection of additional VP8* structural features affecting ligand binding is under way.
The VP8* (amino acids 64–224 from the VP4 protein of rotavirus) belonging to the P[4], P[6], P[8], P[9], P[11], P[14] and P[25] genotypes were cloned into the, pGEX-2T, expression vector (GE Healthcare) in order to express N-terminal GST fusions.
To amplify P[4], P[6], P[8], P[9] and P[14] VP8*s coding region, RNA was extracted from human stool samples (collected at Hospital Clínico Universitario de Valencia) containing rotavirus of known P genotype using the Trizol reagent following the standard procedure (Invitrogen). Viral RNA was retro-transcribed using the SuperScript Reverse Transcriptase (Invitrogen) and random-primers, and the cDNA was amplified by PCR using Pfu polymerase (Stratagene) with primers detailed in S5 Table. The cDNAs were finally cloned into pGEX-2T (GE healthcare) vector after digestion with BamHI (ThermoFisher). The VP8* genes from genotypes P[4], P[11] and P[25] were purchased as synthetic genes from Gene-ART technologies (ThermoFischer). The expression level of the VP8* protein from the clinical sample P[4] VP8* genotype was very low in E. coli and its codon usage was optimized. P[11] VP8* and P[25] VP8* were not available as clinical samples. The recombinants GST::VP8* proteins were expressed in E. coli BL21 (DE3) (Novagen) and purified by affinity chromatography using GSTrap columns coupled to an ÄKTA prime FPLC system (GE Healthcare). All sequences are included as fasta files in the supplementary data.
Selected residues in the GST::VP8* P[8] were replaced for alanine or valine according to the structural data of the LNB binding site. Four mutants (M1–M4) were constructed using a Quick-Change site-directed mutagenesis kit (Stratagene) and appropriate oligonucleotides (S5 Table), and the DNA changes were confirmed by DNA sequencing. Mutant M1, M2 and M3 contained changes in the codons for tryptophan 174, arginine 209 and glutamic 212 residues, respectively, that introduced an alanine at each position. Mutant M4 substituted isoleucine 173 by valine.
A panel of biotinylated sugar antigens including Lea, Leb, Lec (lacto-N-biose; LNB), H type-1, H type-2 and blood group A and B trisaccharides were purchased from Glyconz (Fig 1 and S1 Table). These glycans are biotinylated neoglyconjugates of a poly[N-(2-hydroxyethyl)acrylamide] (PAA) with a size from 30 to 50 KDa. This forms a flexible polymer ideal for a multivalent presentation of glycans. Immobilized streptavidin F96 black plates (Nunc) were coated with the biotinylated oligosaccharides (2 μg/ml) in milli-Q water and incubated during 1 hour at 37°C. After functionalization the plates were washed once with PBS containing 0.05% of Tween 20 (PBS-T) and the VP8* proteins were added (10 μg/ml) and incubated at 4°C overnight. After three washes with PBS-T, a rabbit polyclonal antibody anti GST (1:1,000) (Abcam) was added and the plates were incubated one hour at 37°C. Then, the plates were washed three times with PBS-T and incubated for 1 h at 37°C with 1:10,000 dilution of horseradish peroxidase (HRP)-conjugated goat anti-rabbit (Abcam). After three washes with PBS-T, the binding was detected using QuantaBlue reagent (ThermoFisher) kit, as recommended by the manufacturer. Fluorescence units were registered by a MultiScan microplate reader. All the binding assays were performed in triplicate.
The EC50 binding of the VP8* from the clinical (P[8]c) and Wa (P[8]Wa) genotypes to the H type-1 and to its precursor (LNB, Lec) was determined incubating two fold serial dilutions of the VP8* proteins, ranging from 100 μg/ml to 1.5 μg/ml. The binding assays were performed in triplicate using the protocol described above.
To confirm the binding of the VP8* from the P[8] genotype to the H type-1 precursor LNB, a blocking assay was performed using soluble LNB and its related disaccharide galacto-N-biose (Gal-β-1,3-GalNAc; GNB) produced and purified in our laboratory as previously described [36].
Streptavidin microtiter plates were coated with biotinylated H type-1 antigen or LNB at 2 μg/ml with water and incubated for 1 hour at 37°C, followed by an overnight incubation at 4°C. Blocking assays were performed in parallel, using glass tubes containing the P[8]c and P[8]Wa VP8* protein at their EC50 for each ligand and 20 mM of each of the soluble disaccharides (LNB and GNB) and monosaccharides (D-galactose, GlcNAc and GalNAc). A positive binding control without sugar was also included. The tubes containing the mixes of VP8* with sugars were maintained 1 hour at 37°C, followed by an overnight incubation at 4°C. The next day the coated streptavidin plates were washed with PBS-T, the VP8*-sugar solutions were added to the plates and incubated during 4 hours at 4°C. The plates were washed three times with PBS-T and detection of the interactions was performed as described above. The results are presented as the percentage (%) of binding of each condition compared to the binding of the positive binding control (without blocking sugar). All experiments were performed in triplicate.
African green monkey kidney epithelial cells (MA104 cell line; ATTC #CRL-2378.1) were used for the propagation of rotavirus Wa strain that belongs to the globally dominant human genotype G1P[8]. Briefly, ten MA104 cells confluent 150-cm2 flasks (approximately 1.5 ×107 cells/flask) were infected with Wa strain at a multiplicity of infection (MOI) of ≤ 0.1 and processed as previously described [37]. One hundred ml of medium with 1.5x108 virus/ml were obtained and the viral particles were concentrated by pelleting at 160,000 × g for 1 h at 4°C in a SW 41 rotor (Beckman). The viral pellet was resuspended in TNC buffer (20 mM Tris-HCl pH 8.0, 100 mM NaCl, 1 mM CaCl2) for triple-layered particles (TLP) or in TNE (20 mM Tris-HCl, pH 8.0, 100 mM NaCl, 1 mM EDTA) for double-layered particles (DLP).
An ELISA-like binding assay was employed to determine the binding ability of rotavirus TLP and DLP to the H type-1 antigen and to its precursor LNB. Streptavidin plates were coated with the biotinylated oligosaccharides as described above. After washing with PBS-T, two fold serial dilutions of TLPs and DLP were added to the plate (ranging from 10 μg/ml to 0.078 μg/ml). The TLP were always maintained in TNC-T (20 mM Tris, 100 mM NaCl, 1 mM CaCl2, 0.05% Tween 20, pH 7,4) buffer and the binding and washing steps were always carried out in this solution. DLPs assays were carried out in Tris-buffered saline buffer with Tween 20 (TBS-T, 20 mM Tris, 100 mM NaCl, 0.05% Tween 20, pH 7.4). TLP and DLP were incubated in the plate overnight at 4°C. After binding, the plates were washed three times in TNC-T (for TLP) or TBS-T (for DLP) with 0.05% Tween 20 (TNC-T and TBS-T), and a mouse anti-VP6 antibody was added at 1:100 in TNC-T or TBS-T and incubated 1 h at 37°C. The plates were then washed three times with TNC-T or TBS-T, and a HRP-conjugated anti-mouse IgG was added at 1:10.000 and incubated at 37°C for 1 h. After three final washes the binding was revealed by QuantaBlue reagent (ThermoFisher) following the manufacturer recommendations. Fluorescence units were recorded by a MultiScan microplate reader.
The G1P[8] Wa strain was tested on MA104 cells. The sugars LNB, GNB, GlcNAc, GalNAc, D-galactose and L-fucose were tested for their effect on rotavirus infectivity. The oligosaccharides were previously heat sterilized at 99°C for 10 min and then dissolved in serum-free DMEM containing 1 μg/ml trypsin. Serum-free DMEM containing 1 μg/ml without oligosaccharide was used as a control in each experiment.
The effect of different mono- and disaccharides on rotavirus infectivity was assessed through standard fluorescent focus assays on MA104 cells [37]. The dilution of Wa virus stocks that yielded ~150 focus-forming units/well was first established. Then, sugars were added during virus inoculation at a final concentration of 5 mg/ml, incubated for 1 hour, and unbound virus was removed by washing with FBS-DMEM. The cells were allowed to be infected for 16h, washed once with PBS and fixed with 100% methanol. A mouse anti-VP6 primary antibody (1:50 dilution in PBS containing 3% BSA) was added and incubation proceeded for 30 min at room temperature with gentle rotation. A secondary antibody anti-mouse IgG-FITC (Sigma F4143) diluted 1:128 in PBS containing 3% BSA was added and incubated for 30 min at room temperature with gentle rotation. Individual fluorescence foci were counted on an inverted fluorescence microscope with a FITC-compatible filter. Infectivity in the absence of oligosaccharides served as the control. Each experimental condition was tested a minimum of 2 times, with technical triplicates for each oligosaccharide. The means and SD from a minimum of 6 determinations are represented for each condition. Virus titer measured in the absence of oligosaccharides was considered to be 100% infectivity, and changes in virus titer in the presence of sugars were expressed as percentage of infectivity compared with no sugar treatment.
The affinity assays were based on SPR and performed in a Biacore T100 instrument (GE Healthcare). H1 PAA-biotin and LNB PAA-biotin were diluted to a concentration of 1 mg/ml in water and captured with streptavidin present in a SA sensor chip (GE Healthcare). H1 was immobilized in channel 2 (630 RU) of the sensor chip and LNB was immobilized in channel 4 (624 RU). The channels 1 and 3 were used as the reference surfaces for channels 2 and 4, respectively. The immobilization process was performed by conditioning the sensor chip surface with three consecutive 1-minute injections of 1 M NaCl 50 mM NaOH before biotinylated ligands were immobilized at a flow rate of 15 μl/min. The affinity assays of VP8* polypeptides to biotinylated sugars were performed at 10°C using 1X HBS-EP+ buffer (0.01 M HEPES pH 7.4, 0.15 M NaCl, 3 mM EDTA, 0.005% Surfactant P20), a flow rate of 5 μl/min with 2700 seconds of contact time and a dissociation time of 1800 seconds. The regeneration step consisted in a wash step with 10 mM Glycine-HCl pH 2 for 20 seconds at the same flow rate. The assays were performed with purified VP8* at different concentrations (45; 137; 411; 1,234; 3,703; 11,111; 33,333; 100,000 and 200,000 nM). Each run included three blanks without sample. The affinity data were obtained after analysis of sensorgrams performed with the BIAevaluation 2.0 software (GE Healthcare). Since multivalent oligosaccharides are immobilized on a sensor chip surface, avidity and rebinding effects can take place and apparent affinity constants (Kda) are calculated with this experimental setup. Kda values were obtained from the steady-state kinetics experiment as the ligand concentration needed to achieve a half-maximum binding at equilibrium. The experiments were made in triplicate. Graphical representation of signal/concentration curves were plotted using GraphPad Prism 6 for MacOsX.
The crystals were grown as hanging drops at 21°C with a vapour-diffusion approach. Initial crystallization trials were set up in the crystallogenesis service of the IBV-CSIC using commercial screens JBS I, II (JENA Biosciences) and JCSG+ (Molecular Dimensions) in 96-well plates. Crystallization drops were generated by mixing equal volumes (0.3 μl) of P[8]c VP8* protein solution and the corresponding reservoir solution, and were equilibrated against 100 lμl reservoir solution. Both P[8]c VP8* Apo structures were crystallized at 10 mg/ml. VP8* Apo1 was crystallized in a reservoir solution of 1.2 M (NH4)2SO4, 3% iso-propanol and 0.1 sodium citrate pH 4.6, whereas VP8* Apo2 was crystallized in 1.5 M Li2SO4 and 0.1 M Tris-HCl pH 6.5. In both cases 2 M Li2SO4 was used to cryoprotect the crystal when freezing in liquid nitrogen. For the crystallization in presence of glycans, the ligands were mixed with the protein at 10 mM final concentration of ligand and 10 mg/ml of protein final concentration. P[8]c VP8* LNB was crystallized in a reservoir solution consisting in 25% PEG 3,350 0.1 M Bis-Tris pH 5.5. The cryosolution used for crystal freezing was its reservoir solution increased up to 35% PEG 3,350. P[8]c VP8* H1 was crystallized against a a reservoir solution consisting in 25% PEG 6,000, 0.1 M Na-HEPES pH 7.5, 0.1 M LiCl, and PEG 6,000 was increased up to 35% for cryoprotection. X-ray diffraction was carried out at 100K at Alba (Cerdanyola, Barcelona, Spain) and DLS (Didcot, UK) synchrotrons and the best data sets used to solve the structures were collected at the indicated beamlines and wavelengths (Table 2). Diffraction data was processed and reduced with Mosflm[38] and Aimless[39] programs from the CCP4 suite [40]. The data-collection statistics for the best data sets used in structure determination are shown in Table 2.
P[8]C VP8* Apo1 structure was solved by molecular replacement carried out with the program Phaser [41] and using the structure of VP8* from CRW-8 porcine rotavirus (PDB 2I2S[20]) as a model. Initial phases from the molecular replacement were used to manually build the P[8]c VP8* structure with Coot [42]. P[8]c VP8* Apo1 structure was then used as a model for molecular replacement to solve the P[8]c VP8* Apo2, P[8]c VP8*LNB and P[8]c VP8*H1 structures. All the final models were generated by iterative cycles of refinement using the Refmac [43] and manually optimization with Coot. Data refinement statistics are given in Table 2. The crystals exhibited good quality control parameters and excellent stereochemistry. Atomic coordinates and structure factors have been deposited in the Protein Data Bank (PDB) with ID numbers 6H9W, 6H9Z, 6H9Y and 6HA0 for P[8]c VP8* Apo1, P[8]c VP8* Apo2, P[8]c VP8*LNB and P[8]c VP8*H1, respectively. Structure Superposition and RMSD calculations were carried out with Superpose [44] from CCP4 suite.
To assess statistical differences in the ELISA-like binding experiments where many groups were compared an ANOVA test was performed. To analyze significative differences in the Kda values obtained by SPR an unpaired t-test was applied. All statistical analyses were performed with GraphPad Prism version 6.0 for MacOsx (GraphPad Software). p values <0.05 were considered to be statistically significant.
This study was conducted with the approval of the Ethics Committee of the University of Valencia (code H1544010468380).
The human stool samples from Hospital Clínico Universitario de Valencia were anonymized previously to their inclusion in the present study.
|
10.1371/journal.pntd.0003100 | HIV-Associated Histoplasmosis Early Mortality and Incidence Trends: From Neglect to Priority | Histoplasmosis is an endemic fungal infection in French Guiana. It is the most common AIDS-defining illness and the leading cause of AIDS-related deaths. Diagnosis is difficult, but in the past 2 decades, it has improved in this French overseas territory which offers an interesting model of Amazonian pathogen ecology. The objectives of the present study were to describe the temporal trends of incidence and mortality indicators for HIV-associated histoplasmosis in French Guiana.
A retrospective study was conducted to describe early mortality rates observed in persons diagnosed with incident cases of HIV-associated Histoplasma capsulatum var. capsulatum histoplasmosis admitted in one of the three main hospitals in French Guiana between 1992 and 2011. Early mortality was defined by death occurring within 30 days after antifungal treatment initiation. Data were collected on standardized case report forms and analysed using standard statistical methods.
There were 124 deaths (45.3%) and 46 early deaths (16.8%) among 274 patients. Three time periods of particular interest were identified: 1992–1997, 1998–2004 and 2005–2011. The two main temporal trends were: the proportion of early deaths among annual incident histoplasmosis cases significantly declined four fold (χ2, p<0.0001) and the number of annual incident histoplasmosis cases increased three fold between 1992–1997 and 1998–2004, and subsequently stabilized.
From an occasional exotic diagnosis, AIDS-related histoplasmosis became the top AIDS-defining event in French Guiana. This was accompanied by a spectacular decrease of early mortality related to histoplasmosis, consistent with North American reference center mortality rates. The present example testifies that rapid progress could be at reach if awareness increases and leads to clinical and laboratory capacity building in order to diagnose and treat this curable disease.
| Histoplasmosis is an endemic fungal infection in French Guiana. It is the most common AIDS-defining illness and the leading cause of AIDS-related deaths. Diagnosis is difficult, but in the past 2 decades, it has improved. The objectives of the present study were to describe the temporal trends of incidence and mortality indicators for HIV-associated histoplasmosis in French Guiana. A retrospective study was conducted to describe early mortality rates observed in persons diagnosed with incident cases of HIV-associated histoplasmosis admitted in one of the three main hospitals of French Guiana between 1992 and 2011. Early mortality was defined by death occurring within 30 days after antifungal treatment initiation. Data were collected on standardized case report forms and analysed using standard statistical methods. Among 274 patients there were 46 early deaths (16.8%). The two main temporal trends were: the proportion of early deaths significantly divided four fold and the number of annual incident histoplasmosis cases increased three fold. The present example testifies that rapid progress could be at reach if awareness increases and leads to clinical and laboratory capacity building in order to diagnose and treat this curable disease.
| French Guiana is a French overseas territory, located in the North-Eastern part of South America. The Human Immunodeficiency Virus (HIV) epidemic there is the most preoccupying among French territories [1]. During the Highly Active AntiRetroviral Therapy (HAART) era, disseminated histoplasmosis has remained the most common Acquired Immunodeficiency Syndrome (AIDS) defining illness with an incidence of 15.4/1000 person-years in HIV-infected patients [2].
In immunocompetent patients, Histoplasma capsulatum var. capsulatum infection is typically asymptomatic or pauci-symptomatic and spontaneous resolution is the rule in the great majority of cases [3]. On the contrary, in HIV-infected patients it presents mostly as a disseminated infection. With the worsening of the immunosuppression, the disease progression is often rapid and always fatal in the absence of treatment [4].
Thus, different studies have observed up to 39% of deaths following diagnosis in endemic areas, where it is supposedly well known, and 58% in non endemic areas, where it is perhaps less known [5], [6]. In endemic areas, although there are different outcome measures and inclusion criteria, the death rates observed in AIDS-associated histoplasmosis differ between the USA (12–23%) and South America (19–39%) [6]. Hypotheses advanced to explain these differences are a delayed recognition due to the lack of awareness of physicians, a delayed diagnosis due to the lack of diagnostic facilities and the late presentation of HIV-infected patients in resource limited settings [6], [7], [8]. Delayed treatment due to the unavailability of the most effective therapy in severe cases, the impossibility of monitoring drug concentrations and/or drug-drug interactions with antituberculosis treatments are other possible explanations [6].
In French Guiana, disseminated histoplasmosis has also been the leading cause of death among HIV-infected patients [9]. Despite HIV care and treatment standards close to those in Mainland France, the mortality rate of AIDS-associated histoplasmosis remains high in the HAART era (30.7% at 6 months and 17.5 at 1 month), whereas in Mainland France, a non-endemic area, this mortality rate was divided by two [10], [11].
The objective of this study was to describe the temporal trends of incidence and mortality indicators for AIDS-associated histoplasmosis in French Guiana. This knowledge is important to guide and improve AIDS-associated histoplasmosis diagnosis, care and treatment, and to illustrate that awareness and standard practices in mycology can dramatically change prognosis.
Since 1992, an anonymized database compiles retrospectively and continuously Histoplasma capsulatum var. capsulatum histoplasmosis confirmed incident cases diagnosed in HIV-infected patients according to the case definition of the European Organization for Research and Treatment of Cancer/Invasive Fungal Infections Cooperative Group and the National Institute of Allergy and Infectious Diseases Mycoses Study Group (EORTC/MSG) Consensus Group [12]. The revised EORTC/MSG criteria defining a proven case of histoplasmosis are: recovery in culture from a specimen obtained from the affected site or from blood; and/or histopathologic or direct microscopic demonstration of appropriate morphologic forms with a truly distinctive appearance characteristic such as intracellular yeasts forms in a phagocyte in a peripheral blood smear or in tissue macrophages. By contrast, molecular methods of detecting fungi in clinical specimens, such as Polymerase Chain Reaction (PCR), were not included in the classifications of “proven,” “probable,” and “possible” invasive fungal disease (IFD) definitions because there is as yet no standard, and none of the techniques has been clinically validated.
All HIV-infected patients hospitalized or seen in the outpatient department before admission, suspicious for histoplasmosis and receiving antifungal therapy in one of the three main hospitals of French Guiana (the Centre Hospitalier de Cayenne (CHC), the Centre Hospitalier Médico-Chirurgical de Kourou (CMCK) and the Centre Hospitalier de l'Ouest Guyanais in Saint Laurent du Maroni (CHOG), were identified and checked for a confirmed diagnosis of histoplasmosis in all laboratories where biological samples were sent. Then, they were finally enrolled according to the following inclusion criteria: age >18 years, admission in one of the three hospitals (the inclusion date corresponding to the date of antifungal treatment initiation), confirmed HIV infection (by Western blot), confirmed incident histoplasmosis infection (EORTC/MSG criteria), and baseline blood screening within 7 days prior to antifungal therapy initiation. Non inclusion criteria were: histoplasmosis relapse or diagnosis of histoplasmosis relying only on Histoplasma Polymerase Chain Reaction (PCR). Data were collected on a standardized form and included sociodemographic, clinical, biologic, radiologic, therapeutic and survival information. These data were then entered in an anonymized database. Ethical approval was obtained for the database and related studies (IRB0000388, FWA00005831). A descriptive study of the patients included in this database until April 2007 was published elsewhere [10].
An observational, retrospective and multicentric study was conducted from 01/01/1992 to 09/30/2011, using the French Guiana HIV-Histoplasmosis database described above.
In this study, the primary endpoint was the vital status on day 30 following antifungal therapy initiation. Patients lost to follow up within 30 days following antifungal therapy initiation, or deceased with an unknown date of death, or presenting a relapse of histoplasmosis were excluded from the analysis.
This early death criterion appeared as a good compromise to attribute mortality to the histoplasmosis infectious episode under consideration, in a context of severe immunosuppression favouring multiple opportunistic pathogens, ensuring simplicity and reproducibility of the study.
The statistical analysis was performed using STATA 10.0 (College Station, Texas, USA) (38). Descriptive analysis used proportions, medians and trend χ2 test.
There were 278 patients with AIDS-associated histoplasmosis. Four cases were excluded before the analysis (3 because they were lost to follow up and one because of an unknown date of death). Their socio-demographic characteristics and median CD4 count did not differ from the 274 patients finally selected in this study (data not shown).
Among the 274 patients selected for whom the vital status at 30 days after antifungal therapy initiation was known, there were 124 deaths (45.3%). The median time to death was 110 days (Interquartile Range [IQR] = 13–481) and the median age at the time of death was 39 years (IQR = 33–47). Early death occurred in 46 patients (16.8%) with a median survival time of 7 days (IQR = 3–16) after antifungal treatment initiation. The median age at the time of early death was 37 years (IQR = 32–47).
Figure 1 shows that the proportion of deaths occurring the same year as the diagnosis of incident histoplasmosis cases remained stable around 5 deaths per year until 2005/2006 and then stabilized around 3 deaths per year. Among these deaths cases, almost half were early deaths until 2004. From 2005 onwards there was a notable decline of early deaths along with the overall decline of mortality. In addition, starting in 1998, the number of histoplasmosis cases diagnoses increased, and subsequently the number of incident cases oscillated between 14 and 22 cases per year. Data were incomplete for 2011, the study considering cases only until 09/30/2011.
Thus, three time periods of particular interest have been identified: 1992–1997, 1998–2004 and 2005–2011. Figure 2 summarizes the two main temporal trends observed in Figure 1. First, the proportion of early deaths among annual incident histoplasmosis cases was significantly divided four fold (χ2, p<0.0001). Second, the number of annual incident histoplasmosis cases increased three fold between 1992–1997 and 1998–2004, and subsequently stabilized at the same level.
Table 1 showed that early deaths associated with histoplasmosis occurred mainly in men, late presenters with HIV infection (CD4 count <50/mm3) among whom 10% were on HAART on admission. The incident histoplasmosis cases were mainly disseminated and often recognized as the first AIDS-defining illness in the course of HIV infection. Fungal culture and direct examination were the main methods used for the diagnosis of histoplasmosis cases. The Real Time Polymerase Chain Reaction (RT-PCR) detection method for Histoplasma only became available during the 2005–2011 period. Amphotericin B and itraconazole were the first line antifungal regimen used to treat these patients. During the study period, liposomal amphotericin B and itraconazole became the standard antifungal regimen over deoxycholate amphotericin B and fluconazole, respectively.
This study described 19 years of experience in French Guiana. Three periods of interest and two main trends could be observed from 1998 onwards: the spectacular decrease of early deaths among incident histoplasmosis cases, and a simultaneous marked increase of the annual incidence of histoplasmosis cases. Whereas, during the same period, HIV prevalence in pregnant women was quite stable >1% since the 1990's: 0.8%–1.4% between 1992–1997, 1.2%–1.4% between 1998–2004 and 1.0%–1.2% between 2005–2011 [1], [13].
The increased number of annual histoplasmosis cases can be attributed to the development of medical mycology skills in hospitals laboratories, notably a reference university laboratory specialized in parasitology-mycology established since 1997 in Cayenne Hospital. By the same time, highly active antiretroviral therapy was introduced, which could have led to more patent cases of histoplasmosis due to the immune reconstitution inflammatory syndrome [14]. In addition, a PCR diagnostic method became available for histoplasmosis in 2006 [15]. Unfortunately, urinary antigen detection for histoplasmosis is still unavailable in French Guiana.
The sharp decline of the proportion of early deaths can be attributed to the improvement of the diagnostic capacity along with the improvement of the clinical management of HIV-infected patients following French recommendations [16]. Thus, French Guiana reached HIV-virological suppression levels comparable to those in Mainland France by 2004. In addition, this trend can also be attributed to the improvement of the clinical management of AIDS-related disseminated histoplasmosis cases. The accurate recognition of severe cases and the supply of liposomal amphotericin B since 1998, an effective and less nephrotoxic treatment recommended for severe disseminated histoplasmosis cases, were two important factors behind the progress.
This study had limitations. Data were collected retrospectively, which might have led to selection biases. Determining retrospectively if death was related to AIDS-associated histoplasmosis incident cases under study is challenging, considering the high percentage of concomitant opportunistic infections. Thus, we chose early death as the primary outcome because we thought that retrospectively it was the simplest and most reproducible indicator of histoplasmosis AIDS-related deaths.
Despite its limitations, this study showed that capacity building both in laboratory and clinical practice, effective drug availability both for HIV and histoplasmosis infections, and an effective bench to bed collaboration between actors progressively helped in reducing the burden of overall deaths and early deaths. Mortality indicators are now consistent with those described in North America, where the most effective and non invasive histoplasmosis diagnostic method is available. To further reduce early mortality, reducing diagnostic delays and antifungal therapy initiation is still a major objective. To reach it, a diagnostic method that meets the World Healh Organization's A.S.S.U.R.E.D. (Affordable, Sensitive, Specific, User-friendly, Rapid/Robust, Equipment-free and Delivered) should be developed.
Although our results may seem parochial, they illustrate the rapid progress that took place within a decade. The increased awareness of clinicians, who became more aggressive in their investigations, and the increased laboratory capacity led to find and treat a disease that was present but probably not identified and not treated in time. Thus, histoplasmosis, previously known as a mild disease in immunocompetent individuals, became a public health problem in HIV-infected patients, known by almost all health practitioners in French Guiana. By dealing with the mycology diagnostic tool box limitations and starting prompt presumptive antifungal treatment in HIV-infected patients it was possible to reduce early deaths considerably.
The historical 40% of early deaths observed in French Guiana, where histoplasmosis was known, plausibly reflects a low estimate of what happens in the Amazon region and probably beyond, where histoplasmosis is endemic but probably still widely misdiagnosed for tuberculosis and/or neglected [17]. Although cost effective strategies to prevent the disease and very effective diagnostic methods have been developed and are well known by scattered medical teams in Latin America [18], this knowledge does not percolate to too many HIV care units and hospital laboratories [19].
The present example testifies that rapid progress could be at reach if awareness increased and led to implement clinical and laboratory capacity building in order to diagnose and treat this curable disease before it is too late.
|
10.1371/journal.pntd.0002526 | Population Genetics of Trypanosoma brucei rhodesiense: Clonality and Diversity within and between Foci | African trypanosomes are unusual among pathogenic protozoa in that they can undergo their complete morphological life cycle in the tsetse fly vector with mating as a non-obligatory part of this development. Trypanosoma brucei rhodesiense, which infects humans and livestock in East and Southern Africa, has classically been described as a host-range variant of the non-human infective Trypanosoma brucei that occurs as stable clonal lineages. We have examined T. b. rhodesiense populations from East (Uganda) and Southern (Malawi) Africa using a panel of microsatellite markers, incorporating both spatial and temporal analyses. Our data demonstrate that Ugandan T. b. rhodesiense existed as clonal populations, with a small number of highly related genotypes and substantial linkage disequilibrium between pairs of loci. However, these populations were not stable as the dominant genotypes changed and the genetic diversity also reduced over time. Thus these populations do not conform to one of the criteria for strict clonality, namely stability of predominant genotypes over time, and our results show that, in a period in the mid 1990s, the previously predominant genotypes were not detected but were replaced by a novel clonal population with limited genetic relationship to the original population present between 1970 and 1990. In contrast, the Malawi T. b. rhodesiense population demonstrated significantly greater diversity and evidence for frequent genetic exchange. Therefore, the population genetics of T. b. rhodesiense is more complex than previously described. This has important implications for the spread of the single copy T. b. rhodesiense gene that allows human infectivity, and therefore the epidemiology of the human disease, as well as suggesting that these parasites represent an important organism to study the influence of optional recombination upon population genetic dynamics.
| Trypanosomes are single-celled organisms transmitted by the biting tsetse fly, which cause sleeping sickness in humans in sub-Saharan Africa, but also infect livestock and other mammals. Most trypanosomes cannot infect humans as they die in human serum, but two mutants of Trypanosoma brucei have evolved the ability to survive in human serum. This survival in human serum is conferred by the presence of one gene in the East African human-infective T. b. rhodesiense. How often trypanosomes exchange genetic material (they can mate in the tsetse fly) is debated, but will impact upon the spread of genes (e.g. that which confers human infectivity) through a population. We studied T. b. rhodesiense populations from different geographic locations (Malawi and two locations in Uganda), and over time (Uganda), to see if the populations are stable over time and space, using a panel of variable genetic markers enabling assessment of diversity. Our results suggest that there is significant difference in diversity between locations; those in Uganda are very closely related, increasingly so over time, whereas the Malawi population is very genetically diverse, consistent with the trypanosomes mating. These findings suggest that a greater understanding of T. b. rhodesiense population evolution will inform on sleeping sickness epidemiology.
| Pathogens that can adapt quickly to environmental change often pose the greatest challenge to disease control. A clear example of this is the generation of drug resistance and subsequent rapid spread through a population [1]. The means and dynamics by which any trait spreads will depend upon the population structure and the level of recombination of the organism within individual populations. Therefore, understanding the population genetic dynamics of a pathogen and how often they share and disseminate genetic material is an important component in the development of risk assessment and intervention strategies.
The evolutionary potential of pathogen populations is a product of a number of factors, including the system of reproduction, the potential for gene flow, the effective population size and the mutation rate. Protozoan parasites offer a particular analytic challenge in this regard as many have complex life cycles in both vector and host, with some life cycle stages that expand mitotically and others in which sexual recombination occurs, resulting in mixed reproductive systems. Analyses of pathogenic protozoan populations reveal that there is significant diversity between different species and populations of the same species in terms of the role of genetic exchange, with some species showing clear clonality [2]–[4], while others demonstrate epidemic or panmictic populations. It is likely that the degree of recombination is dependent on local epidemiological factors [5]–[7]. Comprehensive analyses of multiple populations have been carried out for the malaria parasite, Plasmodium falciparum, which undergoes both asexual reproduction and an obligate sexual component of the life cycle, including out-crossing and self-fertilization. As sexual reproduction occurs in the insect vector, the frequency of out-crossing is a consequence of the transmission intensity, thus differences in transmission can result in a spectrum of population structures ranging from effective clonality (due to extensive self fertilization) to panmixia [8]. Thus there is a complex interaction between the epidemiology of the vector, host and parasite that influences the reproductive potential of the parasite. The Plasmodium research demonstrates that sampling from a range of epidemiological situations is necessary to evaluate the role of recombination in shaping the population genetic structure of a particular parasite species.
While mating in Apicomplexan parasites is an obligatory part of their life cycle in the arthropod vector, this is not the case with African trypanosomes. This issue is probably central to the controversy that has surrounded the definition of population structure and the role of mating in natural populations of the zoonotic protozoan parasite, Trypanosoma brucei [3], [9]–[11]. T. brucei is transmitted by tsetse flies (Glossina spp.) and in humans two subspecies, T. b. rhodesiense and T. b. gambiense, cause the often-fatal disease Human African Trypanosomiasis (HAT), also known as Sleeping Sickness. Sexual recombination in T. brucei occurs in the tsetse fly salivary glands and is well characterised under laboratory conditions [12]–[16]. Laboratory analysis has provided robust evidence that alleles segregate in a Mendelian manner [17] and the available data support the occurrence of both cross- and self-fertilisation [18], [19]. However, mating is not obligatory and does not happen with every transmission through a tsetse fly [20]. Thus, the parasite has the capacity for both clonal propagation with no sexual recombination, and also sexual propagation with varying degrees of inbreeding or out-crossing. This means that ‘clonality’ with respect to trypanosomes can be considered in two ways – that of classical mitotic clonality in the absence of sexual recombination [21], and the ‘reproductive clonality’ as has been observed in malaria parasites that undergo obligatory sexual recombination but in areas of both high and low transmission can undergo extensive inbreeding [22]–[24].
Initial isoenzyme analysis of T. brucei isolates from tsetse flies in East Africa indicated a panmictic or randomly mating population structure [9]. This interpretation was subsequently contested when high levels of linkage disequilibrium, lack of agreement with Hardy-Weinberg and the occurrence of identical genotypes at high frequency suggested either a clonal population structure where genetic exchange was very infrequent [2], [3], [25], or an epidemic population structure where there is a background level of frequent sexual recombination with the occasional clonal expansion of a few particular genotypes [26]. However, the interpretation of clonality is difficult with respect to trypanosomes, and counterarguments have centred on the existence of population sub-structuring, due either to geography or host specificity [27]. Genotype bias provided by the amplification of parasites in vitro or in vivo prior to analysis has also been suggested as another possible reason for the departures from expected genotype or allele frequencies [27], [28] and indeed this has been shown to occur [29]–[31].
An additional confounding factor for the study of T. brucei population genetics is that T. brucei consists of three morphologically identical sub-species. T. b. brucei cannot infect humans but causes disease in a wide range of domestic and wild animals, whereas T. b. gambiense is responsible for HAT in West and Central Africa, a chronic disease, and T. b. rhodesiense causes HAT in East and Southern Africa, typically a more acute disease. T. b. gambiense has been subdivided into two groups consisting of a homogeneous group 1 and a less common more heterogeneous group 2 [32]. Domestic and wild animals have been implicated as reservoirs of both human infective sub-species [33]–[35]. Several early studies failed to distinguish between the three sub-species and treated them as a single population, which may explain the detected high level of linkage disequilibrium [2], [3], [25]. From all available data it seems clear that T. b. gambiense group 1 is a clonal organism that undergoes sexual recombination very rarely, if at all [36], [37]. Indeed, T. b. gambiense group 1 is clearly genetically distinct from both T. b. brucei and T. b. rhodesiense [38]–[40]. Microsatellite analysis of 27 T. b. rhodesiense isolates from a range of foci in East and Southern Africa has shown that while isolates from different foci are broadly similar to each other, there is an association of the genotypes with their geographical origin [39]. However, the detailed analysis of the genetic structure within a single focus has not been studied with such markers. Although T. b. rhodesiense is genetically very closely related to T. b. brucei [40]–[42], it is not clear whether genetic exchange occurs in T. b. rhodesiense populations. The basis of human infectivity in T. b. rhodesiense has been understood for some time, and is due to the expression of a single gene, the serum resistance associated (SRA) gene [43]. By using SRA as a marker, the detection of T. b. rhodesiense parasites in non-human hosts has become more straightforward [34], [44], [45]. The genotyping of parasites isolated from foci of human disease have led to the conclusion that T. b. rhodesiense is clonal [10], [46], suggesting that a few parasite genotypes carrying the SRA gene amplified in the human population, resulting in an epidemic clonal expansion. However, these genotypes were also stable over time [10], suggesting that T. b. rhodesiense was not mating with the genetically more diverse sympatric T. b. brucei population, within which evidence for frequent mating was demonstrated. However it is clear that, unlike T. b. gambiense group 1, there do not seem to be biological barriers to T. b. rhodesiense mating with T. b. brucei, as this has been demonstrated in the laboratory in two separate crosses with different T. b. brucei strains [47], [48]. The disparity between laboratory and field data suggests that it is important to analyse further foci of T. b. rhodesiense and so examine populations in different epidemiological settings in order to rigorously address the question of clonality in this human infective sub-species. This will also allow a series of questions to be addressed, such as whether T. b. rhodesiense HAT foci in different geographical regions display similar levels of clonality; whether different foci are genetically distinct from each other, as well as from local T. b. brucei populations; and whether clonal populations of T. b. rhodesiense are stable over space and time.
To clarify our understanding of T. b. rhodesiense populations, we have employed microsatellite markers to determine allelic variation and multilocus genotypes from parasites isolated from three different foci of disease in East Africa, two in Uganda, and one in Malawi. The microsatellite loci were selected from a panel of genome wide markers, which had been used in the construction of the first genetic map of the parasite [16]. We have avoided ascertainment bias by employing a whole genome amplification technique on bloodspots taken directly from infected individuals [49] for all samples collected after 2001, allowing direct assessment of parasite populations by multilocus genotyping. These tools and approaches will allow us to address the following questions; (1) are different foci of T. b. rhodesiense genetically distinct? (2) Are the population structures and the role of genetic exchange similar in different foci? (3) By analysing samples over a period of 45 years from in and around the clonal Tororo focus, are the multilocus genotypes stable over time?
This study was conducted according to the principles expressed in the Declaration of Helsinki. All patients recruited received written and verbal information explaining the purpose of this study and gave informed written consent. All protocols were approved by ethics committees in Uganda (Uganda Ministry of Health) and Malawi (Malawi College of Medicine) as appropriate. Furthermore, the protocols, information forms and consent forms were reviewed and approved by the Grampian Research Ethics Committee (Aberdeen, UK). Ethical consent forms were designed in English and also translated into local languages. Consent was given as a signature or a thumb print after verbal explanation. For those under 16 years of age consent was given by their legal guardian, and for those whose clinical condition prohibited full understanding of the recruitment process, consent was gained from a spouse or other family member.
HAT patients presenting to local hospitals or identified during community surveillance were recruited in South-Eastern Uganda in 2002 and 2003 from an extensive focus of T. b. rhodesiense transmission covering the Tororo, Iganga, Jinja and Busia districts [50]. This will be referred to henceforth as the Tororo focus. The second focus sampled was Soroti, where HAT emerged as a new epidemic in 1998/1999 [51], which was sampled in 2003. During this study we examined 30 samples from the Tororo focus and 88 from the Soroti outbreak. These samples were compared with 52 previously isolated and described samples (from both humans and cattle) collected from the Ugandan/Kenyan border region (including Tororo, Busia, Iganga and Jinja districts in Uganda, and Busia and Nyanza districts in Kenya), covering a period of 36 years (1961 to 1997) prior to the more recent outbreaks in Tororo and Soroti. This set of samples will be referred to as ‘Ug/Ke 61–97’ (for sample details see Table S1). These samples provide a representative snapshot of the wider geographic focus for the decades prior to 2003, and provide a useful reference point as they have previously been described as a temporally stable clonal complex [46]. This will allow us to investigate genetic links and population stability between the 2003 Ugandan outbreaks and the historical T. b. rhodesiense population. Samples were identified as being T. b. rhodesiense if they were isolated from an HAT patient or if they were able to resist the lytic effects of human serum [10], [46].
Twenty eight patients were sampled from the Central Malawi HAT focus and were recruited after admission to Nkhotakota General Hospital between 2002 and 2003. Suspect cases were initially identified by clinical surveillance teams in communities within and on the periphery of the Nkhotakota Wildlife Reserve.
Patient recruitment protocol has been previously described in [52]. Briefly, diagnosis of infection was by microscopic detection of trypanosomes in wet blood films, Giemsa stained thick blood films or in the buffy coat fraction after microhaematocrit centrifugation. Blood was collected by venipuncture from consenting patients, and collected as either 1 ml samples or as 200 µl spots on FTA filter (Whatman) cards. Samples from the Ug/Ke 61–97 focus were grown in mice and have previously been described [46]. A full list of all samples and their geographic and temporal origin is available in Table S1.
For samples isolated on FTA cards, discs of 2 mm diameter were cut from each blood spot using a Harris Micro-punch (Whatman). The discs were washed three times with 200 µl FTA purification reagent (Whatman), and twice with 200 µl 1 mM TE buffer pH 8.0, with incubation for 5 minutes at each wash. The washed discs were then used as substrate for multiple displacement amplification (MDA) whole genome amplification reactions. Whole genome amplification was carried out using the GenomiPhi DNA Amplification kit (Amersham) as described previously [49]. Three independent reactions were carried out for each sample and the reaction products pooled. Where whole blood samples were available DNA was prepared from 1 ml of blood using the Qiagen DNA blood mini kit, following the manufacturer's protocol. MDA products and DNA samples were routinely stored at −20°C prior to use.
One µl of each MDA product or purified DNA was used as PCR template in a volume of 10 µl. The seven microsatellite loci (Ch1/18, Ch2/PLC, Ch3/5L5, Ch3/IJ15/1, Ch4/M12C12, Ch5/JS2 and Ch9/4) have been described previously [16]. Markers Ch3/5L5 and Ch3/IJ15/1, although both on chromosome 3, are 1.2 Mb apart and effectively unlinked [16]. Oligonucleotide primers (both primary and nested) for each marker are detailed in Table S2. PCR conditions were: PCR buffer (45 mMTris-HCl pH 8.8, 11 mM (NH4)2SO4, 4.5 mM MgCl2, 6.7 mM 2-mercaptoethanol, 4.4 µM EDTA, 113 µg.ml−1 BSA, 1 mM of each four deoxyribonucleotide triphosphates), 1 µM of each oligonucleotide primer, and 1 unit of Taq polymerase (Abgene) per 10 µl reaction. For nested reactions, 1 µl of a 1/100 dilution of first round product was used as template in the second round PCR. Microsatellite PCR products were resolved by electrophoresis on a 3% Nusieve GTG agarose gel (Cambrex), and gels were stained with 0.2 µg/ml ethidium bromide and visualised under UV light.
One primer of each pair for the microsatellite nested PCR included a 5′ FAM or HEX modification, allowing size separation of products using a capillary-based sequencer (ABI 3100 Genetic Analyser; Applied Biosystems). A set of ROX-labelled size standards (GS400 markers; Applied Biosystems; Dundee Sequencing Service http://www.dnaseq.co.uk/) was included in the run, allowing accurate determination of DNA fragment size. Data were analysed using Peak Scanner v1.0 software (Applied Biosystems). A multilocus genotype (MLG) for each isolate was defined by the specific combination of alleles across the seven loci (Table S1). Genotypes were defined as heterozygous at a marker if two peaks were detected, whereas homozygotes were represented by a single peak. Mixed infections were defined by the presence of more than two alleles for any one marker.
Analysis of MLGs used Clustering Calculator (http://www2.biology.ualberta.ca/jbrzusto/cluster.php) generating a Phylip Drawtree string (unweighted arithmetic average clustering method, and Jaccard's similarity coefficient), which was converted into a dendrogram by Treeview (http://taxonomy.zoology.gla.ac.uk/rod/treeview.html) [53], with the dendrogram colour coded according to sample origin. Clustering Calculator generated the bootstrap values for dendrograms, using 100 iterations. Marker polymorphism and heterozygosity, Nei's genetic distance (D) and Wright's fixation index (FST) between sample populations, were calculated using GenAlex [54]. Principal Component Analysis (PCA) of the MLGs was performed in GenAlEx following determination of genetic distance with data standardisation. Linkage disequilibrium between paired loci was examined using GDA. eBURST software (http://eburst.mlst.net/default.asp) was used to analyse the clonal expansion of the Ugandan genotypes and identify putative ‘founder’ genotypes [55]. The most stringent setting was used for analysis, in which isolates assigned to the same group are single locus variants (SLV; 6/7 identical loci). In order to use this software, genotypes were treated as described by Stevens and Tibayrenc [28], whereby different combinations of alleles at each locus (for example homozygotes and heterozygotes that share a common allele) are treated as distinct alleles.
One hundred and ninety-eight infected blood samples were examined from three distinct, active HAT foci in 2003, two in South-Eastern Uganda (‘Tororo’ and ‘Soroti’), and one in Malawi. In addition, 52 samples from the Tororo focus were collected between 1961 and 1997 (referred to as‘Ug/Ke 61–97’) and include 28 samples collected in the period 1988–90. During the period between 1990 and 2003, the Tororo focus is considered to have seeded the outbreak in Soroti, which has been linked to the restocking of cattle herds in the region [51]. The shared lineage of the Ugandan samples thus comprises a unique case study, allowing us to examine both the progress of a continuous endemic focus (Tororo) and the establishment of a new, but linked, focus (Soroti). The final focus in Malawi is endemic and still active. However, in contrast to the relatively severe and acute disease observed in Uganda, Malawian HAT is characterised as a chronic disease with slower progression to the late (meningoencephalitic) stage [50], [52], [56]. Thus the Malawi T. b. rhodesiense focus can be distinguished from those in Uganda by both pathogenesis and geography, but they have not been compared genetically.
Comparative analysis of these four populations, therefore, allows us to rigorously examine the role of both space and time in shaping the population dynamics of T. b. rhodesiense. This has been achieved through the use of seven previously described single-locus microsatellite markers, which have been physically and genetically mapped to six different megabase chromosomes of T. brucei. Of the 198 samples, full multilocus genotypes (MLGs) were obtained for 176, with the remainder genotyped for at least four of the seven loci (Table S2). Three samples, one from Tororo (LIRI017) and two from Ug/Ke 61–97 (K237 and UgE90) were identified as mixed genotypes by the presence of three microsatellite alleles for at least one of the seven loci and have therefore been excluded from further analysis (data not shown). A summary of the basic population genetic features of each of the four populations, based on the MLGs, is presented in Table 1. The population from Malawi clearly differs from the Ugandan populations in that the number of distinct MLGs approaches the number of samples whereas the proportion of distinct MLGs is much lower in the three sets of Ugandan samples. This difference is further emphasised by the observed and expected heterozygosities and the values of the fixation index. Thus the Malawi population shows much higher levels of diversity than those from Uganda.
In order to determine if the T. b. rhodesiense population in East Africa was sub-structured due to geographical separation, we compared only those populations that were collected at the same time (2003), to avoid possible temporal sub-structuring. There were more private alleles in the Malawi population (eight) compared to three in Tororo and three in Soroti. Of the private alleles in Malawi five were present at frequencies above 0.1 within the population, whereas only one was above this frequency in Tororo and none in Soroti (Table S3). Nei's unbiased genetic distance (D) and pairwise population FST were measured, indicating that the Ugandan populations are closely related, although the Soroti and Tororo populations are more closely related to each other than either is to the population from Tororo sampled from 1961–97 (Table 2). The Malawi population shows substantial genetic differentiation from the Ugandan samples by both measures (Table 2). The dendrogram of similarity (Fig. 1) confirms the significant separation of the Malawi population from those in Uganda (100% bootstrap support). The Ugandan populations could not be resolved with high confidence, although there is some support for the separation of the Ug/Ke 61–97 population from the Soroti and Tororo 2003 populations (Fig. 1). Principal Component Analysis (PCA) of the MLGs from these populations identified two co-ordinates that accounted for more than 80% of the variation (Fig. 2A). These highlight the separation of the Malawi population and the similarity within the Ugandan populations. Principal coordinate 1, accounting for 70% of observed variation, primarily separates the populations based on country of origin, while coordinate 2 (12% of the variation) partially separates the two Ugandan populations as well as highlighting the high level of diversity within the Malawi focus. The PCA plot indicates that while the genotypes from the Tororo and Soroti foci in 2003 were closely related, the populations are genetically distinct albeit with some overlap. All of these data combine to demonstrate that there is significant genetic differentiation between the Malawian and Ugandan T. b. rhodesiense isolates, indicating population sub-structuring due to geography.
The Ug/Ke 61–97 isolates are representative of the historical population of T. b. rhodesiense present in South East Uganda over a period of 36 years with a significant sample set from 1988/90 [57]. The availability of these historical isolates allowed us to analyse if the trypanosomes had remained genetically stable over time or if new genotypes have appeared, for example by migration or mutation. Twenty-six samples collected from Tororo in 2003 were fully genotyped including one mixed infection identified (LIRI017), which was removed from the analysis, together with a further three partially genotyped samples. Samples from 52 individuals from the Ug/Ke 61–97 sample set were genotyped. Two contained multiple infections and were removed from this study, while 43 of the remaining 50 were fully genotyped for seven microsatellite markers. As these samples were from several geographic locations within the focus (Busia, Busoga and Nyanza – encompassing an area of ∼100 km from Tororo) and collected over several decades we did not attempt to examine this population for indices of mating, to avoid errors due to temporal or geographical sub-structuring. Analysis showed that the dominant MLGs identified in Ug/Ke 61–97 and Tororo2003 were distinct. In Ug/Ke 61–97 the dominant MLGs were MLG 65, 69 and 75, whereas in 2003 the dominant MLGs were MLG 24 and 27 (Tables 3 & S1). One of the most striking differences occurs at locus Ch4/M12C12, which was completely monomorphic in the Ug/Ke 61–97 population. By 2003, three additional alleles had arisen within the population to the point that the predominant allele from Ug/Ke 61–97 was present at a frequency of 0.53, largely as part of a heterozygote pair that dominates the Tororo2003 population (Tables S1 & S3). Additionally, examination of the genetic distance between the populations using Nei's unbiased genetic distance (D) and FST, indicates that while the two populations are highly related they can be distinguished using these measures (Table 2).
In terms of the population structure, an excess of heterozygotes at six out of seven loci was observed in the Tororo2003 samples, while five of the seven loci displayed significant deviation from Hardy-Weinberg predictions, indicating a departure from panmixia (Table 4). However the two markers that displayed agreement with Hardy-Weinberg predictions, Ch2/PLC and Ch5/JS2, had low polymorphism (Table S3) and so could be susceptible to Type 2 error. When duplicate genotypes were removed, two additional markers, Ch1/18 and Ch3/5L5 show agreement with Hardy-Weinberg predictions (Table 4). However, after removal of the repeated MLGs, only 17 individuals remain in the population. Examining the genotypes from this population at each locus, it is clear that six of the loci are predominantly heterozygous for two alleles while the remaining locus is largely homozygous (Table S1) and this genotypic structure precludes any meaningful analysis of linkage disequilibrium. This, coupled with the occurrence of four MLGs (Table 3) that are found multiple times (accounting for 50% of the population), is suggestive of little or no sexual recombination.
The data from Tororo in 2003 suggest little or no mating due to the presence of multiple dominant repeated genotypes and significant disagreement from Hardy-Weinberg expectations at the majority of loci. The data also suggests that the genotypes present in the Ug/Ke 61–97 and Tororo 2003 populations are different. Analysis by PCA of the Ugandan populations (Fig. 3B) provides further evidence for this conclusion with the Ug/Ke 61–97 and Tororo 2003 populations clustering separately.
The Soroti focus, unlike those of Tororo and Malawi, is relatively new as human cases of trypanosomiasis in this district were first reported in 1998. The focus has since been identified as an offshoot of the Tororo epidemic [51]. Subsequent implementation of disease control measures including tsetse trapping and treatment of livestock have been unable to contain the outbreak, with over 400 cases reported between 1998 and 2004 [58]. Fitting with the suggested origins of this disease focus, the population sampled is most closely related (by measurement of Nei's genetic distance and FST) to that of Tororo 2003 (Table 2). While the Soroti population represents the largest sample size, with 84 individuals fully genotyped, the majority of these represent replicate MLGs (Table 3) as only 18 complete and unique MLGs were identified. The most frequent repeated genotype is MLG 49, which is represented 50 times in total. The presence of many parasites with the same genotype constituting more than 59% of the population clearly demonstrates that this population is not panmictic.
Comparison of the genotypes identified in the Soroti population with those from the two Tororo populations using similarity analysis (Fig. 1) shows that they are closely related. While members of each population broadly cluster together but separately from the Malawi population and, with less convincing bootstrap support, the Ug/Ke 61–97 population, there is limited bootstrap support for the Ugandan clusters. However, PCA analysis of the MLGs (Fig. 2B) clearly shows that Soroti and Tororo (2003) populations are closely related to each other but both are more distinct from the Ug/Ke 61–97 population - the two co-ordinates account for 76% of the diversity within this dataset. Furthermore, the relative tightness of the clusters of genotypes from each population reflects the level of diversity within each, with Ug/Ke 61–97 showing a broader scatter reflecting its higher level of diversity. The most frequent MLG in the Soroti population (MLG 49) is not observed in the Tororo 2003 population but the two populations share MLGs 29 and 31 with the latter occurring once in Tororo but seven times in Soroti (Table 3), suggesting the possibility that it might have been a founder genotype in Soroti. To explore the genetic relationships between the genotypes from Soroti and the two Tororo populations and so provide insight into the origins of the Soroti outbreak, the genotypes were analysed using eBURST (Fig. 3). The analysis defines two distinct groups of genotypes one (Group 1) comprising mostly the Soroti and Tororo 2003 isolates (albeit with two MLGs from the Ug/Ke 61–97 population; MLGs 40 and 54), and the second (Group 2) comprising the bulk of the Ug/Ke 61–97 isolates. These results indicate a direct genetic lineage of the Soroti isolates deriving from the Tororo 2003 isolates, consistent with the proposed import into Soroti from Tororo [51]. The predominance of a single clone in Soroti suggests that the import has occurred relatively recently, and probably involved very few MLGs from Tororo, as evidenced by the clonal nature of the Soroti complex, which shows very little genetic divergence in comparison with the more longstanding outbreak in Tororo, where genetic changes have accumulated over time. Group 2 is composed of fewer closely related single-locus variants, resulting from the greater diversity in the Ug/Ke 61–97 population seen by other measures. This may be a reflection of the fact that the number of cases was relatively high in the 1980s into 1990, but then dramatically decreased through the 1990s [59], and this significant reduction in cases and therefore T. b. rhodesiense population offers a potential explanation for the bottleneck effect of the subsequent emergence of a very few surviving genotypes that founded the outbreaks seen in 2003 and onwards.
This, in addition to the close relationship to the 2003 Tororo focus, is consistent with the data that Soroti represents an off-shoot population and suggests the population has been through a recent bottleneck, based on the establishment of a population by a limited number of founder individuals.
The Malawi population, genetically distinct from those in Uganda (Fig. 1 and Table 2), comprises 28 individuals, with 23 fully genotyped with all seven markers. Twenty-one of the 23 MLGs observed are unique within the population. Examination of the markers for agreement with Hardy-Weinberg expectations revealed three loci, Ch4/M12C12, Ch5/JS2 and Ch9/4 that deviate significantly from predictions (Table 4). Disagreement at Ch4/M12C12 and Ch5/JS2 results from heterozygote and homozygote excesses, respectively. For marker Ch9/4 the disagreement arises from the presence of a single individual homozygous for a rare allele. Among the markers both Ch2/PLC and Ch1/18 are dominated by single alleles within the population (Table S3), possibly accounting for the complete agreement at these loci (Type 2 error). While only two repeated genotypes were observed their removal from the population results in Ch4/M12C12 moving to agreement with Hardy-Weinberg. Analysis of the combinations of alleles at pairs of loci showed that only 2 out of 21 loci combinations showed significant evidence of linkage disequilibrium (Tables 3 & S4), which is reduced to a single locus combination (Ch9/4 – Ch3/IJ15/1) once repeated genotypes were removed. The high proportion of unique genotypes observed within this population, coupled with agreement with Hardy-Weinberg and lack of linkage disequilibrium is consistent with the occurrence of a level of recombination within the population. Additionally, the F-statistics for this population suggest that there is an appreciable degree of mating occurring, as the value is close to zero (Table 1), in contrast to the Ugandan populations, where there is significant deviation from zero. Although the number of samples is relatively low (23) and we therefore cannot robustly conclude that the population is panmictic, additional evidence is provided by the fact that the genetic diversity observed within the Malawi cohort is much greater than that in the Ugandan samples (Fig. 1 and Fig. 2A). In summary, the Malawi focus is genetically diverse, displays allelic segregation in the population and there is limited LD consistent with frequent mating. This is the first time that this has been observed for T. b. rhodesiense in the field.
Our results provide evidence that the causative agent of East African Sleeping Sickness, T. b. rhodesiense, can undergo genetic exchange in the field in Malawi, in contrast to previous studies that have described T. b. rhodesiense as a genetically homogeneous variant of T. b. brucei. Unlike the situation in Malawi, the Ugandan populations analysed provided no evidence for the occurrence of frequent genetic exchange and conform with the accepted concept of T. b. rhodesiense as a related set of stable clones in the two foci of disease in Uganda. Thus, the population structure and the role of genetic exchange within this sub-species differs in different geographical regions making it difficult to draw general conclusions about the sub-species as a whole, and so questions the description of T. b. rhodesiense as a genetically homogeneous human infective variant of T. b. brucei.
One question that these findings raise is why mating occurs in the Malawi focus but not in the Ugandan foci. The available laboratory data show that mating can occur between T. b. rhodesiense and T. b. brucei, albeit using a Zambian human infective isolate, and so show that T. b. rhodesiense has the ability to undergo genetic exchange [47], [48]. In Uganda, it is known that both T. b. brucei and T. b. rhodesiense are prevalent in non-human mammalian hosts, notably livestock [10], [34], [46], and are therefore likely to be cycled through the tsetse fly together, providing the opportunity for genetic exchange, particularly as T. b. brucei undergoes genetic exchange itself. In this scenario, one would predict that T. b. rhodesiense would undergo genetic exchange, show high levels of diversity and not be distinguishable from T. b. brucei except by the presence of the SRA gene. The available evidence does not support this as firstly we have shown (in Soroti and Tororo) that the populations are of low diversity with frequent identical genotypes and secondly previous studies have shown that T. b. brucei can be distinguished from T. b. rhodesiense by RFLP and minisatellite markers [10], [60], demonstrating that they are genetically isolated. Based on these considerations, one hypothesis to explain the results is that Ugandan T. b. rhodesiense has lost the ability to undergo genetic exchange. This could be tested by attempting laboratory crosses with these strains. In contrast, our data support the occurrence of genetic exchange in Malawian T. b. rhodesiense and so one would predict that genetic exchange would also occur with local T. b. brucei with human infection occurring when the SRA gene is inherited. Unfortunately no viable Malawian T. b. brucei strains are available and so it is not currently possible to test this hypothesis.
The genotyping of isolates from the two foci in Uganda not only provides important information about the role of genetic exchange in these populations but also information about the temporal genotypic stability in Tororo and the potential origin of the Soroti outbreak. Our data show that genetic exchange is limited or does not occur in these populations based on the lack of agreement with Hardy-Weinberg predictions, high levels of heterozygosity, linkage disequilibrium and the high frequency of identical genotypes. These findings lead to the conclusion that these populations are clonal, primarily evolving by mitotic division and mutation. This conclusion agrees with previous analysis of the Ug/Ke 61–97 population using minisatellite markers [10] where two predominant genotypes represented much of the population and, furthermore, these were stable over time based on the analysis of a few isolates from 1961 [10]. Our data presented here provide a higher resolution analysis by using a larger number of markers and provide a further test of the stability of clonal trypanosome populations in space and time. The genotypic comparison between Ug/Ke 61–97 and Tororo 2003 provides a novel finding that stability over time may not be a feature of these populations. Using similarity analysis (Fig. 1), PCA (Fig. 2B) and eBURST (Fig. 3), the two populations are different – although they do share a small number of common MLGs, the dominant MLGs are different. The two populations show similarity in that they both contain multiple repeated genotypes as well as a number of common alleles (Tables S2 & S3). However, the eBURST analysis separates the two populations into distinct, but related, clusters. As the two populations were sampled 13 years apart and there is no evidence for genetic exchange, we must assume that either mutation accounts for these differences and has occurred at several loci over this time span, or alternatively there has been a degree of migration and introduction of some novel genotypes. This is in marked contrast to the similarities between the Soroti and Tororo 2003 populations, which are highly related by PCA and similarity analysis (Fig. 1 and 2) as well as sharing two MLGs (MLG 29 and 31). The predominant MLG in Soroti (MLG 49) is, however, not observed in Tororo but the eBURST analysis (Fig. 3) shows that this MLG differs by a single allele from a series of the other MLGs in the population by a classical star like relationship characteristic of a clonal population. MLG 49 differs by a single allele from MLG 31 (present in both populations), which occurs seven times in the Soroti population and is related to MLG 29 (also present in both populations) by a further single allelic difference. Based on these data, a hypothesis for the origin of the Soroti focus is that it was seeded by MLGs 31 and 29 from Tororo, which mutated to generate MLG 49 and subsequently the other related genotypes. As Tororo was not sampled at the time point when the cattle were moved into Soroti and initiated the outbreak, this hypothesis cannot be tested directly. However the genotype data add strong support to the conclusions reached by Fevre et al. 2001 [51] as to the origin of the Soroti outbreak. Even though the two populations are very similar and do not undergo significant levels of recombination, it is again clear that the genotypes are not wholly stable in time and place but form a clonal complex often dominated by a single or a few highly related genotypes.
These findings have implications for our understanding of recombination as an evolutionary driving force in trypanosomes. It is clear that mating plays different roles in different species, with T. vivax and T. b. gambiense being clonal [36], [37], [61], whereas T. b. brucei and T. congolense can undergo frequent mating [10], [62]. However, T. b. rhodesiense provides evidence for these differences being displayed within a sub-species. The identity of the trigger for whether mating occurs or not within these species or subspecies is obviously a key question to address, but it seems reasonable to assume that it is likely to depend upon certain epidemiological scenarios (e.g. transmission intensity, reservoir host population, tsetse species etc). This plasticity in the use of sexual recombination within a genus, and particularly within a species (T. b. rhodesiense versus T. b. brucei presenting a prime example), makes trypanosomes a unique paradigm for studying the evolution of sexual recombination, and the role that mating plays in shaping the responses to epidemiological selective pressures.
|
10.1371/journal.pcbi.1003020 | A Two-State Model for the Dynamics of the Pyrophosphate Ion Release in Bacterial RNA Polymerase | The dynamics of the PPi release during the transcription elongation of bacterial RNA polymerase and its effects on the Trigger Loop (TL) opening motion are still elusive. Here, we built a Markov State Model (MSM) from extensive all-atom molecular dynamics (MD) simulations to investigate the mechanism of the PPi release. Our MSM has identified a simple two-state mechanism for the PPi release instead of a more complex four-state mechanism observed in RNA polymerase II (Pol II). We observed that the PPi release in bacterial RNA polymerase occurs at sub-microsecond timescale, which is ∼3-fold faster than that in Pol II. After escaping from the active site, the (Mg-PPi)2− group passes through a single elongated metastable region where several positively charged residues on the secondary channel provide favorable interactions. Surprisingly, we found that the PPi release is not coupled with the TL unfolding but correlates tightly with the side-chain rotation of the TL residue R1239. Our work sheds light on the dynamics underlying the transcription elongation of the bacterial RNA polymerase.
| Pyrophosphate ion (PPi) release is a critical step in the nucleotide addition cycle of transcription elongation. Despite extensive experimental studies, the kinetic mechanism of the PPi release in bacterial RNA polymerases (RNAP) still remains largely a mystery. As a cellular machine, RNAP contains more than 3000 residues, and thus it is challenging for all-atom molecular dynamics (MD) simulations to directly capture the process of the PPi release. In this study, we have simulated the dynamics of the PPi release at microsecond timescale using the Markov State Models (MSMs) built from extensive MD simulations in explicit solvent. MSM is a powerful kinetic network model and can predict long timescale dynamics from many short MD simulations. Our results suggest a simple two-state model for the PPi release in RNAP, which sharply contrasts with the more complex four-state hopping model in the yeast RNA polymerase (Pol II). We also observe a 3-fold faster dynamics for the PPi release in RNAP compared to Pol II, consistent with the faster transcription rate in the bacterial systems. Our results greatly improve our understanding of the PPi release, and also provide predictions to guide future experimental tests.
| The DNA-dependent RNA polymerase is the main enzyme that participates in the transcription process transferring the genetic information from DNA to messenger RNA (mRNA) [1]. Crystallographic structures of the multi-subunit RNA polymerases in eukaryotes [2]–[4] and bacteria [5]–[8] engaged in transcription elongation process have been obtained. These atomic-level structures provide static snapshots of the transcription cycle [9]–[14].
In each nucleotide addition cycle (NAC) of the multi-subunit RNA polymerase, the post-translocation state first allows the substrate NTP to bind to the active site [6]. Then, a critical domain, named trigger loop (TL), can fold then expel the solvent from the active site [15]–[17], and finally form direct contacts with the substrate NTP. Substitution of a conserved TL histidine can significantly decrease the polymerization rate [18]–[21]. Recent mutagenesis studies have shed light on the roles of the TL on the nucleotidyl transfer [20], [21], and the reverse intrinsic hydrolysis process [22]. Previous MD simulation studies also provided information on TL dynamics and its potential regulatory roles during the translocation process [23], [24]. After the catalytic reaction, PPi forms and releases from the active site [25], [26]; then the TL opens and allows the template DNA to translocate so that a new NAC can start. Extensive biochemical and theoretical studies have been performed to understand the specific steps in the NAC, such as motions of the TL [17], catalysis [26]–[30], translocation [23], [24], [31]–[33] and NTP binding [34], [35].
PPi release in single subunit T7 RNA polymerase is proposed to be tightly coupled with the translocation [36] but the same coupling is not observed in Escherichia coli (E. coli) RNA polymerase [37]. Interestingly, recent fluorescence and biochemical studies found that the PPi release in the E. coli RNA polymerase occurs shortly before or concurrently with the translocation [33]. Nonetheless, the interplay between the PPi release step and the TL opening motions at molecular level is still elusive. Previously, we used MD simulations to study the PPi release in the eukaryotic RNA polymerase II (Pol II) [25]. We proposed a hopping model for Pol II in which PPi release was coupled with the TL tip motion through the interactions between the TL residue H1085 and the (Mg-PPi)2− group, and subsequently hopping among several positive charged residues in the secondary channel. Our model further suggested that the PPi release is a fast dynamic process so that it may not be able to induce the fully TL opening motion.
A comparison of the secondary channel and TL structure between Pol II and bacterial RNA polymerase (RNAP) from T. thermophilus (Tth) displays substantial differences (See Figure 1) [3], [7]. In Pol II, the TL contains a long loop domain (from the Rpb1 residue T1080 to T1095) [3]. However, the TL in RNAP consists of two alpha helices connected by a short turn in the closed state [6]. This structural difference suggests that the dynamics of the TL folding in these two systems are likely to be different. Moreover, in addition to the conserved Tth TL residue H1242, the Tth TL residue R1239 also interacts with the substrate NTP [6]; this residue is absent in Pol II and mutation of the counterpart residue in E. coli (R933A) can reduce the nucleotide addition rate [20]. Moreover, the secondary channel in Tth RNAP is much shorter than that in Pol II (See Figure 1), and exhibits a different layout of the positively charged residues. Specifically, in Pol II, the four residues, K619, K620, K518 and K880 are located at relatively separated sites (See Figure 1A). However, the positively charged residues in Tth RNAP: K908, K912, K780 and K1369 are close to each other in a continuous region (See Figure 1B). Given these structural differences, it is of interest to compare the dynamics of PPi release in RNAP with that in Pol II.
Although conventional all-atom MD simulations can provide the dynamic information for biological macromolecules at atomic resolution, it is still challenging to capture the biologically relevant timescales in microseconds or even longer. Markov State Models (MSMs) constructed from a large number of short simulations provide one way to overcome this timescale gap [38], [39]. MSMs have been successfully applied to model the long timescale dynamics that cannot be directly accessed by conventional MD simulations in studying the conformational changes of biological macromolecules [39]–[41], including our previous study of PPi release in Pol II [25].
In this study, in order to reveal the mechanism of the PPi release in RNAP, we constructed a MSM from extensive all-atom molecular dynamics (MD) simulations in explicit solvent with a system size of nearly 300,000 atoms and aggregated simulation time of ∼1 µs. Our results reveal that the PPi release in Tth RNAP adopts a simple two-state model with a fast dynamics over a few hundred nanoseconds. Surprisingly, we found that the PPi release is not coupled with the secondary structure unfolding of TL but only with the side-chain rotation of the TL residue R1239.
To study the release mechanism of the (Mg-PPi)2− group in RNAP, we modeled the PPi-bound RNAP complex by directly cleaving the Pα-O bond in the ATP-bound RNAP complex that is derived from the Tth RNAP crystal structure (See SI Figure S1 for the two structures and the Methods section for the modeling details) [6]. This modeled PPi-RNAP complex was used as the starting structure for the steered MD (SMD) simulations to obtain the initial release pathways. To eliminate the bias in SMD simulations, we have then performed 100 10-ns MD simulations, and these simulations have widely sampled the region in the secondary channel (See SI Figure S2). Finally, we have constructed a MSM from these simulations to obtain the dynamics and other thermodynamic properties of the PPi release (See the Methods section for details).
Our MSM shows that the PPi release in Tth RNAP adopts a simple two-state model. In addition to the initial state with the PPi in the active site (S1 state in Figure 2A), only one additional metastable state is identified (S2 state in Figure 2A), and this state is ∼7-fold more populated than the S1 state (See Figure 2B). The S2 state locates in an elongated region where several positively charged residues can stabilize the (Mg-PPi)2− group. These results contrast with our previous findings that the (Mg-PPi)2− group in Pol II hops through four clearly separated metastable states [25].
When the (Mg-PPi)2− group is in the active site (See Figure 2C), three positively charged β′ residues R1029, H1242 and R1239 can interact with the negatively charged (Mg-PPi)2− group. The residue R1029 locates at the exit of the active site, and thus it may play similar roles on the PPi release with its corresponding residue K752 in Pol II (See Figure 2D) [25]. Interestingly, the location of the conserved TL residue H1242 is different from its counterpart residue H1085 in Pol II, though both of them are in direct contact with the (Mg-PPi)2− group. Both before and after chemistry, H1242 interacts with the Pα-O atom of the NTP in RNAP, whereas H1085 is in contact with Pβ-O atom in Pol II (See Figure 3) [3], [6]. To achieve this, the H1242 in RNAP has to locate deeper in the active site compared to H1085. Finally, R1239 in RNAP locates at the same position as H1085 in Pol II, suggesting that these two residues may play similar roles in the PPi release.
After escaping from the active site, the (Mg-PPi)2− group reaches the S2 state with an elongated shape. In this state, multiple positively charged residues on the secondary channel (K780, K908, K912 and K1362) can provide favorable electrostatic interactions with the negatively charged (Mg-PPi)2− group (See Figure 2C). In contrast, the (Mg-PPi)2− group in Pol II is found to transfer through several hopping sites where groups of positively charged residues are spatially well separated (See Figure 1A) [25]. From the S2 state, the (Mg-PPi)2− group will directly enter the solvent. In order to elucidate the specific roles of the three important residues: R1029, H1242 and R1239 in the PPi release (See Figure 2D), we performed additional mutant simulations starting from several different conformations from the S1 and S2 states.
The potential of mean force (PMF) profile along the distance between the (Mg-PPi)2− group and the Mg2+A is displayed in Figure 4A. The PMF plot shows two major free energy basins that are consistent with the two metastable states identified by our MSM. The starting structures chosen for the mutant simulations fall into two different regions in the PMF profile (P1 and P2 sites in Figure 4A). The P1 site is located in the S1 state, while the P2 site is located in the S2 state but near to the boundary between the S1 and S2 states. Initial conformations from these two sites allow us to examine the roles of the residues involved in different stages of the PPi release.
The mutant simulation results indicate that both residues R1239 and R1029 can facilitate the escape of the (Mg-PPi)2− group from the active site to S1 state (See Figure 4B). Here, we use the distance between the (Mg-PPi)2− group and the Pα atom of the 3′-terminal nucleotide of the RNA chain (dαβ) to describe the extent of the PPi release from the active site. In the WT simulations (See P1 in Figure 4A), the (Mg-PPi)2− group can move towards the exit of active site with the dαβ value increasing from 6 Å to around 8 Å (See the left panel of Figure 4B). However, the R1239A and R1029A mutants lead to a weaker tendency for the (Mg-PPi)2− group to escape the active site (the dαβ value fluctuates around 5.5 Å, middle and right panels in Figure 4B). On the other hand, the R1029K mutant is shown to have a similar effect to help the (Mg-PPi)2− group to leave the active site as in WT(see Figure S4A). These results indicate that positively charged residues play a crucial role to facilitate the PPi to release from the active site.
Notably, the H1242A mutant can dramatically promote the PPi release from P1 site (See SI Figure S4C), suggesting that H1242 may prevent the PPi release from the active site. In contrast, the TL residue H1085 in Pol II was previously found to facilitate the PPi release from the active site [25]. This difference may be due to the different locations of these two residues in the active site. Compared with H1085 in Pol II, H1242 in RNAP locates significantly deeper inside of the active site (See Figure 3D). Therefore, it will be more difficult for H1242 to rotate and help the (Mg-PPi)2− group to leave the active site. Instead, H1242 can provide an attractive interaction to prevent the PPi release.
Next, we evaluated the roles of residues R1239 and R1029 in PPi release when the (Mg-PPi)2− group is at the S2 state (P2 in Figure 4A). In the WT system, the (Mg-PPi)2− group fluctuates around its initial location within our simulations at a few nanoseconds, which was also observed in the R1029K mutant simulations initiated from P2 (Figure 4C). Intriguingly, R1029A and R1239A substitutions lead to dramatic, but opposite, effects. The R1029A substitution facilitates the PPi release toward the solvent (Figure 4C, right panel). Combined with the previous observations, we conclude that the R1029 may facilitate the PPi release from the active site but prevents the PPi release when it arrives at the S2 state. Thus R1029 plays a similar role as the corresponding residue K752 in Pol II (See Figure 2D) [25]. In contrast, the R1239A substitution drives the (Mg-PPi)2− group back to the S1 state, suggesting that R1239 is critical for the (Mg-PPi)2− group to escape the active site (middle panel in Figure 4C). This indicates that R1239, rather than H1242 residue, plays the role in PPi release most equivalent to that played by H1085 in Pol II. Compared with its counterpart residue H1085 in Pol II, the R1029 has a longer and more flexible side chain. In addition, it can form a stronger salt bridge with the (Mg-PPi)2− group. Therefore, the side-chain rotation of R1239 alone may be sufficient to facilitate the PPi release.
Based on extensive unbiased MD simulations, we built a MSM for the PPi release in RNAP to elucidate its long timescale dynamics. The MSM identified a two-state model for the PPi release (See Figure 6A). The mutant simulations indicate that the β′ residues R1239 and R1029 can facilitate the escape of the (Mg-PPi)2− group from the active site after the catalytic reaction (See Figure 4B). Then the (Mg-PPi)2− group transfers to the S2 state, where it forms favorable interactions with four positively charged residues on the secondary channel: K908, K912, K780 and K1369 (See Figure 6A). More strikingly, our work suggests that the PPi release does not induce the TL unfolding but tightly couples to the side-chain rotation of the TL residue R1239, which in turn makes the TL tip region more flexible. Furthermore, our control simulations show that TL is stable in Pol II, but can quickly unfold (within 200 ns) when exposed to the solvent. We thus speculate that the rotation of R1239 that accompanies the PPi release may allow solvent to re-enter the active site and promote the overall movements of the TL domain; This TL movements would further lead to its exposure to solvent and eventually allow TL unfolding. However, the timescales for this solvent-induced TL unfolding may be significantly longer than that of the PPi release so that we didn't observe it in our simulations.
We found that the TL in RNAP may be more difficult to unfold than that of Pol II, since its secondary structures barely unfold upon the PPi release. Therefore, if the open state of the TL is a pre-requisite step for the translocation as recently suggested by both experimental [33] and computational studies [24], it is intriguing that the transcription rate for bacterial RNAP is much faster than that of Pol II [27]. Despite that the more stable secondary structure of the TL in RNAP may slow down its opening motion, its reverse closing motion may be spontaneous and fast. This fast closing motion may further accelerate the nucleotide addition process to achieve an overall fast transcription rate.
Finally, our MSM indicates that the PPi release in bacterial RNAP is faster than that in Pol II [25]. This faster dynamics is due to several factors: First, the secondary channel of RNAP is shorter than the that of Pol II due to the absence of the funnel region, therefore the (Mg-PPi)2− release path is shorter, which leads to a faster PPi release from the active site of RNAP. Next, in RNAP, the PPi only needs to overcome a single free energy barrier before it can be released to the solvent (See Figure 6A). In contrast, the PPi release in Pol II was found to go over multiple free energy barriers sequentially before it could be released (See Figure 6B). Furthermore, in our model for RNAP, the S2 state (in the pore) has a population over seven times larger than that of the S1 state (in the active site). Thermodynamically, this difference will favor release of PPi from the active site. However, in Pol II, the equilibrium population of the S2 state (the first state in the pore) is comparable to that of the S1 state (in the active site) [25]. This difference may be due to the fact that the S2 state in bacterial RNAP is greatly stabilized by four positively charged residues that are spatially close to each other (K908, K912, K780 and K1369), but in Pol II, these positively charged residues locate at relatively separate sites (See Figure 1A). Finally, R1239 in RNAP can substantially facilitate the (Mg-PPi)2− release from the active site all the way to the solvent due to its longer and more flexible side chain (See Figure 5). However, its counterpart residue for PPi release from Pol II, H1085, only promotes the PPi escape from the active site to the first metastable state, S1, rather than all the way to solvent [25].
We constructed the MSM to study the PPi release in RNAP, and our algorithm consists of the following steps: (1) Model the PPi bound complex. (2) Generate initial release pathways using SMD simulations. (3) Seed unbiased MD simulations from these initial pathways, and (4) Construct the MSM to identify metastable intermediate states and obtain both thermodynamics and kinetics of the PPi release.
In order to obtain the initial PPi release pathway, we applied steered MD simulations [51] to pull the (Mg-PPi)2− group out of the active site. The pulling was performed along three directions with the aim of considering all the possible PPi release pathways. Three groups of residues were used to determine the pulling directions: β′ subunit residues 1136–1145, 908–914 and 1246–1253 (named as group I, II and III respectively). Two sets of pulling simulations were along the wall of the secondary channel: one was pulled towards the center of the Cα atoms of group I residues, and the other was directed to the center of Cαatoms of both group I and group II residues. The third set of pulling simulations pointed to the center of the Cαatoms of group II and group III residues, and toward the solvent. The external force was only applied on the center of mass of the PPi group with a force constant of 0.5 kJ mol−1 Å−2 and pulling rate of 0.01 Å/ps. For each pulling direction, five independent steered MD simulations were performed starting from the final snapshot derived from 5 parallel MD simulations of the PPi-bound RNAP complex.
We first divided the conformations from SMD simulations into 20 clusters using the K-center clustering algorithm [52]. In the clustering, the distance between a pair of conformations was set to be the RMSD value of three PPi atoms (the bridge oxygen and two phosphate atoms). To compute RMSD, the structure was aligned to the energy minimized PPi-RNAP complex by the Cαatoms of the bridge helix domain. We then randomly selected 5 conformations from each cluster (a total of 100 conformations) to conduct unbiased MD simulations. Each simulation was run for 10 ns and the snapshots were saved every 2 ps. Altogether, we obtained an aggregation of ∼1 µs simulations with 500,000 conformations.
In MSM, the conformational space is divided into a number of metastable macrostates and the fast motions are integrated out by coarse graining in time with a discrete unit of Δt. The model is markovian if Δt is longer than the intra-state relaxation time. In other words, the probability for the system to be at a given state at time t+Δt only depends on the state at time t. In MSM, the long timescale dynamics can be modeled by the first-order master equation.(1)Where P(nΔt)is the state populations vector at time nΔt, and T is the transition probability matrix. Δtis the lag time of the model. To calculate T, one can normalize the transition count matrix generated by counting the number of transitions between each pair of states at the observation interval of Δt from MD trajectories. MSM has been successfully applied to model conformational changes that occur at timescales that cannot be directly accessed by conventional MD simulations such as protein folding [39], [40], [53]–[55].
To construct the MSM, we have followed a splitting and lumping procedure [52]:
In order to check if the model is markovian, we have plotted the implied timescales (τk) as a function of the lag time τ:(2)where μk is the eigenvalue of the transition probability matrix T with the lag time τ. The implied timescales correspond to the average transition times between two groups of states, and thus indicate the dynamics of the system. If τ is sufficiently large, the model is markovian, and the predicted implied timescales will not change upon the further increase of the lag time. In our system, the implied timescale plots reach the plateau at the lag time of ∼4 ns (See SI Figure S3A). Therefore, we select the lag time of 4.5 ns to construct the final MSM.
To further validate the model, we predicted the probability for a given macrostate to stay within it after a certain lag-time based on our MSM, and this predicted values are in good agreement with those obtained from the original MD simulations (See SI Figure S3B).
In order to investigate the stability of the TL in free solution, we have performed a 300 ns control simulation with the isolated TL domain (A1225 to A1265) in solution (with ∼6900 atoms, See SI Figure S6B). However, it is difficult to extend individual MD simulations of the complete transcription complex (nearly 300 K atoms) to hundreds of nanoseconds due to its high computing cost. Therefore, we have also performed simulations with a truncated RNAP complex containing all the motifs surrounding the TL domain (See SI Figure S7A), including β subunit residues 381–569, 831–1049, β′ residues 604–794, 901–1470, 10 upstream hybrid DNA-RNA base pairs, 6 downstream DNA base pairs and Mg2+A in the active site. The final solvated system only contains ∼118 K atoms, but it still takes more than 2 months to perform one 300-ns simulation using 24 CPU cores.
The explicit SPC water model was used for the MD simulations, 1 and 29 Na+ ions were added to neutralize the isolated TL and truncated RNAP model respectively. The other setups for the MD simulations were the same as the seeding MD simulations. We have performed one 300-ns simulation for the isolated TL and the other two 300-ns simulation for the truncated RNAP. For the truncated RNAP model, several terminal residues that are truncated from the complete model were fixed in the simulations in order to avoid undesired unfolding.
|
10.1371/journal.pcbi.1003924 | A Neural Population Model Incorporating Dopaminergic Neurotransmission during Complex Voluntary Behaviors | Assessing brain activity during complex voluntary motor behaviors that require the recruitment of multiple neural sites is a field of active research. Our current knowledge is primarily based on human brain imaging studies that have clear limitations in terms of temporal and spatial resolution. We developed a physiologically informed non-linear multi-compartment stochastic neural model to simulate functional brain activity coupled with neurotransmitter release during complex voluntary behavior, such as speech production. Due to its state-dependent modulation of neural firing, dopaminergic neurotransmission plays a key role in the organization of functional brain circuits controlling speech and language and thus has been incorporated in our neural population model. A rigorous mathematical proof establishing existence and uniqueness of solutions to the proposed model as well as a computationally efficient strategy to numerically approximate these solutions are presented. Simulated brain activity during the resting state and sentence production was analyzed using functional network connectivity, and graph theoretical techniques were employed to highlight differences between the two conditions. We demonstrate that our model successfully reproduces characteristic changes seen in empirical data between the resting state and speech production, and dopaminergic neurotransmission evokes pronounced changes in modeled functional connectivity by acting on the underlying biological stochastic neural model. Specifically, model and data networks in both speech and rest conditions share task-specific network features: both the simulated and empirical functional connectivity networks show an increase in nodal influence and segregation in speech over the resting state. These commonalities confirm that dopamine is a key neuromodulator of the functional connectome of speech control. Based on reproducible characteristic aspects of empirical data, we suggest a number of extensions of the proposed methodology building upon the current model.
| Our knowledge of brain activity and network organization during complex motor behaviors in humans relies mainly on neuroimaging studies. However, the majority of available brain imaging methods are not feasible for quantifying the neural processes that occur on very short time-scales at the microscopic level. To address this shortcoming of functional MRI, we designed a mathematical model, which simulates brain activity using local ensembles of neurons and physiologically meaningful variables, such as cellular membrane potentials and ion channel relaxation times. We further incorporated dopaminergic function into our model as a neuromodulator of the dynamic organization of brain networks. We applied our model to examine brain networks controlling human speech and language production. We present a rigorous mathematical proof, which establishes the theoretical validity and solvability of the presented model, and we discuss the influence of dopaminergic transmission on simulated brain activity. We show that our model successfully reproduces characteristic changes seen in empirical data between the resting state and speech production. Our results indicate that the proposed mathematical model may be used as a platform for future studies to investigate the specific impact of certain pathologies within the dopaminergic pathways and their effect on global network dynamics.
| Computational neuroscience takes a grounds-up approach to understand complex neural phenomena by investigating the underlying activity of neurons themselves. Starting from the basic neural model of Hodgkin and Huxley [1], which described the electric activity of a neuron in terms of the cell's constituent ionic currents, computational neuroscientists began to study temporal input-output relations of neural units. One of the most famous models created during this era was the “point unit” by McCulloch and Pitts [2]. Among the first modelers who incorporated not only temporal but also spatial aspects of neural processing was Rall, who used compartment models to show the strong impact of dendritic arborization on neural processing of synaptic inputs [3]. His work laid the ground for the first neural network modeling based on rather complex single neuron models. Next, the so-called Wilson-Cowan units [4] allowed for simulation of rather realistically macroscopic responses of entire brain regions on the scales corresponding to measurements obtained by non-invasive in vivo human imaging techniques. Indeed, neural simulations did fit well with data from various human imaging modalities, such as magnetoencephalography (MEG) [5], positron emission tomography (PET) [6], and functional magnetic resonance imaging (fMRI) [7], [8]. Many neural states have been modeled, resulting in a rich literature on resting-state brain activity as well as behavior-specific activities, such as visual [6], memory [9], sensorimotor [10] and auditory [11] processing. However, while it is generally accepted that differences may be seen between the resting state and task conditions as well as between healthy and patient data and models, the fundamental question in modeling and data analysis methodology, the significance of differences between modeled conditions, remains unclear [12].
Given that functional activity may be affected by neurotransmitters (see [13] for a review), recent modeling efforts have been undertaken to integrate neuromodulators, such as dopamine, into task simulations. Dopaminergic neurotransmission has been implicated in cognition, learning, motor control, and, more generally, sensorimotor integration [14], [15], [16]. Chadderdon and Sporns proposed a large-scale computational model of prefrontal cortex to show the effects of dopamine release on the onset and performance of working memory tasks, which could be confirmed by behavioral, single-cell and neuroimaging data [17]. Determining how dopamine may regulate the functional connectivity observed during a behavioral task is a critical next step in addressing the ambiguity of task-specific functional connectivity.
To that end, we present a biologically-informed, large-scale model, which is based on neurobiological considerations, to simulate neuronal function and connectivity modulated by dopamine release in the human brain. The present paper applies this model to investigate speech production, one of the most complex human behaviors, which can be studied in a neuroimaging setting. Speech production is known to integrate several neural networks, ranging from auditory processing to motor control of articulatory movements [18], [19]. Our recent study has demonstrated that dopaminergic modulation may play a role in left-hemispheric lateralization of functional brain activity and connectivity during speech production in the absence of lateralized structural networks [20]. Here, we propose an extension of a non-linear model presented by Breakspear et al. [21] to allow for the simulation of brain activity due to dopaminergic modulation. We introduce the original model, which is a system of stochastic differential equations (SDEs), and rewrite it in terms of a multi-dimensional time-continuous stochastic process. Coupling between the brain regions with respect to regional neural firing rates is incorporated within the framework of Ito processes [[22], Chap. 7]. A dopamine release model is developed and integrated into the model linking the basal ganglia and the laryngeal motor cortex based on previous studies [23]. We further present a mathematical proof establishing the existence and uniqueness of solutions to the extended model. Exploiting specific structural properties of the model, a computationally efficient scheme for numerical approximation of solutions is also presented. We show simulations of resting-state and dopamine-modulated BOLD signals and analyze the associated functional connectivity networks as related to corresponding real fMRI data obtained from healthy volunteers during the resting state and speech production. Finally, we discuss merits and limitations of the proposed model.
All participants provided written informed consent before participation in the study, which was approved by the Institutional Review Boards of the Icahn School of Medicine at Mount Sinai and National Institute of Neurological Disorders and Stroke, National Institutes of Health.
Our goal was to simulate a large-scale neural population using coupled small scale local models, each replicating neural activation in a specific brain region (), while incorporating neuromodulator release in a region-specific manner. Every regional subsystem consisted of interconnected excitatory and inhibitory neurons, which were assumed to be representatives of the local neural ensemble within a region. Thus, all quantities were understood as mean values across the considered region. The dynamics of regional state variables were governed by voltage-gated ion channels, functional synaptic couplings and neurotransmitter release. Thus, the temporal evolution of the entire population was determined solely by the interaction of its regional subsystems. In contrast to other approaches, the model discussed here was not based on coupled oscillator systems like the widely-used Kuramoto model (compare, e.g., [24] or [25]), but was based on neurobiological considerations. Below, we first detail the theoretical aspects of the model, including the Wilson-Cowan and dopamine dynamics, then describe the integration of the model with data.
Following Breakspear et al. [21], we denote the average membrane potential of neurons in region by , which we assume to be governed by voltage-gated potassium (K), sodium (Na) and calcium (Ca) ion channels together with the passive conductance of leaky (L) ions. Thus, for let denote the fraction of open j-ion channels and let be the ion population's maximum conductance for (i.e., when all j-ion channels are open). The basic model describing current flows across neural membranes in region is a balance equation of the form (assuming unit neural capacitance) (1)where denote the respective Nernst potentials and are neural activation functions. To adequately reflect relaxation times of potassium channels, is characterized by an exponential decay (2)with being the value of at the initial time , denoting a temperature scaling factor and being the relaxation time. For brevity, we introduce the shorthand notation . The other neural activation functions are defined as and .
Assuming that the ion channel specific opening-thresholds are normally distributed with mean and variance across the considered neural population, the fraction of open channels in region may be computed as (3)
Note that the basic model (1) consists exclusively of uncoupled equations, i.e., the membrane potential of neurons in region is entirely independent of neural firing in neighboring regions. Coupling is introduced by considering firing rates of excitatory and inhibitory neurons across the whole population. Thus let be the mean membrane potential of inhibitory interneurons in region , and define average excitatory and inhibitory firing rates as follows (4)where and denote the mean values and and are the variances of membrane threshold potentials of excitatory and inhibitory neurons, respectively (assuming a normal distribution of thresholds across the neural population). Regional membrane potentials are altered by excitatory and inhibitory cell firing via synaptic feedback loops. Thus, functional synaptic factors , and are introduced to scale excitatory-to-excitatory, inhibitory-to-excitatory and excitatory-to-inhibitory couplings, respectively. Furthermore, to reflect firing rate dependent glutamate neurotransmitter release (opening additional calcium ion channels) a supplemental scaling parameter (the ratio of NMDA to AMPA receptors) is used. Non-specific input to excitatory and inhibitory neurons is modeled using random noise, which gives rise to a system of coupled stochastic differential equations (SDEs). Thus let denote a scalar Wiener process [[26], Sec. 1.6] and let and be -dimensional Ito processes. To avoid notational overhead we understand the auxiliary quantities (3) and (4) to be obviously adapted to (replace by the components of in the respective definitions). In the following we establish a vectorial representation of the basic model equations given in [21]. Thus, for a vector let denote a diagonal matrix with the components of on its main diagonal. Further, we introduce the -dimensional vectors , , , and . With denoting a (global) coupling parameter and , we define a function () with components given by (5)and (6)such that . Similarly with we introduce a vector defined by (7)where and denote synaptic factors corresponding to non-specific excitatory/inhibitory input, is a noise scaling parameter and is a vector of ones. Setting we thus obtain the SDE (8)which is the Ito version of the original multi-compartment neural dynamics model presented by Breakspear et al. [21].
The original model (8) uses a scalar parameter to parameterize excitatory coupling between regions. This means, in the framework considered here, that inter-regional connectivity strengths constant throughout the entire brain. To relax this restrictive assumption, we assign each pair of regions coupling parameters and representing the connectivity strengths and , respectively. We collect the inter-regional coupling parameters in a matrix and incorporate it in the model (8) as follows. Instead of calculating excitatory-to-excitatory neural feedback by relying on a mean firing rate , we scale neural firing using weight information from the coupling matrix . Thus, we assume that firing of brain areas connected to region impact the membrane potential in region according to . Hence, (5) is modified to be (9)
We set to reflect local excitatory input within a region. Note that we do not impose any restrictions on the directionality of regional couplings. Depending on the application considered, the above formulation allows the use of directed connections (i.e., a non-symmetric coupling matrix ) or undirected connections ( symmetric).
A second extension to the original model was incorporated to simulate the effects of speech-induced dopamine release, as shown previously in real data [20]. We were especially interested in the effect of dopaminergic neurotransmission on the primary motor cortex [27] and its direct influence on the activity of the laryngeal motor cortex (LMC), which is a final common cortical pathway of speech control [28], [29]. Keeping in mind the biologically-inspired channel model adopted in the present paper, elevated dopamine levels in the striatum (without a differential effect on either D1 or D2 type of dopamine receptors) were assumed to increase the probability that potassium, sodium and calcium channels of LMC neurons open, thus making these neurons more likely to fire. Hence, we simulated both D1- and D2-type modulatory effects on these channels [30].
We modeled the direct dopaminergic pathway from the substantia nigra, pars compacta (SNc) to the LMC [28]. Thus, we assumed that dopamine release was solely driven by neural activity in the SNc. Hence, let be a two-dimensional Ito process with components and denoting the dopamine concentration in the left () and right (r) LMC respectively. We assume is governed by two simple mass balance equations (10)
To reflect the positive feedback of neural firing in the SNc on dopamine release we define (11)where denotes the neural firing rate in the left/right substantia nigra as defined in (4), and is a (time-defpendent) production rate. We assume that attains a maximum value during speech production and is equal to a (positive) minimum value otherwise. The precise value of the uptake rate is taken to be a reasonable value from previous studies of extracellular dopamine levels [31]. Following [32], dopamine re-uptake was presumed to be governed by a Michaelis-Menten type kinetics equation (12)where denotes the maximal uptake rate and is the Michaelis-Menten constant. Thus, a closed form representation of the considered dopamine model is (13)
As mentioned above, dopamine was assumed to affect the firing of the LMC by altering neural ion channel permeability. Thus, the effect of dopamine on potassium channels can be seen as a dependence of the gain in on dopamine concentration. Hence we modify the equation governing the fraction of open potassium channels in the LMC as follows (14)where denotes a dopamine dependent gain. In the absence of dopamine we want a gain of unity, i.e., . Conversely, we also like to impose an upper bound on the gain. To achieve this, consider the expression (15)where is an antagonism parameter controlling the overall impact of dopamine on the gain . Obviously, if then , thus, by setting , a unity gain for a dopamine concentration of zero is established. Since and assuming physiologically meaningful dopamine concentrations, i.e., , sets an upper bound for the gain.
Finally, we modeled the impact of dopamine on calcium and sodium channels in the LMC using the gain . We expressed the dopamine dependence of the permeability of those channels via varying the LMC's excitatory-to-excitatory functional synaptic coupling by introducing (16)
In the absence of dopamine we have and thus . Rising dopamine levels increase and, in turn, until reaches its previously established upper bound , which gives . Thus, we have the estimate (17)
To establish a closed form representation of the full model, let be defined by , where is given by the right hand side of (9) with LMC components defined by (18)for . Let further be given by the right hand side of (6) and define (19)
Similarly, let be given by with and as defined in (7) and . Then with the full neural dynamics model can be written as (20)where is called drift (or deterministic force) and is the diffusion (or random force) of the model. In the following section we discuss an efficient strategy to numerically approximate solutions to the model (20). A rigorous mathematical proof establishing existence and uniqueness of those solutions is presented later.
We used time discrete approximation techniques to simulate sample paths of the SDE system (20). Extensive numerical experiments revealed pronounced non-linear dynamics of the model, which motivated the use of a higher order solution scheme. We encountered numerical instability of the widely used strong order 1.0 Milstein scheme [33]. Using a strong order 1.5 explicit Runge-Kutta (RK15) method, however, proved to be reliable. An explicit strong order 2.0 scheme yielded no notable improvements over the RK15 method but required a considerably higher computational effort. Thus, a RK15 scheme was specifically adapted to the model (20).
To establish a time discrete approximation of the solution to (20), we started by defining a discretization of the interval . For , let be a step-size and define discrete time points for . We introduce the Markov chain to approximate the stochastic process that satisfies (20). Thus we set and . Note that (20) is a -dimensional non-autonomous SDE with constant additive scalar noise. This latter property is exploited to construct a highly efficient recursive solution scheme that has a considerably reduced computational cost compared to a general purpose SDE solver.
The following considerations are based on the family of solution schemes presented in [[26], Sec. 11.2]. The vector form of an explicit order 1.5 strong scheme for a non-autonomous SDE with constant additive scalar noise is given by (21)where (22)and is a random variable representing the following double stochastic integral (23)
Rearranging terms, (21) can be simplified to (24)
Note that and is also normally distributed satisfying (25)as shown in [[26], Chap. 10]. These properties play a key role in practice since they allow us to generate the pair of correlated random variables and in an efficient and straight forward manner: let and be independent distributed random variables, then (26)
Thus, an approximate solution of (20) was recursively computed following scheme (24) with auxiliary quantities (22) and (26).
Note that (24) requires three evaluations of the drift term per step. In contrast, the Milstein method adapted to model (20) reduces to (27)and thus requires only one function evaluation per step. However, unlike RK15, (27) reduces to an explicit Euler scheme in the absence of noise (zero diffusion). Thus numerical instability of the Milstein scheme for a model like (20) exhibiting pronounced non-linear characteristics in the drift term was predictable. Note that it is possible to enforce convergence of (27) by substantially reducing the step-size . However, this in turn dramatically increases the total number of time-steps making the overall computational performance of the Milstein method significantly worse than that of RK15 (24). Hence RK15 was the solver of choice for all simulations presented below.
In order to produce measureable changes in extracellular dopamine levels, which reflect rapid phasic dopamine release during a behavioral task or a pharmacological challenge, the dopaminergic axons must be stimulated at frequencies of 10-20 Hz or greater [34]. Because phasic dopamine release may reach high concentrations for brief periods due to concerted burst firing of dopamine neurons [34], [35], [36], we tested our model at a neural firing rate>20 Hz with different time-step sizes. We found that a small step-size of 0.1ms had the highest numerical robustness and showed the optimal temporal resolution of neural firing in order for dopamine release/re-uptake to set in gradually, without jumps.
The simulations shown below have been run on a Mid 2010 Mac Pro (2×2.66 GHz 6-Core Intel Xeon, 24GB DDR3 ECC RAM) under OS X 10.9.1. All codes have been written in Python [37] making extensive use of the packages NumPy, SciPy [38] and Matplotlib [39]. Computationally expensive sections of the code have been converted to C extensions using Cython [40].
Below, we start by showing existence and uniqueness of solutions to the model (20) (Theorem 1). Once this fundamental result is established, we present simulations generated by the model and analyze it with respect to empirical fMRI data.
The following result guarantees unique solvability of the model (20).
Theorem 1. For and let . Then the system (20) has a unique -continuous solution .
Proof. If we show boundedness and Lipschitz continuity of and on , then existence and uniqueness of a solution to (20) follows from Theorem 5.2.1 in [22].
We start by proving that and are Lipschitz continuous. Obviously as a constant trivially satisfies a Lipschitz condition. Since linear and trigonometric functions are differentiable (and thus Lipschitz continuous), we only have to show that the Michaelis-Menten kinetics equation (12) is Lipschitz continuous with respect to dopamine. Thus, for a straight-forward calculation yields (28)where we used the reverse triangle inequality and the fact that for all . Hence, is Lipschitz continuous in and thus all component functions of are Lipschitz which makes the entire mapping Lipschitz continuous on .
Next, we show that and satisfy (29)where is a positive constant and denotes some vector norm on Since all norms on a finite dimensional linear space are equivalent, we prove (29) for the maximum norm . We start by showing boundedness of all components of given by the right hand side of (9) with LMC equations (18). First, note that all firing rates (4) are bounded by and respectively. Furthermore, the rates of open ion channels defined by (3) and (14) respectively are bounded by 1. Thus, the neural activation function for potassium channels (14) satisfies (30)
Weights connected to region may be estimated by (31)where denotes the i-th row of the matrix . Thus, let be a vector in and consider the i-th component of as given by the right hand side of (9) with LMC components (18). To simplify notation we introduce a vector such that (32)with given by (16) for . Hence, by (17), all components of satisfy . Thus we obtain the following estimate for the first term of (33)where we used (31) and the fact that . Note that all terms subsumed in the constant are independent of and . Similarly, we establish (34)
Using (30) we further obtain (35)
Finally, we establish (36)and due to (37)for . Thus combining (33) - (37) yields (38)where . Analogously to (37) we compute (39)and hence readily obtain (40)
Finally, by (12) we have , and thus we get the following estimate for (13) for any (41)where we used and . Thus we obtain (42)
Combining estimates (38), (40) and (42) for hence yields (43)
This together with the definition (7) of eventually gives (44)which establishes (29) with and concludes the proof.
Having established existence and uniqueness of solutions to the model (20), we now present simulations corresponding to the resting state and dopamine modulation and compare them to empirical fMRI data.
Using the coupling matrix described above, brain activity was simulated corresponding to the resting state and task-induced dopamine release. A list of all used parameters is provided in Table 1, which were taken from literature and scaled appropriately to reflect units used in this work or manually estimated based on previously published values [21], [17], [31]. Physiological variations across simulated brain regions were modeled by normally distributing inhibitory-to-excitatory and non-specific-to-excitatory synaptic coupling strengths using a fixed random number generator seed across simulations. This introduced the possibility of regionally desynchronized temporal dynamics in the model allowing simulated neural nodes to evolve non-identically over time in the absence of inter-regional coupling. Note that all simulations below were run with the same initial conditions and parameter values, i.e., starting values and parameters were identical for the resting state and dopamine-modulated speech-related simulations.
In both resting-state and task simulations, complex spatio-temporal patterns of activity emerged. Fig. 1B illustrates the temporal dynamics of the left LMC with and without dopamine modulation. The left panel shows the time-course of the left LMC's excitatory membrane potential overlayed with the corresponding time-evolution of . While shows similar behavior during rest and task simulations in the absence of dopamine, the time-course is being visibly altered as soon as release increases. Thus, increasing dopamine levels in the task simulation changed LMC membrane potentials noticeably, which in turn raised the firing rates of LMC neurons. This increase in () was distributed throughout the entire network, subsequently changing local neural dynamics of other brain areas. The right panel of Fig. 1B shows the time-course of for fifty simulated speech cycles. Note that the propagation of firing rate changes acted as a neural feedback loop on the SNc itself in that repeated dopamine release caused different activity patterns than preceding cycles. In the task simulation, the LMC exhibited on average slightly higher firing rates than during rest (rest: , task: , compare also Fig. 1B) in agreement with the initial modeling assumption. To highlight that the proposed dopamine release model indeed shaped the dynamics of the entire neural population, the following section discusses changes in the correlative structure of simulated brain activity under dopamine modulation relative to the resting state.
The raw model output was converted to BOLD signals as detailed above. Fig. 2 shows simulated and real BOLD signals for a selection of speech-related ROIs (Fig. 2A,B). Simulated BOLD signals with and without dopamine modulation were compared to empirical resting-state and speech production fMRI data, respectively, in order to assess the global effects of dopamine modulation on the entire simulated neural population. To do so, we employed graph theory analysis to quantify variations in functional connectivity between the resting state and speech production. Thus, we first had to quantify statistical similarity between two time-series. We chose the normalized mutual information (NMI) [47] as statistical metric. Hence, for two random variables and , let and denote their respective Shannon entropies [48] and define
(45)where denotes the raw mutual information between and . Hence, unlike the original formulation of the mutual information , which is not bounded from above [[49], Chap. 2], the NMI is normalized by the geometric mean of the entropies and . Thus, takes values between zero (two signals are independent) and one (two signals mutually depend on each other), permitting unambiguous comparison of values across data sets.
Pairwise interactions in the simulated BOLD signals with and without dopamine modulation were quantified by computing NMI coefficients for each pair of ROI time-series. Analogously, NMI matrices were computed for the group-averaged resting-state and speech production BOLD data. This gave rise to four NMI-matrices (model rest, model speech, data rest, data speech) (Fig. 2C,D). Visual inspection of the matrices revealed larger variability in the model's correlative structure than in the corresponding empirical data. This might be partly explained by the fact that the empirical data were averaged across twenty subjects in an attempt to minimize subject-specific effects. Averaging a number of simulation runs would possibly decrease variability in the model; however, the aim of this study was to establish a qualitative assessment of the presented dopamine release model with respect to global effects seen in empirical data. In that respect, the proposed model, incorporating a single dopaminergic link between the SNc and laryngeal motor cortex, modulated neural activity of the whole brain to an extent that differences were observed between the structure of model's functional connectivity during dopamine release and the resting state. In addition, the model's prediction of empirical functional connectivity during speech production was in good alignment with the data.
In the following, we discuss simulated and empirical functional connectivity using the framework of graph theory. Interpreting functional connectivity matrices as graphs allowed us to not only reveal the functional topology and connectivity architecture of data and model but to also rigorously quantify the observed differences using well-established network metrics (see Supporting Information). By interpreting the 70 ROIs as nodes of a network with the associated NMI-coefficients representing the weights of the graph's edges, we constructed four weighted undirected graphs. Note that with the NMI being always non-negative (contrary to the classical zero-lag Pearson correlation coefficient) a graph-theoretical analysis of NMI networks is straight-forward. Without the need to either extend classical metrics to negatively weighted graphs or consider negative and positive edges separately, most graph measures can be readily applied to NMI networks.
The four weighted, undirected networks were analyzed following the concepts of functional integration, segregation, and influence [50]. As a measure of integration, we considered the local efficiency of a node , , quantifying a node's local communication performance in terms of inverse shortest path lengths within its neighborhood [51]. The degree of functional segregation was estimated using the weighted local clustering coefficient , which was calculated as the average geometric mean of edge weights in triangular motifs around [52]. Nodal influence was approximated based on nodal strength and nodal degree . A node's strength is the sum of attached edge weights, while its degree is defined as the number of connected edges. Clustering coefficient and efficiency were also compared to corresponding values of 100 conservatively-configured, null-model random networks. Normalized clustering coefficient and efficiency were computed by dividing and by the respective random network values. Statistical significance of differences in network metrics between the resting state and task production was determined using a paired two-sample permutation test at adjusted for family-wise errors (FWE) using the maximal statistic [53]. All graph metrics were calculated based on the full networks in their original density without applying any thresholding strategy. Since density-reduction techniques may severely deter network topology [54], [55], [56], [57] and might thus dilute subtle differences between simulated and empirical functional connectivity patterns, the presented analysis is focused on the full un-thresholded NMI networks as suggested by [58]. Graph metrics were computed using a Python port (pypi.python.org/pypi/bctpy) of the Brain Connectivity Toolbox for MATLAB [59].
We presented an extension of a model of neural assemblies proposed by Breakspear et al. [21] to simulate dopamine release in the human brain during complex voluntary behaviors. In contrast to other large-scale neural modeling techniques based on coupled oscillator systems, our approach was grounded in neuroanatomy and physiology and thus allowed us to design a dopamine release model guided by biological considerations. We established unique solvability of the proposed model and demonstrated a computationally efficient strategy to numerically approximate its solutions.
In the context of the model, we assumed the difference between the resting state and speech production to be solely given by a modulation of dopamine levels in the LMC via a direct input from the SNc. Thus, the model was oblivious to task-related effects caused by any neurotransmitter other than dopamine. Importantly, we observed pronounced differences between the resting state and task production in simulations. This finding may be interpreted as an indication of the profound physiological impact of dopamine on brain dynamics.
It is remarkable that altered neural firing rates within the bilateral LMC only were sufficient for the entire simulated neural population to exhibit changes in its temporal dynamics. We attribute these observed task differences to dopamine driving neural dynamics via the coupling matrix . Given real structural connectivity data as input, the strength of the model lies in its ability to reproduce observed properties of connectivity during a dopamine-modulated activity with a biophysical prescription for dopamine neurotransmission.
To quantify the impact of dopamine release on the entire neural mass, we used functional connectivity and graph theory analysis and interpreted the results as a stationary synopsis of the global effects of dopamine modulation. The graph theoretical analysis of the functional connectomes revealed a number of similarities between the model and data. Due to the slightly higher variability of NMI coefficients, especially during the rest, (Fig. 2) nodal strengths of the model without dopamine modulation showed more fluctuations than corresponding values for the data. Nonetheless, strength values of simulated and empirical networks were in good agreement with the model exhibiting slightly larger values. Clustering coefficients of both model and data also showed qualitatively similar attributes, when comparing the resting state to dopamine modulation. Thus, the model showed characteristics comparable to those of the data with respect to functional segregation and nodal influence.
Similarly, local efficiency showed good qualitative agreement between simulated and empirical connectomes, thus the model mimicked functional integration patterns seen in the experimental data. However, both model and data failed to show increased network segregation and integration compared to random graphs during rest but showed consistently larger values than null model networks during speech production (all and greater than one for ). This may support earlier findings indicative of pronounced changes in network organization for speech control [63], [64]. It should be emphasized that normalized efficiency is computed using the notion of shortest paths within a network. Due to the absence of zero-weighted edges in the considered networks, the shortest path between any two nodes in the graphs was given by their connecting edge, effectively side-stepping the notion of paths in a graph. A thresholding strategy to eliminate ‘weak’ edges (i.e., edges corresponding to small NMI coefficients) may have somewhat remedied this problem. It should be noted, however, that interpreting efficiency values obtained from functional networks that are based on statistical similarity between brain areas is not an immediately evident approach. Indeed, since NMI networks express not only direct but also all indirect couplings between regions, a path-based metric, like efficiency, may yield ambiguous results (see, e.g., [65], [66]).
Nevertheless, decreasing connection densities in the networks would also yield non-trivial nodal degree distributions, opening another perspective on nodal influence within the networks. However, weight-based thresholding must be performed with considerable precautions so as to not deteriorate topological properties of a network. Since this work was mainly concerned with establishing a biologically-informed, large-scale model with optional dopamine neuromodulation, no thresholding strategy was applied to the constructed functional networks. A future study focused exclusively on the graph-theoretical analysis of functional networks should address this issue.
A visual inspection of the simulated and empirical functional connectomes (Fig. 2) revealed that the model tended to slightly overestimate regional pairwise interaction during both resting state and dopamine modulation. This finding was not surprising considering the fact that the simulated BOLD signals were generated by 140 structurally identical equations that only differed in some parameter values. This is an apparent limitation of the presented approach. However, one of the advantages of the presented model is that the employed strategy enabled us to perform large-scale simulations of brain activity based on considerable neurobiological detail without becoming too complex to be practically unfeasible.
Coupling between regions in the model was achieved via scaling excitatory neural firing rates by entries of the coupling matrix (compare eqn. (9)). Thus, all modeled axonal connections were excitatory, which is a simplification that ignores the effects of feedforward inhibition. In particular, firing of connected areas impacted a region's membrane potential through excitatory projections targeting local populations of NMDA and AMPA receptors. In other words, inter-regional coupling was not modeled as an explicit consequence of changes in neural voltages of neighboring areas. Instead, the influence of other regions on the local membrane potential was mediated by changes in neural firing rates. In this context it should be noted that the proposed model did not include an explicit representation of inter-regional axonal conduction delays. To some extent, however, the employed form of indirect coupling in the model may be interpreted as a lumped representation of conduction time delays.
Having tested the model for its efficacy in reproducing essential features of real data from healthy humans during speech production, the next step should be an examination of clinical relevance of the proposed neural population model. This may be achieved by incorporating ‘lesions’ into a simulated network of interest to investigate the extent of inter-regional influences coupled with dopaminergic transmission in a range of neurological and psychiatric disorders, such as Parkinson's disease, dystonia, schizophrenia, etc.
Furthermore, the model's use is not limited to human applications [67] and may be applied equally well to animal models of disease and normal behavior, taking into account appropriate modifications for differences in animal and human dopaminergic innervation [68].
Since most parameters used here were taken from literature, some inferences about trajectories of isolated regions can be formulated based on the exhaustive analytical treatment of the original model [21]. In the original model, inter-regional coupling is introduced using a scalar parameter that acts on the spatially averaged excitatory firing rates of all modeled nodes, i.e., . The approach presented here uses not a scalar, , but a matrix, , to introduce coupling and thus expands the scaled mean field firing to be . This can be seen as a weighted average of firing rates. Thus, in the absence of dopamine and for a diagonal coupling matrix the dynamic behavior of an isolated node can be reduced to the cases discussed by Breakspear et al. [21]. However, for a general non-diagonal coupling matrix, the dynamics become increasingly more complex. Moreover, as our simulation results indicated, dopamine modulation also had a pronounced impact on the overall behavior of the model. Thus, the extensions proposed here changed the dynamics of the original model in a non-trivial way. Thus, a rigorous dynamical analysis of the presented model would require a thorough study of the non-linear relation between and () and an assessment of the influence of dopamine-related parameter choices on the temporal evolution of the LMC nodes in terms of a full sensitivity analysis [69], [70], [71]. It was not within the scope of this work to present such an exhaustive analytical treatment of the model. Nonetheless, this poses an interesting direction for future studies.
Given the demonstrated differences in functional connectivity across the entire experimental time in simulations of resting versus speech conditions, the question arises as to what extent dopamine altered function on small versus long time scales within the tasks. Our results indicate that dopamine may influence dynamics on long time scales. This may suggest that rapid temporal release of dopamine, as evidenced by the spontaneous dopamine release incorporated during each time-step in the model, may be involved in slow plastic responses. Thus, it is tempting to speculate that a future adaption of the proposed dopamine model might yield further insight into the learning and adaptation involved in voluntary behaviors, particularly given dopamine's involvement in learning and motivational behavior in other tasks.
Finally, models simpler than the one considered in this paper are capable of reproducing empirical functional connectivity. In fact, a recent study showed that a stationary model of resting-sate functional connectivity explains functional connectivity better than more complex models [72]. In modeling empirical functional connectivity as accurately as possible, the application of a complexity reduction technique [73] to the introduced highly non-linear model should be considered in order to derive a set of considerably simpler equations of statistical moments. On the other hand, it has been shown [73] that functional connectivity is essentially state-dependent and that local changes of activity in a set of cortical areas (due to external inputs, attention, neuromodulation, or learning) change the dynamical state of the brain network, thus modifying the correlations between the brain areas and introducing various levels of complexity. Along this line, while simpler models have a number of computational advantages (e.g., reduced computational load, easier estimation of parameters, simpler relationship between structure and function), their ability to simulate complex temporal activity patterns at various cognitive scales (and in the context of simulated dopamine modulation) may be somewhat limited. This motivated the development of the proposed complex model to better understand empirical data and to make predictions about the different states of dopamine-modulated brain activity during voluntary behavior. Future work should be directed to a possible simplification of this model, while assuring its ability to accurately reproduce the complex biological patterns of voluntary behaviors.
We conclude that a regional model that includes dopamine release, reuptake, and modulation of ion channels significantly alters the behavior of an otherwise unmodulated, resting state neural population model. This work thus combines a small-scale basic cellular biology understanding of dopamine to alter macroscopic behavior of neuronal systems with nontrivial structural circuity, and presents meaningful global simulated fMRI network behavior. Region-specific analysis warrants the identification of specific effects of neuromodulation on task-based networks for speech and other dopamine-modulated voluntary behaviors.
|
10.1371/journal.ppat.1002604 | Chemoenzymatic Site-Specific Labeling of Influenza Glycoproteins as a Tool to Observe Virus Budding in Real Time | The influenza virus uses the hemagglutinin (HA) and neuraminidase (NA) glycoproteins to interact with and infect host cells. While biochemical and microscopic methods allow examination of the early steps in flu infection, the genesis of progeny virions has been more difficult to follow, mainly because of difficulties inherent in fluorescent labeling of flu proteins in a manner compatible with live cell imaging. We here apply sortagging as a chemoenzymatic approach to label genetically modified but infectious flu and track the flu glycoproteins during the course of infection. This method cleanly distinguishes influenza glycoproteins from host glycoproteins and so can be used to assess the behavior of HA or NA biochemically and to observe the flu glycoproteins directly by live cell imaging.
| Enveloped viruses such as the influenza virus cause significant disease in humans. In order to cause a productive infection, the virus particle must interact with the host cell using the viral proteins encoded within its genome. For many such viruses, it is possible to directly observe the early steps in infection, yet for technical reasons it has been extremely difficult to study the genesis of daughter virions as they bud off of infected host cells. Here we devised a chemoenzymatic labeling strategy to site-specifically append probes to the influenza hemagglutinin (HA) and neuraminidase (NA) proteins using the bacterial sortase A enzyme. Because labeling is confined to surface exposed HA and NA in the context of live, infected cells, it is possible to study budding biochemically and microscopically in real-time. Using this system, we can observe budding of flu virions from discrete sites at the cell surface. Our work will enable detailed investigation into the birth of viruses from infected host cells and can likely be applied to viruses other than influenza that have been similarly resistant to real-time microscopic observation during budding.
| Enveloped viruses are composed of elements produced by and recruited from the infected cell. The formation of new virus particles occurs either on intracellular membranes or at the plasma membrane. The assembly of a nascent virion requires the coalescence of the envelope (glyco)proteins embedded in a proper lipid environment, and the recruitment of matrix and nucleocapsid proteins together with the viral genome [1]–[3]. How apposition of envelope components and viral genomes are controlled as a means of ensuring production of infectious progeny is not well understood.
The influenza virus particle contains a segmented, negative stranded RNA genome encoding 11 proteins, two of which the virus uses to interact with the host cell membrane [4]. Hemagglutinin (HA), a type I transmembrane protein, binds to sialoglycoconjugates on the surface of the host cell and mediates entry of the viral particle [5], [6]. HA also mediates fusion of the viral and host cell membranes to effectuate genome delivery to the cell to be infected [7]. Neuraminidase (NA), a type II membrane protein, is a sialidase that assists in release of virions from the infected cell [8].
The inability to label either flu HA or NA in a manner that allows continuous monitoring of surface disposition, surface distribution, and release has hampered the study of flu particle biogenesis. The use of antibodies, while feasible in principle, requires their introduction as fluorophore-conjugates that would crosslink viral proteins unless used as monovalent F(ab) fragments. Moreover, this labeling method is indirect. Studies that address particle biogenesis have also mostly used fixed cells and by design have not addressed virus release in real time. Visualization of the influenza glycoproteins in living cells demands a method for site-specifically modifying HA and NA, at the exclusion of all host proteins inserted into the very same membrane. We know of no successful attempts to achieve this by genetically tagging the flu glycoproteins with fluorescent proteins or with other methods that yield visible HA or NA by covalent modification in the context of an infectious virus.
We and others have developed a site-specific labeling method that exploits sortase transpeptidases found in gram positive bacteria [9], [10]. These enzymes cleave the five amino acid LPXTG recognition sequence between the threonine and glycine residues, forming an acyl-enzyme intermediate that is resolved by nucleophilic attack by the N-terminus of an oligoglycine peptide, forming a new amide bond. This reaction is portable: upon incubation with recombinant sortase A, proteins that carry an LPXTG motif are readily labeled with oligoglycine-based probes bearing a broad range of functionalities [11], [12]. The incoming nucleophile may carry any desired substituent for attachment, including fluorophores, biotin, lipids, or may even consist of other polypeptides- see [13], for review.
Here we report the creation of two influenza A/WSN/33 strains bearing the sortase cleavage site in the HA and NA proteins respectively. Infection of host cells with such strains allows site-specific labeling of HA or NA and allows us to observe the products of influenza infection in real time. We can thus visualize and examine biochemically the events that immediately precede viral release from the host cell surface, as well as the release of newly formed virus particles. The ability to execute sequential labeling reactions employing distinct tags allowed us to observe preferred sites from which virus particles are released.
While purified virus particles can be labeled with lipophilic dyes, the dequenching of which reports on fusion of the incoming viral envelope with target endosomal membranes [14], [15], the production of new virions is more difficult to visualize. Neither flu neuraminidase nor hemagglutinin tolerate fusion to fluorescent proteins or other modules that allow site-specific covalent attachment of fluorophores—attempts to do so are incompatible with virus production and assembly. For N-terminal fusions to NA, this is likely the result of failure to insert into the ER during its biosynthesis. The bulky GFP moiety, when fused to the C-terminus of NA, likely compromises its functional activity and oligomerization. For HA, the only viable option would be to place GFP at the C-terminus, but fusions as small as a (His)6 epitope already impair virus assembly and fusogenic activity of HA [16]. Fusions to the N-terminus of mature HA have been reported [17], but with the caveat that such fusions undergo significant proteolysis and yield substantial amounts of wild-type HA protein. The inability to specifically label the flu glycoproteins for biochemical and visual observation has hampered an analysis of the virus budding process.
We devised a sortase-based labeling method to overcome at least some of these limitations, and our findings are likely to be more generally applicable to other viruses with problematic labeling characteristics when relying on fusion with fluorescent proteins. The sortase labeling (sortagging) method is particularly well-suited for labeling of type II membrane proteins, as has been done for CD40L [11], CD74 and Dectin-1 [18] at the extracellularly exposed C-terminus. We installed an LPETG motif at the C-terminus of the A/WSN/33 NA protein, followed by an HA epitope tag (this epitope is absent from the A/WSN/33 hemagglutinin) (Figure 1A). Because the portion distal to the cleavage site is lost upon sortagging, the presence of the epitope tag allows monitoring of material not accessible to the enzyme in intact cells. We used the 12 plasmid reverse genetics system to generate recombinant A/WSN/33 flu particles bearing this sortaggable NA construct [19], and found that the resulting virus was infectious and indistinguishable from wild type virus in its ability to replicate in vitro (Figure 1B). We refer to this strain as NA-Srt.
HA is a type I membrane protein, synthesized as an HA0 precursor, which requires proteolytic cleavage by a trypsin-like activity to generate the disulfide-bonded HA1 and HA2 subunits. This cleavage exposes a key glycine residue at the N-terminus of HA2 that is essential for HA2 to retain its fusogenic activity [5]. We created a version of HA that allows trypsin cleavage in the loop that connects HA1 and HA2, with concomitant exposure of the sortase recognition sequence (placed immediately upstream of the trypsin cleavage site). Cleavage is likely to improve accessibility of the LPXTG motif, a requirement for efficient labeling [11], [20]. We therefore generated a recombinant A/WSN/33 flu strain bearing this sortaggable HA, and found that this strain, too, was not attenuated in vitro (Figure 1C). We refer to this strain as HA-Srt. The behavior of the NA-Srt virus was indistinguishable from that of the WSN parental strain when assessing virulence by monitoring weight loss in mice. Mice infected with sublethal doses of the HA-Srt virus also showed weight loss, albeit somewhat reduced compared to mice infected with wild-type virus. (Figure 1D). We conclude that the installation of a sortase tag on either NA or on HA does not seriously impair virus assembly, virus release and infectivity in vitro and in vivo.
We obtained NA-Srt virions by sedimentation from the supernatants of infected MDCK cells, and subjected this material to sortagging with a biotinylated probe, followed by detection of biotinylated material by immunoblotting. Only in the presence of sortase and probe did we detect specific labeling, accompanied by the loss of the HA epitope tag, as expected. Based on the intensity of the HA-positive materials recovered, we estimate that the labeling efficiency of HA in the sedimented virus is approximately 70–80% for the conditions used (Figure 2A). This value is not atypical for sortase-mediated labeling reactions, which usually proceed to near-completion [11], [21]. Using sortase we similarly installed TAMRA or Alexa647 dyes on intact virions pelleted from tissue culture supernatant or further purified through a sucrose gradient (Figure 2A and Figure 2B). Dimerization of the NA-Srt protein incorporated into gradient-purified virions is unaffected by the LPETG tag (Figure 2C)
We infected MDCK cells with the NA-Srt virus and at different times post infection, we subjected cells to sortagging with a biotinylated triglycine-based probe (Figure 2D). We detected biotinylated, surface accessible NA by immunoblotting using streptavidin-HRP. We assayed for the unlabeled and cell-internal pools of NA by reactivity with an anti-HA epitope antibody. NA was first detectable at ∼1 hr post-infection and its levels peaked at ∼4 hrs, after which we observed no further increase. Sortase-mediated surface biotinylation of NA was detectable at 4 hrs post-infection and steadily increased over the duration of the experiment (7 hrs). Specificity of labeling is excellent: we observed no host cell proteins modified with the biotinylated probe.
Cells infected with the HA-Srt virus can be similarly labeled. Glycosidase digestions confirm that sortase labels only the cell surface pool of flu glycoproteins (Figure 2E), as follows. We labeled intact cells infected with either the HA-Srt or NA-Srt viruses with sortase using a biotinylated probe, followed by lysis and digestion with either Endoglycosidase H or PNGaseF. Immunoblotting showed that all of the biotinylated HA-Srt protein is partially EndoH-resistant, indicating successful traversal of the secretory pathway. Because mature NA and HA carry both complex-type and high mannose-type oligosaccharides [22], [23], resistance to digestion with EndoH is always partial, as seen by comparison with the PNGaseF digestion product. The entire unlabeled, anti-HA reactive pool of NA-Srt protein was fully EndoH-sensitive, however, as evident from a comparison with the PNGaseF digestion products. The fraction of NA-Srt inaccessible to sortagging is thus indeed composed of cell-internal NA-Srt protein. We conclude that only the cell surface pool of influenza glycoproteins, poised for incorporation into nascent virions, is labeled upon incubation of infected cells with sortase and a suitable probe.
Having established the specificity of labeling of the sortagging method, we examined the biogenesis of flu virions and their release from infected cells through biochemical analysis and by live cell imaging. We labeled flu HA-Srt protein on the surface of infected and metabolically labeled cells by exposure to trypsin, followed by sortase-mediated installation of a single biotin at the C-terminus of HA1. We performed a pulse-chase experiment to examine the kinetics of arrival of HA-Srt protein at the cell surface, and its subsequent release from the infected cell as assembled virions. We did not ascertain the presence of all subgenomic RNA fragments in the material released from the infected cell, as we have no means of testing whether individual particles carry a full complement of subgenomic RNAs, or whether the released materials contain substantial amounts of defective particles with incomplete sets of RNAs. However, our data are consistent with the release of HA-Srt into the medium corresponding to assembly and release of progeny virions (see below).
We initiated metabolic labeling with 35S labeled Cysteine/Methionine at ∼5 hours post-infection, a time when robust viral protein synthesis is ongoing. For some experiments, we infected cells at a low multiplicity of infection (MOI) (Figure 3A), and started metabolic labeling at 14 hrs post-infection, when most of the cells are infected by progeny HA-Srt virus produced by the cells infected initially. In this setting, we observe a similar if not greater amount of labeling than for cells infected at a high MOI (Figure 3D), indicating that the LPETG-tag is neither lost nor interferes with virus replication (compare Figure 3B and Figure 3E). During the chase, we performed sortase labeling for 30 minutes at the indicated time points. We lysed the 35S-labeled sortagged cells and subjected them to affinity purification on a neutravidin-agarose matrix to recover the biotin-modified HA1 and associated proteins (Figure 3B). At the 0 min time point of this experiment (it includes a 30 minute incubation with sortase, during which intracellular transport of glycoproteins continues), no labeled HA1 is recovered, indicating that the newly synthesized pool of HA requires at least 30 minutes to reach the cell surface. We observe a steady increase in surface exposed (sortase accessible) HA-Srt protein as well as a minor fraction of associated, uncleaved HA0. We attribute the presence of this HA0 to incomplete cleavage of the HA trimer by trypsin added to the medium. In this manner we recover -as part of a trimer- some HA0 devoid of biotin, along with the sortase modified, biotinylated HA. As expected, we do not observe biotinylated HA0 by streptavidin blot (data not shown). We do observe a small amount of HA0 at the cell surface at the 0 min timepoint. This we attribute to a minor portion of labeled HA1 not detectable via autoradiography. The HA0 recovered at early time points is composed of both the mature HA0 and the high mannose intermediate (Figure 3C). As expected, HA1 and HA2 are recovered together because of their covalent association, which persists after cleavage of HA0. We examined the behavior of HA-Srt protein at later chase times (Figure 3E and Figure 3F) and again observed an increase in labeling, after which the amount of labeled HA-Srt protein decreases. By 10 hours of chase, less than 20% of the material that successfully reached the cell surface and is labeled by sortase is retrieved from the cells, indicating that most, but not all HA labeled in the course of the pulse is released from the cell, presumably as intact virions. As infected cells show clear signs of cytopathic effects at late time points post-infection, cellular functions required for virion assembly are likely to be compromised, thus preventing complete release of all viral products, a situation that likely applies in vivo as well.
To determine the fate of surface-labeled HA-Srt, we subjected infected cells to metabolic 35S –Met/Cys pulse labeling, followed by a chase period of 2 hours to allow radioactive HA-Srt protein to accumulate on the cell surface (Figure 4A). We labeled intact infected cells with sortase A and a biotinylated probe and recovered biotinylated HA-Srt protein from cell lysates as well as from the media (Figure 4B and Figure 4C). We observe a gradual increase in HA-Srt released into the media over time, corresponding to the rate of loss of biotinylated HA-Srt from the cell surface (Figure 4D bottom panel). However, we detect more released HA-Srt protein than is accounted for by the loss from the cell surface (Figure 4D, top panel). We attribute this difference to the fact that biotinylated HA-Srt is tightly associated with non-biotinylated HA-Srt in intact virus particles, which are retrieved by the neutravidin-agarose matrix along with the sortagged fraction. Metabolically labeled HA-Srt does not bind non-specifically to this matrix, as virus-infected cells exposed to the biotinylated probe in the absence of added sortase do not show any signal.
Does the decrease in cell-associated, biotinylated HA1, accounted for by the appearance of biotinylated HA-Srt protein in the supernatant, correspond to the release of virus particles? To demonstrate this, we recovered from tissue culture supernatants radiolabeled viral proteins not modified by sortase through adsorption of released virions onto chicken erythrocytes (Figure 4E), and visualized the adsorbed materials by SDS-PAGE and autoradiography. Because binding of the virions to erythrocytes occurs via HA, any other protein recovered by low speed sedimentation of erythrocytes must be part of an adsorbed virus particle. Indeed, we detect the other viral proteins upon suitable exposure of the autoradiograms. These include polypeptides with the assigned molecular masses of the RNA polymerase subunits, as well as NP and M1, all of them in quantities proportional to their methionine/cysteine content and to the reported copy numbers in intact virions [24]. We next compared the kinetics of HA accumulation in the media for the fractions recovered via Neutravidin-agarose or on chicken erythrocytes. Levels of total HA at the 1 hr time point were quantified and all other time points were normalized to these values (Figure 4F). We observe indistinguishable kinetics for HA-Srt accumulation in the media, underscoring our conclusion that biotinylation does not affect budding of HA-Srt.
Having established the specificity of the labeling reaction and the ability of the labeled flu glycoproteins to be incorporated into virions and released into the culture supernatant, we next visualized virus budding and release by labeling biotinylated, surface exposed HA-Srt protein with streptavidin-modified quantum dots. To examine the behavior of surface disposed HA-Srt and its release from infected cells using a similarly modifiable control protein as a reference, we generated an MDCK cell line stably transduced with CD154/CD40L, equipped with a sortase tag [11]. Like the NA-Srt and HA-Srt proteins, this molecule is readily labeled with biotin using sortase, yet should not be actively incorporated into nascent virions and so allows for a direct comparison with flu HA-Srt.
Incubation of MDCK cells that display biotinylated HA-Srt or CD154 were readily labeled with quantum dots. We examined the fluorescence intensity as a function of time after labeling by cytofluorimetry of infected, labeled cells (Figure 5A). Whereas the levels of fluorescence recorded for labeled CD154/CD40L were constant, those for labeled HA-Srt decreased exponentially over the first few hours of incubation. When we infected labeled CD154/CD40L-expressing cells with wild type WSN virus, we also observed constant staining intensity but at a lower level, presumably because host protein synthesis was much reduced in the virus-infected cells. Labeled CD154/CD40L is obviously not incorporated into budding virions.
We labeled HA-Srt infected MDCK cells grown on glass coverslips with sortase A and biotinylated oligoglycine probe, followed by staining with streptavidin-functionalized quantum dots. Cells prepared in this manner were directly imaged by spinning disc confocal microscopy to observe the behavior of the HA-Srt protein during the course of virus production and release. The influenza glycoproteins are known to be inserted into host cell membrane regions referred to as lipid rafts, operationally defined as insoluble in non-ionic detergents. One key component of lipid rafts is the ganglioside, GM1, the levels of which oscillate with cell cycle status [25]. We co-stained HA-Srt labeled cells with fluorescently labeled cholera toxin (CTx), a marker for GM1, and observed variable CTx staining between MDCK cells, presumably because of asynchronous growth and rapid internalization of CTx [20]. Uninfected MDCK cells show similar heterogeneous CTx staining, indicating that virus infection is not the cause of this variability (data not shown). We observe co-localization of CTx with patches of quantum dot staining in HA-Srt infected cells (Figure 5B), but in CD40L/CD154 labeled cells, quantum dot staining shows only partial overlap with CTx staining, reflecting broad distribution of labeled CD154 in the cell membrane (Figure 5B). Where there are patches of quantum dot-stained HA-Srt, we see evidence of CTx colocalization, while for quantum dot labeled CD154, this is not always the case. Of note, cells that do not stain with CTx nonetheless label perfectly well with streptavidin-modified Qdots (and hence correspond to flu-infected cells) in a pattern that is indistinguishable from that seen in the adjacent, CTx positive cells. Although GM1 is a raft component, the organization of the plasma membrane apparently does not require its presence in a CTx-reactive form for the organization of HA-Srt.
We conducted a pulse labeling experiment using sortase to determine the fate of HA-Srt protein labeled with biotin and streptavidin modified Qdots, followed by a second round of site-specific labeling with an oligoglycine based Alexa fluor 488 probe [18] to distinguish this second pool of labeled material from the first round of labeled HA-Srt, at a later stage of maturation and virus budding. Infected cells were first labeled with sortase and a biotinylated probe, followed by staining with streptavidin quantum dots (HA-Qdot). After 30 minutes, this first pulse was followed by a second round of labeling using sortase A and an oligoglycine-Alexa 488 probe (HA-488) at 4°C to inhibit endocytosis of free dye (Figure 6A, Figure S1). When labeling with the Alexa488 probe is initiated directly after quantum dot staining, we observe clear colocalization of the Qdot signal with the Alexa488 signal for most of the patches, and this is reflected as an increase in the overlap coefficient of the Qdot signal with the Alexa488 signal over time (Figure 6A and Figure 6B, Supporting Figure S1). Given the near quantitative labeling we observe for HA at the cell surface, the Alexa488-labeled pool of HA must therefore correspond to HA molecules inserted at sites where HA, labeled with Qdots in the first round of sortagging, has coalesced. These insertion sites appear as discrete dots. When Alexa 488 staining is initiated at later time points, HA-488 is located not only in previously established HA-Qdot patches, but we observe the presence of an increasing number of new Alexa488 spots of HA-Srt outside of the Qdot patches. This results in a decrease in the overlap coefficient of the HA-488 signal with the Qdot signal over time (Figure 6B).
The majority of HA-Qdot remains co-localized with HA-488 over the course of the experiment (Figure 6B), suggesting that new HA-Srt is continuously being exported to the same membrane patches during budding. However, while colocalization with HA488 persists, both the number of HA-Qdot patches as well as their intensity decreases over time when compared to HA-488 (Figure 6A, Figure S1). Although it is a formal possibility that budding may not have been completed during the 2–3 hour interval, these observations may also suggest that patches of HA-Srt on the cell surface serve as sites of multiple budding events. At the 3 hr timepoint, little of the initial HA-Qdot signal remains, paralleling exactly the decrease observed by flow cytometry (Figure 5A). A substantial HA-Qdot signal remains and is found inside cells instead of at the cell surface at later timepoints (Figure 6C). The release of HA-Qdot tagged virus in the confined environment of this tissue culture experiment unavoidably leads to adsorption and internalization of labeled virus by adjacent, uninfected as well as onto already infected cells. Labeling with Alexa-488 probe increases at every time point, consistent with continued output of HA-Srt.
We next studied the behavior of HA-Srt at the cell surface in real time by timelapse imaging (Figure 7A). We labeled HA-Srt with biotin probe and streptavidin-Qdots at 4 hours post infection and acquired images over the following 60 minutes. We see a clear disappearance of HA-Qdot from the infected cells over this time course (Figure 7A). We do not see this loss of Qdot signal in CD154/CD40L-expressing cells labeled in the same fashion (Figure 7B). This loss of Qdot signal is apparent when the total sum of pixel intensities is plotted (Figure 7C). The initial increase seen in CD154/CD40L control cells is likely due to a slight movement of the cells which cannot be restricted in our system and indicates that the decrease in HA-Qdot signal seen must be a minimum estimate of HA-Qdot release. We also observe an increase in clustering of the Qdot patches over time in the CD154/CD40L control cells. This may also account for the increase in signal intensity, as clustering of quantum dots is known to increase the on-time of blinking Qdots [26] Timelapse movies show a similar decrease in Qdot signal for the infected cells relative to the CD154/CD40L-expressing control cells (data not shown). We detect both an apparent decrease in the total number of patches as well as a decrease in the intensity. This is in agreement with the observations made by flow cytometry (Figure 5A) and the timecourse of pulse labeling with sortase (Figure 6A, Figure S1).
For viruses such as vaccinia virus, which tolerates GFP extensions in many of the proteins it encodes, direct visualization of viral replication has provided major insights into the interactions between host cell and pathogen [27]. For viruses such as flu that are refractory to labeling with fluorescent proteins, other methods are urgently needed. We here devised a system to observe the behavior of influenza glycoproteins in cells infected with a fully functional virus. Our approach leverages the sortase labeling technique through the generation of influenza viruses that carry the short sortase recognition sequences in their glycoproteins, resulting in minimally modified viral gene products in the context of an infectious particle. Upon infection of tissue culture cells, the engineered viruses show behavior identical to the parental strain in terms of infectivity, replication kinetics, and viral protein synthesis. Observation of the sortase-labeled glycoproteins thus reflects very well the behavior of their wild-type counterparts. By combining this labeling approach with live-cell imaging, we can monitor the behavior of the influenza glycoproteins in real time. We observe extensive colocalization of surface-disposed flu glycoproteins with lipid rafts, as inferred from staining with CTx. In MDCK cells, flu HA and NA are enriched in lipid rafts based on their transient insolubility in cold TX100 and on colocalization with CTx [28].
Several aspects deserve emphasis. First, the specificity of the reaction and its reliance on an active enzyme limit all labeling to proteins that bear the sortase recognition sequence. We have not detected spurious incorporation of label into proteins not specifically designed to serve as sortase substrates. This applies not only to our earlier results with CD154/CD40L [11], dectin-1 and CD74 [18], but is extended here to flu NA and HA. The requirement for proteolytic conversion of HA0 into HA1 and HA2 suggested the possibility of installing a sortase motif upstream of the trypsin cleavage site without affecting the folding or fusogenic activity of the HA1-HA2 heterodimer, and our results validate this approach. Second, the glycosidase digestions performed on HA and NA unequivocally demonstrate the selectivity of the sortase reaction for fully mature surface-disposed proteins. In this regard, the method compares favorably in terms of ease and specificity with other chemical (iodination, biotinylation) or enzymatic (lactoperoxidase-catalyzed iodination) surface labeling methods, whose products require additional purification steps to enable analysis of the protein(s) of interest. Third, the size of the substituents introduced is minimal compared to that of a fluorescent protein such as eGFP or other enzymatically active modules used for site-specific covalent modification [29]. Given the failure to construct infectious flu when incorporating GFP into any of its structural proteins including NA and HA, the small size of our probes and the choices of label available have enabled, for the first time, the visualization of flu release from the surface of infected cells in real time. Fourth, it is possible to perform sequential labeling experiments and install labels that distinguish between each pool of labeled protein, and so independently monitor each pool of labeled product. This allowed us to generate a starting population of HA-positive cells labeled with quantum dots, from which we generated, at set intervals, a second population of cells with a distinct label installed on HA. Using this approach we demonstrate that HA-Srt, labeled in a first round of sortagging, identifies patches into which HA, tagged in a subsequent round of labeling, is inserted. Although it is possible that budding may not have been completed within this time interval (2–3 hours), an alternative interpretation is the existence of specialized sites that serve as a platform for coalescence of viral glycoproteins (a typical flu virion requires some 400 HA and 30 NA molecules). Whether the identity of such patches corresponds to lipid rafts or specializations of lipid rafts is not clear, but the ability to visualize such sites opens the possibility of identifying host factors that control the construction of these sites as launching pads for new virions. We anticipate that this system will yield a robust method to visualize the kinetics of particle formation and, in combination with perturbations of host cells, will reveal host proteins that contribute to the process of influenza virion biogenesis.
All animal protocols were conducted in accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. All animals were maintained according to the guidelines of the MIT Committee on Animal Care (CAC). These studies were approved by the MIT CAC (protocol #1011-123-14). All infections were performed under avertin anesthesia, and all efforts were made to minimize suffering.
Sortase was produced as described (Popp et al., 2007). Peptide probes were produced as described [18].
Mutant viruses were generated by reverse genetics using plasmids as described [19]. The hemagglutinin and neuraminidase plasmids were modified by standard molecular biology techniques to carry the sortase cleavage site. All viruses, including the wild-type WSN virus used were rescued as described [19]. Nucleotide and protein sequences for modified portions of flu glycoproteins are below.
Protein Sequence (Trypsin cleavage site is in bold)
…CPKYVRSTKLRMVTGLRNIPSIQYRGLFGAIAGFIEGGWTGMIDGWYGYHHQNEQGSGYAA…
Nucleotide Sequence (Trypsin cleavage site is in bold)
…tgcccaaaatatgtcaggagtaccaaattgaggatggttacaggactaagaaacatcccatccattcaatacagaggtctatttggagccattgctggttttattgaggggggatggactggaatgatagatggatggtatggttatcatcatcagaatgaacagggatcaggctatgcagcg…
Protein Sequence (Sortase recognition site is in italics, trypsin cleavage site is in bold)
…CPKYVRSTKLRMVTGLRNIPSIQYLPETGGRGLFGAIAGFIEGGWTGMIDGWYGYHHQNEQGSGYAA…
Nucleotide Sequence (Sortase recognition site is in italics, trypsin cleavage site is in bold)
…tgcccaaaatatgtcaggagtaccaaattgaggatggttacaggactaagaaacatcccatccattcaatacctgcccgagaccggcggcagaggtctatttggagccattgctggttttattgaggggggatggactggaatgatagatggatggtatggttatcatcatcagaatgaacagggatcaggctatgcagcg…
Protein Sequence
…SGSIISFCGVNGDTVDWSWPDGAELPFTIDK-
Nucleotide Sequence
agtgggagcatcatttctttttgtggtgtgaatggtgatactgtagattggtcttggccagacggtgctgagttgccgttcaccattgacaagtag
Protein Sequence (Sortase recognition site is in italics, HA epitope is in bold)
…SGSIISFCGVNGDTVDWSWPDGAELPFTIDKGGGGSLPETGGYPYDVPDYA-
Nucleotide Sequence (Sortase recognition site is in italics, HA epitope is in bold)
agtgggagcatcatttctttttgtggtgtgaatggtgatactgtagattggtcttggccagacggtgctgagctcccgttcaccattgacaagggcgggggcggatcccttcctgaaactggtggatacccatacgatgttccagattacgcttag
CD154/CD40L bearing an LPETG tag [11] was cloned into pLHCX and used to make retrovirus as described [30]. Retrovirus was used to infect MDCK cells as described [30] and cells were selected in 250 µg/ml Hygromycin B.
Viral titer was assessed by plaque assay on MDCK cells as described [28]. For multi-step replication assays, MDBK cells were infected at an MOI of 0.001 and incubated for the indicated times in viral growth medium (VGM, DME with 0.3% BSA) supplemented with 0.5 µg/ml TPCK-treated trypsin (NA-Srt) or 1 µg/ml TPCK treated trypsin (HA-Srt). For NA-Srt multi-step replication, 100 µl of media was plaqued at the indicated time points on MDCK cells grown with 0.5 µg/ml TPCK-trypsin. For HA-Srt, HA-Srt and wild-type WSN virus (n = 3) were used to infected MDCK monolayers at an MOI = 0.001 and viral supernatant was analyzed via standard hemagglutination assays.
Viral particles from tissue culture supernatant were concentrated by pelleting through a 20% sucrose cushion (Sigma) at 25000 rpm in an SW-28 rotor for 120 minutes. Where indicated, virus was further purified through a continuous 15%–60% sucrose gradient, centrifuged at 30,200 rpm in an SW-40.1 rotor for 3 hours.
For biotin and TAMRA labeling, virions were pelleted from tissue culture supernatant without a sucrose cushion. The pellet was resuspended in 1× sortase buffer and labeled with 150 µM sortase A/5 mM probe for 1 hour at 37°C. For Alexa647 probe labeling, sucrose gradient purified virions were mixed with 200 µM sortase A and 500 µM probe at 37°C for 2 hours.
MDCK cells were plated in a 24 well dish at 70% confluency the night before the experiment. Cells were infected at an MOI of 1 and 0,1,2,3,4,5,6 or 7 hours post infection, cells were incubated with 100 mM sortase A and 100 mM G5K-biotin probe [11] in VGM for 30 minutes at 37°C. Cells were washed extensively in PBS, collected, and lysed in 1% SDS with protease inhibitor cocktail (Roche). A BCA assay was performed (Pierce) and 20 µg of lysate was loaded for western blotting.
Mice (n = 4 in each group) were inoculated intranasally with 40000 pfu of the indicated virus and body weight was monitored at the indicated intervals. Balb/C mice and B6129SF2/J mice were purchased from the Jackson Laboratory (stock# 000651 and 101045 resp). Mice were anesthetized with Avertin and infected intranasally with 40.000 PFU WT, HA-Srt or NA-Srt virus. Infection was followed by daily monitoring of weight loss and animals were euthanized with C02 when weight loss exceeded 20% of initial body weight.
MDCK cells were infected at an MOI of 0.4 overnight and labeled for 1 hour at 37°C with 100 mM sortase and 500 mM biotin probe. Cells were then lysed in glycoprotein denaturing buffer (New England Biolabs) and total protein in lysates were quantiatated by BCA assay (Pierce). Five micrograms of cell lysate was digested with either PNGase F or EndoH according to manufacturer's directions (New England Biolabs), resolved by 12.5% SDS-PAGE, transferred to nitrocellulose, and used for western blotting with the indicated antibodies.
MDCK cells were grown in 6-well tissue culture dishes and infected with HA-Srt at either an MOI 0.05 for 14 hrs or MOI 0.5 for 4.5 hrs as indicated in figure legends. Cells were starved with methionine- and cysteine-free DMEM for 45 minutes at 37°C followed by a 20 minute pulse labeling with [S35]Cysteine/Methionine (perkin elmer) at 0.77 mCi/ml in methionine- and cysteine-free DMEM. Chase was initiated by addition of VGM supplemented with 1 mM methionine, 0.2 mM cysteine and 1 µg/ml TPCK treated trypsin. At indicated timepoints during chase, cell surface HA molecules were labeled with 0.25 mM G3K-Biotin and 0.1 mM SrtAureus for 30 minutes at 37°.
MDCK cells or CD154 control cells were cultured o/n in 24-well tissueculture dishes. Cells were infected with HA-Srt or WT virus at an MOI of 1. At 4 hrs post-infection, cell surface molecules were labeled by addition of 0.1 mM Srtaureus and 0.25 mM G3K-Biotin for 30 minutes at 37°C. Cells were washed in PBS and incubated for 5 minutes with either 20 nM qdot655 (MDCK cells and CD154 cells infected with WT-WSN) or 10 nM qdot655 (CD154 cells). Cells were incubated at 37°C and at indicated timepoints collected via addition of trypsin and kept on ice. Cells were analyzed immediately on a FACSCalibur flow cytometer (BD Biosciences) and FlowJo software.
MDCK cells or CD154/CD40L expressing MDCK cells were cultured in 35 mm glass bottom dishes (MatTek Corporation) and infected with HA-Srt or WT virus at an MOI of 0.35–0.5 for 4 hours. Cell surface molecules bearing the LPETG motif were labeled by incubation with 0.1 mM SrtAaureus and 0.25 mM G3K-biotin for 30 minutes at 37°C. Cells were washed and incubated with 20 nM (MDCK) or 10 nM (CD154) qdot655 (Invitrogen) for 5 minutes. After extensive washing, cells were incubated with VGM containing 1 µg/ml trypsin-TPCK. For surface labeling with (G)3K-Alexafluor 488 probe, cells were incubated with 0.1 mM SrtAaureus and 20 nM probe for 1 hr at 4°C to inhibit non specific endocytosis of free dye. For staining of lipid rafts, cells were stained with 20 µg/ml Alexa fluor 594 conjugated choleratoxinB dye (invitrogen) for 5 minutes at room temperature immediately following Qdot labeling and imaging performed directly after.
Images were acquired using an Andor Revolution spinning disk system with Yokogawa CSU-X1 spinning disk head, Andor iXon+ EM-CCD camera, 488 nm diode laser for excitation, emission discrimination with an emission filterwheel, Piezo Z100 z-stage on a Nikon Ti-E motorized microscope stand with a 100× 1.49NA Plan Apochromatic objective all controlled with the Andor iQ2 software (version 2.0). Temperature, CO2 and humidity was controlled with a LiveCell stage-top incubation system (Pathology Devices).
Analysis was performed using Imaris, Volocity and ImageJ software. For colocalization analysis, background correction was applied using ImageJ background correction with a rolling ball radius of 20 pixels. Images were further analyzed with Volocity colocalization analysis software. Background threshold was manually set using a background ROI to correct side effects of possibly remaining background pixels.
Imaris software was used for analysis of timecourse Z-stack series. Brightness of images was adjusted using the linear stretch algorithm with the maximum set to 55. Background was corrected for using a 17 µm filter width. Imaris software was used to create 3D images as well as the quantification of pixel intensities.
|
10.1371/journal.pntd.0005618 | Development of a movement-based in vitro screening assay for the identification of new anti-cestodal compounds | Intestinal cestodes are infecting millions of people and livestock worldwide, but treatment is mainly based on one drug: praziquantel. The identification of new anti-cestodal compounds is hampered by the lack of suitable screening assays. It is difficult, or even impossible, to evaluate drugs against adult cestodes in vitro due to the fact that these parasites cannot be cultured in microwell plates, and adult and larval stages in most cases represent different organisms in terms of size, morphology, and metabolic requirements. We here present an in vitro-drug screening assay based on Echinococcus multilocularis protoscoleces, which represent precursors of the scolex (hence the anterior part) of the adult tapeworm. This movement-based assay can serve as a model for an adult cestode screen. Protoscoleces are produced in large numbers in Mongolian gerbils and mice, their movement is measured and quantified by image analysis, and active compounds are directly assessed in terms of morphological effects. The use of the 384-well format minimizes the amount of parasites and compounds needed and allows rapid screening of a large number of chemicals. Standard drugs showed the expected dose-dependent effect on movement and morphology of the protoscoleces. Interestingly, praziquantel inhibited movement only partially within 12 h of treatment (at concentrations as high as 100 ppm) and did thus not act parasiticidal, which was also confirmed by trypan blue staining. Enantiomers of praziquantel showed a clear difference in their minimal inhibitory concentration in the motility assay and (R)-(-)-praziquantel was 185 times more active than (S)-(-)-praziquantel. One compound named MMV665807, which was obtained from the open access MMV (Medicines for Malaria Venture) Malaria box, strongly impaired motility and viability of protoscoleces. Corresponding morphological alterations were visualized by scanning electron microscopy, and demonstrated that this compound exhibits a mode of action clearly distinct from praziquantel. Thus, MMV665807 represents an interesting lead for further evaluation.
| Tapeworms (cestodes) are a medically important group of helminths that infect humans and animals all around the globe. The clinical signs caused by intestinal infection with adult cestodes are mostly mild, in contrast to the more severe disease symptoms inflicted by infection with the tissue-dwelling larval stages of the same species. Praziquantel is the main drug in use against intestinal cestode infections. Development of resistance and treatment failures have been reported in trematodes, and are expected to become a problem in the future also in the case of cestode infections. Therefore, new treatment options against intestinal helminths are needed. To date, there is no in vitro-based whole-organism screening assay available that allows screening of candidate drugs with potential activity against adult cestodes. We established and characterized of a screening assay in 384-well format, which serves as a model for adult stage parasites by using Echinococcus multilocularis protoscoleces and their loss of motility as a read-out. This novel assay showed that drugs with known activity against adult cestodes inhibited motility of protoscoleces. The movement-based assay identified MMV665807 as a novel compound with profound activity against protoscoleces, and potentially also adult cestodes. Light- and electron microscopical assessments of protoscoleces treated with praziquantel and MMV665807 point towards different modes of action of the two drugs.
| Helminths are separated into the two major phyla of nematodes (roundworms) and platyhelminths (flatworms), including trematodes and cestodes, and they are important causes of disease in humans as well as animals. An estimated one billion people are infected with at least one helminth in developing countries of Africa, Asia and America [1]. Infection of livestock by helminths, small ruminants in particular, has an enormous economic impact on productivity in farming [2]. Despite the large number of infected individuals and enormous economic losses due to helminth infections in animals, there are still not many drugs registered for their treatment [1]. Present efforts ongoing to discover new anthelminthic drugs are focused on gastrointestinal nematodes, schistosomes and filariae [3], as they comprise the highest prevalences. However, most of the adult stages of trematodes, and all cestodes, are not being considered in the current drug screening efforts. Intestinal cestodes might be considered as parasites of lower relevance as they usually cause few clinical signs, but they are of high relevance as source of infection of diseases caused by the larval stages of these parasites [4,5].
For the treatment of nematodes, a variety of drugs are in use, and new ones have been introduced to the market recently. However, spread of resistance is a major problem in the veterinary sector [6]. For treatment of cestode and trematode infections, praziquantel (PZQ) is the drug of choice against most species [7]. For cestode-treatment, the alternatives available are epsiprantel that is exclusively applied in animals, and niclosamide, whose marketing status is currently discontinued [4,8]. PZQ is generally very well tolerated, even though it tastes bitter. It induces only mild adverse reactions, but rare events of allergy and hypersensitivity reaction have been described [7]. However, there is increasing evidence on resistance of schistosomes against PZQ [9] and treatment failures of PZQ against Taenia saginata are described as well [10]. A major problem is that PZQ is the only drug in use against many platyhelminths and mass drug administrations all over the globe might select for resistant platyhelminth strains in the future. In helminths, as compared to for example bacteria, resistance development takes more time as their generation time is much longer. Nevertheless, drug resistance is already a major problem for many diseases caused by nematodes and trematodes, and it will be only a question of time until resistance to PZQ spreads also to cestodes [11,12]. Therefore, it is of crucial importance to search for new anthelminthic drugs, including compounds against platyhelminths.
Whole-organism screening is still the most widely accepted method for anti-parasitic drug discovery, despite the fact that target-based screening approaches have been widely introduced [3]. However, this approach is challenging as it relies on a profound knowledge of parasite life-stages and their biology, the availability of in vitro culture techniques, and reliable methods for compound efficacy assessment [3]. Caenorhabditis elegans is a frequently used model worm for whole-organism in vitro screenings, but over the last decades also other models have been successfully established. Classically, the viability and morphology of whole parasites upon treatment is assessed by light microscopy, which harbors drawbacks such as time-consuming and subjective evaluation procedures, and automatization is not possible. However, no specialized equipment is required [13]. Other assays include dyes that indicate viability or loss of viability such as trypan blue and eosin or even fluorescent stains, but nevertheless, these are low-throughput methods and they represent only indirect indicators of viability [14,15]. For objective larger-scale screenings, new methods were implemented over the last decades, such as MTT (3-(4,5-Dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide) assay, alamar Blue assay or fluorescent labeling that allow an automated readout [3,14].
The observation that anthelminthics reduce larval motility in nematodes led to the development of a motility-based assay for assessing the effects of certain compounds [16]. Non-image-based methods to measure nematode motility include measurements of the fluctuations in electrical currents by xCELLigence [17] or isothermal microcalorimetry [18]. In addition, image-based methods have been introduced, such as the Parallel Worm Tracker [19], the WormAssay [20], the WormScan [21], the Worminator [22] or the whole-organism screening by Preston et al [23].Adult cestodes have so far been excluded from in vitro screening, since for most of these parasites in vitro culture methods have not been established, or they are too large to be cultured in tissue culture devices suitable for screening assays.
Adult E. multilocularis tapeworms live in the small intestine of final hosts, such as foxes and dogs, and they release infectious eggs in the faeces of these hosts. Upon ingestion of eggs, intermediate hosts, such as rodents and small mammals, get infected with the parasite and a multivesicular metacestode will develop mainly in their livers. After 2–4 months, brood capsules with protoscoleces will form within these metacestodes. Once a final host feeds upon an intermediate hosts, protoscoleces get ingested as well, and they will develop into adult tapeworms in the intestine of these hosts [24]. The fox tapeworm Echinococcus multilocularis has become an important model for the study of cestodes, since the genome has been sequenced and corresponding data is publically available, including also the close relative E. granulosus [25,26], advanced molecular tools have been developed for the study of the disease-causing metacestode stage [27,28], and metacestode in vitro drug screening assays have been implemented [29–31]. However, it is known that drugs with activity against larval stages are not necessarily active against adult stages and vice versa [3,14]. Albendazole is the drug of choice against alveolar echinococcosis caused by E. multilocularis metacestodes, but the drug is ineffective against adults. On the other hand, PZQ is the most widely used drug against adult cestodes, but it is not active against metacestodes. Adult E. multilocularis worms have not been studied extensively, due to high risk of infection for the experimenter and lack of suitable laboratory models. There are a number of studies on protoscolicidal substances for the treatment of E. granulosus infections. However, none of these have considered protoscoleces as a potential model for adult tapeworms. Protoscoleces are easily generated and purified in large numbers, they are relatively small (150–350 μM in length), move actively and they represent precursors of adult tapeworms. Thus, we describe here an in vitro drug screening assay that is easy to perform, inexpensive, and which is based on the semi-automated, quantitative assessment of E. multilocularis protoscolex motility. In addition, we also compare the method of motility assessment to the classical trypan blue staining method.
If not stated otherwise, all chemicals were purchased from Sigma (St. Louis, MO, USA). Dulbecco’s modified Eagle medium (DMEM) and fetal calf serum (FCS) were from Biochrom (Berlin, Germany). Plastic ware was from Sarstedt (Nümbrecht, Germany). All drugs tested in the present study were obtained from Elanco (St. Aubin, Switzerland) and they were all delivered as stocks of 10 g/L in DMSO. The drugs applied for experiments of Supplementary S3 Fig were purchased from Sigma, and prepared as stocks of 10 g/L as well.
E. multilocularis protoscoleces were extracted under sterile conditions from E. multilocularis metacestodes kindly provided from strain maintenance surplus by Dr. R. Fiechter (permission number ZH139/2015, Institute of Parasitology, Zürich, Switzerland). The parasite originated from naturally infected monkeys from the German primate center, and they had been under in vivo passage in Mongolian gerbils for up to five years. The protocol described by Brehm et al was applied to purify protoscoleces from metacestodes with minor changes [32]. In short, the metacestode material was pressed through a tea sieve, washed out with PBS and further broken mechanically by vigorous shaking in a QIAGEN tissue lyzer for 10 minutes (8 shakes per second). The resulting suspension was passed through a sieve with 250 μM mesh size. The flow through was then passed through a second sieve with 50 μM mesh size and thoroughly washed with PBS. The protoscoleces that were collected in the 50 μM sieve were transferred into a 50 mL tube and washed repeatedly with PBS until all remaining vesicle tissue was removed and the protoscoleces sedimented fast. For further purification, the protoscoleces were given into a petri dish (14 cm diameter) containing 70 mL DMEM with 10% FCS. The dish was moved in circles so the protoscoleces could be collected in the centre of the plate and they were again washed in PBS before being immediately subjected to activation (see below).
The standard procedures for evagination of protoscoleces are based on incubation in pepsin (0.5 mg/mL), pH 2.0 for 3 hours at 37°C [33–35], or pepsin (0.05%), pH 2.0 for 30 minutes and subsequent incubation in 0.2% Na-taurocholate for 3 hours at 37°C [36,37]. In order to simplify this method for a subsequent motility-based screening, we incubated freshly extracted protoscoleces also for 3 hours at various concentrations of DMSO (0, 1, 2.5, 5, 10, and 20%) in DMEM, including 10% FCS. All incubations were done in triplicates. Subsequently, the protoscoleces were washed in PBS and were allowed to recover overnight in DMEM with 10% FCS at 37°C and 5% CO2. Thereafter, the numbers of invaginated and evaginated protoscoleces were counted manually under the light microscope within one field of vision (40 x magnification, > 170 protoscoleces per view) and the percentage of evaginated protoscoleces was determined. Averages and standard deviations were calculated in Microsoft Office Excel 2010. The experiment was repeated three times.
Based on the results obtained above, a standardized activation protocol was established: a maximum of 25’000 protoscoleces per well were seeded in a 6-well pate in DMEM containing 10% FCS and 10% DMSO, at 37°C, and 5% CO2 for 3 hours. Thereafter, the protoscoleces were washed in PBS twice at room temperature and incubated in DMEM with 10% FCS and antibiotics (100 U/mL penicillin, 100 μg/mL streptomycin) at 37°C, and 5% CO2, overnight. For all subsequent manipulations, pipettes and tubes were pre-rinsed with FCS in order to avoid sticking of protoscoleces to the tubing walls.
Photomicrographs of protoscoleces were taken in a live cell imaging system (Nikon TE2000E microscope connected to a Hamatsu ORCA ER camera) with the software NIS Elements Version 4.40 and the additional module JOBS (JOBS program given in S1 File). For motility measurement, two separate images were taken of each well at a 10 seconds interval and at 40 times magnification. Motility was assessed at various time points (1, 6, 12, 18, and 24 h) after addition of compounds. The motility index for each well, resulting from differences in pixels within the 10 seconds interval, was calculated with a pixel grey value threshold of 230 in ImageJ version 1.49 (the respective Macro including details of calculation is given in S2 File). Averages of movement indices (excluding the minimal and maximal values of each of the six replicas) and respective standard deviations were calculated in Microsoft Office Excel 2010.
In order to check the correlation of number and movement of protoscoleces in the 384-well format, various numbers of protoscoleces (1 to 58) were tested. Protoscoleces were seeded in a total of 20 μL of DMEM without phenol red, including 10% FCS and 1% DMSO, in a 384-well plate (order nb. 788095–128, Greiner bio-one, via Huberlab, Aesch, Switzerland) and sealed by a pressure sensitive seal (order nb. 7676–070, Greiner bio-one, via Huberlab). The total motility corresponding to each number of protoscoleces was assessed as described above, but without normalization and the correlation was determined in R (version 3.3.0).
To determine the optimal temperature for the motility assay, the same general protocol was used as in the number to motility ratio experiment (see above): 20–30 protoscoleces per well were seeded in a total of 20 μL of DMEM without phenol red, including 10% FCS and 1% DMSO. Automated seeding was performed by application of a peristaltic pump (multidrop combi, Thermo Fisher Scientific, Reinach, Switzerland) equipped with a standard size cassette (tubing inner diameter 1.3 mm, tip inner diameter 0.5 mm) at low speed. After incubation for 1 hour at various temperatures (25, 30, 37, and 41°C), corresponding motilities of 10 replicas were assessed as described above. The average value of absolute movement and corresponding standard deviations, and Wilcoxon rank-sum test for determination of p-values, were calculated in R. The experiment was repeated three times.
Several experiments were performed to measure the effects of different compounds (each in six replicas) on the motility of protoscoleces. In an initial experiment, freshly activated protoscoleces (activated by incubation in 0 (DMEM control) to 20% DMSO, pepsin, or pepsin and Na-taurocholate, see above) were tested for their motility as described in this section. In an additional DMSO control experiment, various concentrations of DMSO (0, 0.1, 0.3, 1, 3, and 10%) were incubated with a total number of 20–30 protoscoleces per well in DMEM (without phenol red, containing 10% FCS) for 12 h. Therefore, respective DMSO dilutions were added first to each well and protoscoleces were distributed using the peristaltic pump with the standard size cassette at low speed. The plates were then sealed by pressure sensitive seal and incubated in the incubation chamber of the Nikon live imaging system at 37°C and the motility assessed as described above.
For incubation of activated protoscoleces with various drugs, the compounds were pre-distributed in 5 μL aliquots into 384-well plates as 4x stocks in DMEM (without phenol red, containing 10% FCS and 4% DMSO). To measure the effects of different enantiomers of praziquantel (PZQ), PZQ racemate, (R)-(-)-PZQ and (S)-(-)-PZQ were added to final concentrations of 100 to 0.0006 ppm in a 1:3 dilution series. Further standard compounds (see Fig 1) with known activity (niclosamide [4], nitazoxanide [38]) and known inactivity (albendazole and monepantel), as well as the compound MMV665807 from the open-access malaria box [31], were prepared to final concentrations of 100 to 0.4 ppm in a 1:3 dilution series. Wells with 4% DMSO in DMEM (without phenol red, 10% FCS) served as negative controls. Subsequently 20–30 protoscoleces per well were added in a total of 15 μL DMEM (without phenol red, containing 10% FCS), in order to reach a final DMSO concentration of 1%. They were distributed using a peristaltic pump. Motility was assessed as described above and expressed as percentage of the DMSO control in order to normalize for parasite batch-variations. The experiment with standard compounds was repeated 3 times. The minimal inhibitory concentration (MIC) was determined by calculating the critical concentration where the slope of the dose-response curve changed significantly according to students T-test (p-value < 0.05) in Microsoft Excel 2010.
To evaluate different protoscolex-activation protocols with regards to sensitivity to standard drugs, we compared also the movement of protoscoleces activated with 10% DMSO, pepsin only, or pepsin and Na-taurocholate, and incubated them with the standard compounds as mentioned above.
In order to compare the effects of the standard drugs as assessed by motility assay also to protoscolex viability, trypan blue staining of protoscoleces was performed. The protoscoleces were extracted and activated as described above. After overnight recovery in DMEM including 10% FCS, the protoscoleces were distributed into 96-well plates of approximately 100 protoscoleces in DMEM without phenol red containing 10% FCS. Subsequently, standard drugs (PZQ, niclosamide, nitazoxanide, albendazole, monepantel and DMSO) and the compound MMV665807 were added to final concentrations of 100 to 0.4 ppm in a 1:3 dilution series with a final DMSO concentration of 1% and in triplicates. The protoscoleces were then incubated with drugs for 18 h at 37°C and 5% CO2. For viability staining, trypan blue was added to a final concentration of 0.5% trypan blue during 5 minutes at room temperature. Thereafter, the protoscoleces were washed once in PBS and pictures were taken of each well at 40 times magnification. At last, the trypan blue-stained protoscoleces and the non-stained protoscoleces were counted manually for each picture and the resulting average percentage of viabilities and standard deviations were calculated in Microsoft Excel 2010. This experiment was repeated three times. In order to compare the relative motility of protoscoleces treated with the above mentioned standard drugs to the relative viability of protoscoleces as assessed by trypan blue staining, linear exponential regression was applied in R (version 3.3.2, function lm) for each drug.
To compare the morphological effects induced by different activation protocols, or by the drugs PZQ and MMV665807 in more detail, SEM was performed as described by Hemphill and Croft [39]. To compare the morphology of differently activated protoscoleces (DMEM only, 10% DMSO, pepsin, or pepsin and Na-taurocholate), protocols were applied as described above, and protoscoleces were allowed to recover overnight in DMEM including 10% FCS at 37°C, 5% CO2. Thereafter, 500 protoscoleces per condition were washed in 0.1 M cacodylate buffer (pH 7.3) and fixed for 2 hours at room temperature in 2 mL glutaraldehyde (2% in 0.1 M cacodylate buffer), before being further treated as described below. To analyze the drug-induced effects on the morphology of protoscoleces, 500 DMSO-activated protoscoleces per well were incubated in a 96 round well plate with 1% DMSO, PZQ (100, 10, 1 and 0.1 ppm, in 1% DMSO) or MMV665807 (100, 10, 1 and 0.1 ppm, in 1% DMSO) in DMEM without phenol red and 10% FCS for 18 hours at 37°C and 5% CO2. Thereafter the protoscoleces were washed in 0.1 M cacodylate buffer (pH 7.3) and fixed for 2 hours at room temperature in 2 mL glutaraldehyde (2% in 0.1 M cacodylate buffer). Following three washes in 1 mL of cacodylate buffer (0.1 M, pH 7.3), samples were postfixed in 1 mL osmium tetroxide (2% in 0.1 M cacodylate buffer) during 2 hours, washed twice in water, and specimens were dehydrated in increasing concentrations of ethanol (30, 50, 70, 90, and 3x 100%), and subsequently 1 mL hexamethyldisilazane (HMDS) was added and incubated for 2 minutes. After removal of the HMDS, protoscoleces were resuspended in another 50 μL of HMDS, and were spotted onto a glass coverslip. After evaporation at room temperature, the fixed protoscoleces samples were sputter coated with gold and inspected in a Hitachi scanning electron microscope S-3000 N operating at 25 kV.
All figures were prepared in R (version 3.3.2) and formatting as well as compiling was done in Adobe Illustrator (2015.1.0). The compound structures in Fig 1 were depicted by MolView V2.4.
Different concentrations of DMSO were applied in addition to the standard method (using pepsin, or pepsin and Na-taurocholate), in order to improve the protocol for evagination and activation of motility of protoscoleces. The isolation procedure itself already led to evagination of 69.1 ± 6.1% of protoscoleces (Fig 2A). Incubation in 20% DMSO for 3 hours yielded the highest (92.3 ± 1.8%) evagination efficiency, while the presence of 1 to 5% DMSO seemed to inhibit rather than promote evagination (Fig 2A). Motility was most pronounced upon incubation in 10% DMSO (2.1 ± 0.1 x 104 changed pixels), and superior to the standard pepsin (0.7 ± 0.2 x 104 changed pixels) or pepsin/Na-taurocholate (1.3 ± 0.1 x 104 changed pixels) methods Fig 2B. Upon activation in 20% DMSO, motility was drastically reduced (0.8 ± 0.3 x 104 changed pixels, Fig 2B). In Fig 2C, SEM micrographs of differently activated protoscoleces are depicted, and they show that the morphology was not affected by the here presented activation methods. A good discrimination of moving and non-moving parasites is crucial for the motility-based drug-screening applied here. Therefore, we concluded to induce evagination and activation by incubation in 10% DMSO for 3 h in all subsequent experiments.
As all tested drug formulations were dissolved in DMSO, we investigated whether the presence of DMSO would affect the motility of protoscoleces. As shown in Fig 3A, protocoleces incubated in the presence of up to 3% DMSO did not exhibit significantly reduced motility, while protoscoleces incubated in 10% DMSO or higher for 12 h showed largely impaired movement. For practicability of subsequent assays, a final concentration of 1% DMSO was defined for all subsequent drug-testing experiments.
Fig 3B shows the correlation between protoscolex number and movement (r = 0.86, R2 = 0.75). For further assays, a representative number of 20–30 protoscoleces per well was chosen, as with this number wells are not overfilled with parasites and individual parasites can be discriminated.
In order to determine the optimal temperature for the assay, comparative runs with protoscoleces incubated for 1 hour in 1% DMSO at different temperatures (25, 30, 37, and 41°C) were performed. As shown in Fig 3C, incubation at a physiological temperature of 37°C yielded the highest motility.
For a first validation of the test, the standard drug in use (PZQ) as well as its R- and S-enantiomers were tested in the protoscolex motility assay. As shown in Fig 4A and S1 Fig, neither the racemic mixture of PZQ, nor its enantiomers (R)-(-)-PZQ or (S)-(-)-PZQ led to complete inhibition of motility at concentrations up to 100 ppm and up to 24 h of incubation time. After 1 h, drug-induced effects did not yet reach their maximum, but as shown in S1 Fig, after 12 h the maximum effects were largely reached. Therefore, in all subsequent experiments, 12 h drug incubations were applied. Based on the drug-response curves of 12 h of drug-incubation, the MICs were determined: (R)-(-)-PZQ and the racemic mixture of PZQ had a MIC of 0.02 ppm (0.06 μM), whereas (S)-(-)-PZQ had a MIC of 3.7 ppm (11.8 μM) (Fig 1). Corresponding micrographs showing morphological changes induced by each drug at 0.02 ppm after 12 h of drug incubation are depicted in Fig 4B. Movies for comparison of effects of PZQ as compared to DMSO are shown in S1 Movie.
Additional reference compounds with known activity against adult cestodes or protoscoleces (niclosamide and nitazoxanide) and without activity against adult cestodes or protoscoleces (albendazole and monepantel) were selected, in order to further validate the protoscolex motility assay. Also for the use of these standard compounds, maximal drug-effects were reached after 12 h of incubation on (see S2 Fig). Therefore, results corresponding to the 12 h time points are provided in Fig 5. One exception is nitazoxanide that, interestingly, lost its activity slightly over time. Niclosamide and nitazoxanide reduced the motility of protoscoleces in a dose- and time-dependent manner (Fig 5A–5E, S2 Fig). Morphological effects induced at a drug concentration of 33.3 ppm are shown in Fig 5A–5E. Corresponding MICs are given in Fig 1. Both albendazole and monepantel did not inhibit the motility of parasites at any of the concentrations tested, except some slight reduction in motility at 100 ppm. We also compared the above described standard drugs against protoscoleces that were activated by 10% DMSO, pepsin, or pepsin and Na-taurocholate. As shown in S3 Fig, protoscoleces followed the same drug-responses independent of which activation method was used. However, differences between active and inactive drugs were highest when protoscoleces were activated with DMSO, as visible in the absolute movement data (S3 Fig).
In addition, a compound previously described to exhibit parasiticidal activity against the metacestode stage of E. multilocularis, MMV665807 [31], was assessed for its motility-reducing activity on protoscoleces. As shown in Fig 5F (and S2 Fig), the drug induced a strong inhibition of motility with a MIC of 3.7 ppm (11.7 μM) after 12 h of drug incubation. Corresponding morphological effects (at 33.3 ppm) are visualized in Fig 5F and a representative movie is provided in S1 Movie.
For comparative reasons, all drugs described above were also assessed by trypan blue staining, which is the standard method of protoscolex viability testing (Fig 6A). The viability of protoscoleces was dose-dependently lowered upon incubation with niclosamide, MMV665807 and nitazoxanide as shown in Fig 6A. The drugs albendazole and monepantel had only slight effects on the viability of protoscoleces. The drug currently used for therapy, PZQ, as well as its enantiomers, showed only minor effects on the viability of protoscoleces at all tested concentrations. The correlation between motility and viability followed an exponential correlation curve for all active drugs except PZQ (Fig 6B and S1 Table for functions and R2 values). Interestingly, the motility of protoscoleces had already been completely inhibited, when viability was still at 50%, which highlights that a reduced motility does not necessarily indicate a reduction or loss of viability. PZQ and its enantiomers did not follow an exponential correlation curve, as they never killed the protoscoleces (Fig 6B).
The standard drug PZQ and the newly identified MMV665807 were also compared regarding their effects on the morphology of protoscoleces. Control protoscoleces that were incubated in 1% DMSO showed the characteristic structures of body and head with rostellum and 4 muscular suckers. The protoscolex surface was comprised of microtriches (Fig 7A). All PZQ-treated protoscoleces showed the same effects, regardless of the tested concentration (100 to 0.1 ppm): they were contracted, especially in the neck region, and started forming blebs all over the body. The microtriches, suckers and the rostellum were not visually affected by PZQ (Fig 7B). Treatment with a concentration of 100 and 10 ppm of MMV665807 (Fig 7C) also led to contraction of the protoscolex body, but to a lower extent compared to PZQ. In addition, the microtriches and the tissue mesh covering the rostellum were both lost and therefore the hooks were not present anymore. The suckers were not visibly affected by MMV665807. At 1 ppm, MMV665807 showed still some of these effects, but at 0.1 ppm, all protoscoleces were visually indistinguishable from the control sample. These morphological findings further confirmed the loss of activity of MMV665807 at concentrations lower than 3.7 ppm (Fig 5).
The present paper describes for the first time an in vitro-based screening test that will allow researchers to evaluate compounds for activity against adult cestodes. In contrast to most other species, E. multilocularis multiplies asexually in its intermediate hosts. Therefore, the generation of parasite material for this test is easy, as metacestode-containing protoscoleces are usually available in excess in laboratories maintaining E. multilocularis in gerbils. The purification of protoscoleces from metacestodes is simple and straight forward and allows high numbers of protoscoleces to be purified from a batch of metacestode material originating from one euthanized animal.
In this assay, motility is used as a phenotypic readout of worm activity, which is often, but not always, coincidental with viability. A reduction of motility to zero could theoretically be seen in a worm that is still alive. However, a viable worm that cannot move, or is severely limited in movement, will be expelled from the gastrointestinal tract [40–42]. Therefore, we consider motility as a suitable readout for the screening of new drugs against adult cestodes. We induced evagination and motility by incubation in DMSO, and did not follow the standard pepsin or pepsin/bile acid methods. DMSO-activation resulted in higher resolution of different motility rates, did not alter the drug-response, did not induce any visible damage to the parasite, and, from the technical point of view, DMSO activation is easy and highly reproducible, since no pH adjustments are necessary. In addition, we also compared the quantified readouts of the motility assay to the subjective and laborious trypan blue viability staining, and we found an exponential correlation between the methods. The lag phase of this correlation curve highlights that inhibition of motility does not necessarily lead to loss of viability.
Described targets of PZQ are voltage-gated calcium-channels and adenosine receptors, both leading to an influx of calcium ions that induces paralysis in the worm [7]. Thus, reduction of protoscolex movement by PZQ-treatment was expected. Our tests revealed that drugs with known anthelmintic activity (PZQ, niclosamide and nitazoxanide) were active and reduced motility of protoscoleces, and those known to be ineffective against adult Echinococcus (albendazole and monepantel) were not. For PZQ we determined a MIC of 0.02 ppm, which is in line with previous observations on Echinococcus protoscoleces [43,44] as well as a variety on various adult cestodes [45,46]. Niclosamide was described to be active against Hymenolepis nana adults at 0.1 ppm within 30 minutes [46], and we observed a slightly higher MIC of 3.7 ppm against E. multilocularis protoscoleces. Nitazoxanide showed a MIC of 33.3 ppm after 12 h against E. multilocularis protoscoleces, which is comparable to results found earlier against E. granulosus protoscoleces (5 ppm after 3 days [47]), as well as the general activity of nitazoxanide against a variety of nematodes with MICs ranging between 10 to 100 ppm [48,49].
Treatment with PZQ led only to a partial paralysis of E. multilocularis protoscoleces. This observation has, to the best of our knowledge, been unknown so far. Viability assessment by trypan blue further confirmed this finding, as PZQ reduced the viability only slightly. Interestingly, after one hour of drug-incubation, PZQ rather increased parasite motility, which could be a first sign of stress response towards the drug. Further, the motility assay allowed also to discriminate between the activity of PZQ enantiomers, and (R)-(-)-PZQ was confirmed as the more active enantiomer, as previously reported for Schistosoma [50]. However, it cannot be concluded whether the (S)-(-)-PZQ has some slight intrinsic activity, or whether this activity simply resulted from impurities in the enantiomer preparation (purity stated by the manufacturer ≥ 95%). In addition, the electron microscopical assessments revealed clear morphological alterations in protoscoleces treated with PZQ. These changes were largely in line with previously described effects of the drug when applied on adult Echinococcus worms, as also in protoscoleces immediate contraction, neck shortening and bleb formation were observed [7,51]. In contrast to descriptions for adult worms, hooks were only partially lost and the suckers did not expand into a convex shape in the present study [51].
One interesting new compound identified by this novel assay: MMV665807, a derivative of niclosamide that is provided within the MMV malaria box [52]. MMV665807 showed a MIC in the same range as niclosamide, and it reduced the viability of protoscoleces to the extent that is comparable to niclosamide. Future in vivo confirmatory studies, such as in the Hymenolepis mouse model, will show, whether the compound exhibits also good in vivo anti-cestode activity. Electron microscopical assessment of protoscoleces treated with MMV665807 showed a clear difference to the changes observed by PZQ, which implies that another mode of action is involved. The mode of action of MMV665807 against E. multilocularis is currently under investigation.
Motility assessment by comparison of change in pixels of two single pictures over a 10 seconds interval was chosen, as such a simplified setup reduces the computational power and analytical complexity needed for data analysis. Previous publications on other motility-based tests of helminths used to employ video-based analyses (e.g. [22]), but due to their analytical complexity they are not easily transferable to higher throughput assays [13]. An approach that could be of interest for increased throughput is the cheaper model described by Marcellino et al [20], where a whole test plate is measured at once. However, in this assay only 24-well plates were applied, as the filarial worms tested with this approach were too large to allow for smaller culture devices. Another approach, based on a real-time monitoring device from Roche, was used for Haemonchus contortus, Strongyloides ratti, hookworms and blood flukes and it allows entire plate assessments as well [17]. However, this device is expensive and so far also restricted to the 96-well format. The more simple Wormscan is relatively cheap, but restricted to the 12-well format [21]. Microfluidic-based platforms for screening of anthelminthics and resistance allow the in-depth and real-time study of the response of worms to drugs, and are therefore of high interest for the study of selected compounds [53]. However, none of the mentioned assays was shown to be applicable for the screening of adult cestodes. Another study by Camicia et al. (2013) showed that E. granulosus protoscolex motility can be monitored in the worm tracker method developed for C. elegans by Simonetta et al. [54], and they also showed that the activity of the neurotransmitter inhibitor citalopram could be detected by this system [55]. Any further application of this system on Echinococcus protoscoleces has, to the best of our knowledge, not been published and the test has also not been applied as a cestode-screening system. An alternative approach that could be used for larger-scale screening of drugs against adult cestodes, and anthelmintics in general, would be stem cell-based. For E. multilocularis, such an assay would theoretically be feasible, as stem cell cultivation [28] and drug testing on stem cells [31] has already been implemented. However, compared to the simplicity of protoscolex generation and purification, as well as the presence of intact parasite structures in the protoscolex-based approach, we consider the whole-organism protoscolex screening to be more suitable.
In conclusion, the screening assay described in this paper is based on protoscoleces of E. multilocularis. Protoscoleces are sufficiently small to allow testing in 384-well format with multiple parasites per well and the setup is inexpensive. Furthermore, it can basically be carried out with any microscope that allows digital images to be taken, thus no highly specialized equipment is needed. The motility-based assay will allow objective and medium-throughput screening of substances against cestodes, and at the same time enables researchers to visualize morphological effects. As such, the motility-based assay will be further applied for the testing of drug libraries, novel compound classes, and for the in-depth characterization of MMV665807, a new compound of interest with respect to activity against adult cestodes.
|
10.1371/journal.pgen.1007676 | 3 minutes to precisely measure morphogen concentration | Morphogen gradients provide concentration-dependent positional information along polarity axes. Although the dynamics of the establishment of these gradients is well described, precision and noise in the downstream activation processes remain elusive. A simple paradigm to address these questions is the Bicoid morphogen gradient that elicits a rapid step-like transcriptional response in young fruit fly embryos. Focusing on the expression of the major Bicoid target, hunchback (hb), at the onset of zygotic transcription, we used the MS2-MCP approach which combines fluorescent labeling of nascent mRNA with live imaging at high spatial and temporal resolution. Removing 36 putative Zelda binding sites unexpectedly present in the original MS2 reporter, we show that the 750 bp of the hb promoter are sufficient to recapitulate endogenous expression at the onset of zygotic transcription. After each mitosis, in the anterior, expression is turned on to rapidly reach a plateau with all nuclei expressing the reporter. Consistent with a Bicoid dose-dependent activation process, the time period required to reach the plateau increases with the distance to the anterior pole. Despite the challenge imposed by frequent mitoses and high nuclei-to-nuclei variability in transcription kinetics, it only takes 3 minutes at each interphase for the MS2 reporter loci to distinguish subtle differences in Bicoid concentration and establish a steadily positioned and steep (Hill coefficient ~ 7) expression boundary. Modeling based on the cooperativity between the 6 known Bicoid binding sites in the hb promoter region, assuming rate limiting concentrations of the Bicoid transcription factor at the boundary, is able to capture the observed dynamics of pattern establishment but not the steepness of the boundary. This suggests that a simple model based only on the cooperative binding of Bicoid is not sufficient to describe the spatiotemporal dynamics of early hb expression.
| During development, the first thing that an embryo needs to know is the orientation of its body. We study how the head-to-tail axis forms in the fruit fly embryo. To position the axis, the embryo relies on proteins called morphogens broadcasting instructions to other genes, so that cells know, depending on where they are along the axis, what they should become. The concentration of the Bicoid morphogen is much higher in the head of the embryo and lower towards the tail. Bicoid activates hunchback, which divides the embryo in two parts. By visualizing this process in real time in living embryos, we see the hunchback gene activated very efficiently in the head part, where each nucleus expresses the gene, whereas it is not expressed at all in the tail part. Intriguingly, the boundary separating the two domains is precisely positioned and becomes very steep in no more than three minutes. We use a modeling approach to understand how this is achieved so rapidly. Given the parameters of the system, we find that although our model is able to reproduce the fast dynamics of the process, it fails to reproduce the steepness of the boundary suggesting that a more complex approach is needed to capture the additional mechanisms involved.
| Morphogens are at the origin of complex axial polarities in many biological systems. In these systems, positional information is proposed to be provided by morphogen concentrations, which allow each cell to measure its position along the embryo’s axes and turn on expression of target genes responsible for the determination of its identity. Although the existence of these gradients is now well established [1], the quantitative details of their functioning (i.e. how small differences in morphogen concentrations are precisely and robustly interpreted into a threshold-dependent step-like response) remains largely debated [2].
To address this question, we study the Bicoid morphogen, which specifies cell identity along the antero-posterior (AP) axis of the fruit fly embryo [3]. The Bicoid concentration gradient reaches its maximum value at the anterior pole [4] and is distributed steadily in an exponential gradient along the AP axis after one hour of development [4–6] (Fig 1A). Bicoid is a homeodomain transcription factor that binds DNA. Bicoid binding sites are found in the regulatory sequences of Bicoid target genes and are both necessary [7–9] and sufficient [10–12] for Bicoid-dependent expression. Changes in Bicoid dosage induce a shift of the expression boundary of these genes along the AP axis [9, 13] indicating that Bicoid provides concentration-dependent positional information to the system [14]. In young syncytial embryos (nuclear cycles 9 to 13), the major Bicoid target gene hunchback (hb) is expressed under Bicoid control in a large domain spanning the anterior half of the embryo [15]. At cycle 14, hb is also expressed in a narrow posterior domain [16] and in the parasegment 4 [17] and even later during development in the nervous system [18]. Expression of endogenous hb is initiated at two different promoters, the distal promoter P1 and the proximal promoter P2 [19]. P1 is responsible for maternal and late blastoderm expression in parasegment 4 and in the posterior stripe [16, 17]. P2 mediates the early Bcd-dependent expression of hb [17, 19]. At nuclear cycle 14, expression of endogenous hb is also controlled by two distal enhancers, the shadow enhancer and the stripe enhancer, which both contribute to the robustness of hb expression [20, 21].
Here, we focus exclusively on the most early zygotic expression of hb, which is Bicoid-dependent, driven by the P2 promoter and occurring from nc11 to nc13. Just 30 min after the onset of zygotic transcription and the steady establishment of the Bicoid gradient, endogenous hb already exhibits a step-like expression pattern [15]. At this developmental time period, this pattern is characterized by an anterior domain containing almost exclusively hb transcriptionally active nuclei and a posterior domain containing exclusively hb silent nuclei [3] (Fig 1B). The boundary separating expressing and non-expressing nuclei is very steep despite the stochastic nature of transcription in eukaryotic cells [22], the short interphase duration (~5 min) and the subtle difference in the Bicoid concentration (10%) on either side of the boundary [15]. How the hb pattern so rapidly acquires such a steep boundary with high levels of expression in the whole anterior domain is unclear. It could involve a purely quantitative threshold-dependent process, in which Bicoid, acting as a direct transcription activator, is the main source of positional information. Alternatively, additional mechanisms including activation by maternal Hb [12, 15, 23] or posterior inhibitors such as those acting later during development at cycle 14 to set the position of the boundary along the AP axis [7, 24, 25] could be considered. In any case, the mechanism involved should account for a Bicoid dose-dependent effect on the positioning of the boundary along the AP axis. Also, given the early timing of development, most actors in this process are likely to be already present at the onset of zygotic transcription and therefore maternally provided.
To shed light on the formation of the hb expression boundary during these early steps of development, we have previously adapted the MS2-MCP approach to developing fly embryos [26]. This approach allows the fluorescent tagging of RNA in living cells and provides access to the transcription dynamics of an MS2 reporter locus [27, 28]. In our first attempt, we placed the hb P2 proximal promoter region (~ 750 bp) upstream of an MS2 cassette containing 24 MS2 loops and analyzed expression of this reporter (hb-MS2) in embryos expressing the MCP-GFP protein maternally [26]. The hb-MS2 reporter was expressed very early as robustly as endogenous hb in the anterior half of the embryo. However, unlike endogenous hb, the reporter, which only encompasses the hb P2 proximal promoter region (~750 bp), was also expressed in the posterior, though more heterogeneously and more transiently than in the anterior. The different expression of the hb-MS2 reporter and endogenous hb in the posterior suggested that, the 750 bp of the hb P2 proximal promoter were not sufficient to recapitulate the endogenous expression of hb at the onset of zygotic transcription. The most obvious interpretation of these discrepancies was that the hb-MS2 reporter was missing key cis-response elements allowing repression of endogenous hb by an unknown repressor in the posterior [26].
Here, we first show that the homogenously distributed Zelda transcription factor is responsible for the expression of the hb-MS2 reporter in the posterior. BAC recombineering indicated that the MS2 cassette itself mediates Zelda posterior expression and in silico analysis reveals the presence of about 36 putative Zelda binding sites in the MS2 cassette. A new reporter (hb-MS2ΔZelda), placing the hb P2 proximal promoter (~750 bp) upstream of a new MS2 cassette (MS2-ΔZelda), in which those unfortunate Zelda binding sites have been mutated, faithfully recapitulates the early expression of endogenous hb observed by RNA FISH [15]. Thus, the hb P2 proximal promoter (~750 bp) is sufficient for a robust step-like expression of the reporter. Quantitative analysis of the MS2 time traces of this new reporter reveals a transcription process dividing the anterior of the embryo in a saturating zone with stable features and a limiting zone closer to the boundary, with more variable features. A high probability for the promoter to be ON (PON) is reached faster in the anterior where the concentration of Bicoid is higher than close to the boundary where Bicoid concentration is lower. In each interphase, full step-like response is established in not more than three minutes, after which the expression boundary is locked at a given position along the AP axis. To understand this observed dynamics, we used a simple model of position readout through the binding/unbinding of a transcription factor to N operator sites on the hb promoter [29]. The model that best fits the data is able to capture the very fast dynamics of establishment of the boundary. However, high steepness of the experimental pattern (coefficient of the fitted Hill function NHill ~ 7, based on various features of the MS2 time-traces) is not achievable assuming only N = 6 Bicoid binding sites, which is the number of known Bicoid binding sites in the canonical hb promoter [8]. This indicates that a simple equilibrium model only taking into account a single activator and varying degrees of cooperativity between the binding of several molecules on the hb promoter in the reporter, from which a steep hb pattern can emerge [29], is not sufficient to capture its activity and fit the data. It suggests that additional mechanisms are required to define the steepness of the boundary.
In our first attempt of using the MS2 system to study the transcriptional response downstream of Bicoid, we placed the hb P2 proximal promoter region (~ 750 bp) upstream of a classical MS2 cassette [30]. The reporter (hb-MS2) was expressed in the posterior of the embryo and did not recapitulate expression of endogenous hb [26]. The Zelda transcription factor is a major regulator of the first wave of zygotic transcription in fruit fly embryos [31] and is involved in the transcriptional regulation of the hb gene [32]. To determine how Zelda contributes to the expression of the hb-MS2 reporter, we analyzed expression of the reporter by live imaging (S1 Movie) and double RNA FISH (using hunchback and MS2 probes, Fig 2A) in embryos expressing various amounts of maternal Zelda. As expected [26], wild-type embryos show high expression of the hb-MS2 reporter both in the anterior and the posterior domain ([26] and Fig 2B, top panel), with an average of 80% of expressing nuclei dispersed through the posterior domain (Fig 2B, bottom panel). In embryos from zelda heterozygous mutant females (zldMat-/+), expression of the hb-MS2 reporter is reduced by 5 fold in the posterior (with only 15% of active nuclei). In embryos from zelda mutant germline clones, completely devoid of Zelda maternal contribution (zldMat-/-), expression boundaries of endogenous hb and hb-MS2 reporter are shifted towards the anterior (Fig 2D). Moreover, posterior expression of the hb-MS2 reporter is reduced to less than 1% of posterior nuclei (Fig 2D). We confirm thus that Zelda maternal proteins contribute to the early expression of endogenous hb [32] and conclude that posterior expression of the hb-MS2 reporter in early embryos is mostly due to Zelda.
To understand the discrepancies of expression between endogenous hb and the hb-MS2 reporter at early cycles, we first aimed at identifying the minimal sequence sufficient to recapitulate early expression of the hb locus. While a transgene carrying 18 kb of the hb locus was shown to recapitulate hb anterior expression at nc13 and nc14, its expression at earlier cycles had not been documented [20, 21, 33]. RNA FISH, using a hb probe on embryos carrying a single insertion of the hb-18kb BAC, reveals ongoing transcription at the three hb-encoded loci and indicates that expression of the hb-18kb BAC and endogenous hb largely overlap in the majority of anterior nuclei (Fig 3B). No RNA FISH signals are detected within the posterior domain of these embryos indicating that the hb-18kb BAC encompasses all the regulatory sequences to spatially control hb expression during early nuclear cycles.
Taking advantage of BAC recombineering, we generated MS2-hb-18kb transgenes carrying insertions of the MS2 cassette either in the 5’UTR within the intron of hb (5’MS2-hb-18kb) or in the 3’UTR of hb (3’MS2-hb-18kb). Expression of these new MS2 reporters was assessed by live imaging (S2 Movie) and double RNA FISH using a hb probe and an MS2 probe. In the anterior of nc11 embryos either homozygous for the 5’MS2-hb-18kb transgene (Fig 3C) or heterozygous for the 3’MS2-hb-18kb transgene (Fig 3D), the hb probe reveals at most four spots of ongoing transcription : one at each of the two endogenous hb and two (C) or one (D) at the MS2-hb-18kb loci. In most cases, two of the hb spots co-localize with two (C) or one (D) MS2 spot(s), which specifically label ongoing transcription at the MS2-hb-18kb loci. In these embryos, we also detect hb and MS2 spots in posterior nuclei (Fig 3C and 3D), which co-localize for most of them, thus revealing ongoing transcription at the MS2-hb-18kb loci. Posterior expression of the MS2-hb-18kb loci is also detected in living embryos expressing the MCP-GFP (S2 Movie). Altogether, these data strongly argue that posterior expression of the MS2-hb-18kb transgenes is mediated by the MS2 cassette and suggest that Zelda-dependent posterior expression of the hb-MS2 reporter is mediated by cis-acting sequences in the MS2 cassette.
Given the trans-acting effect of Zelda on the posterior expression of the hb-MS2 reporter (Fig 2) and the enhancer-like behavior of the MS2 cassette for posterior expression (Fig 3), we searched for potential Zelda binding sites in the MS2 sequence, using the ClusterDraw2 online algorithm [16] and Zelda position weight matrix [34]. The canonical Zelda binding site is a heptameric motif (CAGGTAG, Fig 4A) over represented in the enhancers of pre-cellular blastoderm genes [32, 35]. Strikingly, in the sequence of the MS2 cassette we find the motif CAGGTCG (a single mismatch with the canonical Zelda site) repeated 12 times and the motifs TAGGTAC (two mismatches) and TAGGCAA (three mismatches) each repeated 12 times (Fig 4B). This in silico analysis indicates that the MS2 sequence contained a total of 36 potential Zelda binding sites all located within linkers between MS2 loops (Fig 4B). Although we cannot be conclusive about the affinity strength of these various binding sites for Zelda, all of them share high similarity with TAGteam motifs [32]. We thus engineered a new MS2 cassette mutating the 36 putative Zelda binding sites of our original hb-MS2 reporter and inserted this new MS2-ΔZelda cassette under the control of the hb canonical promoter as in the original hb-MS2 reporter. Expression of this new hb-MS2ΔZelda reporter was assessed in living embryos expressing the MCP-GFP protein (S3 Movie). We do not detect any MS2 spots in the posterior of the embryo (Fig 5A) indicating that unlike our original hb-MS2 reporter [26], the new hb-MS2ΔZelda reporter is expressed exclusively in the anterior at early cycles 10 to 11 as detected for endogenous hunchback by RNA FISH [3].
As RNA FISH are performed on fixed embryos whereas the MS2 data are obtained from live material, we wondered whether these two different approaches provide consistent quantification of hb transcriptional activity when focusing on the “steepness” of the expression boundary. Therefore, the probability to be active for hb-MS2ΔZelda loci at a given position along the AP axis and at a given time (PSPOT(t)) was extracted from movie snapshots and compared to RNA FISH data of endogenous hb expression [3]. To compare data from several embryos, embryos were aligned fixing the origin of the AP axis when the probability for a locus to experience transcription at any time during the interphase (PON) is equal to 0.5 (for the definition of the boundary see details in S3 Text). This embryo alignment allows us to compensate for input noise (variability in the Bicoid gradient), which was shown to be of about 2–3% EL [6], and focus only on the output noise (noise in the transcriptional response). Throughout the paper, we refer to average measures of expression by computing the probability of the locus to be ON as a function of position along the AP axis and as a function of time. We named this time-dependent probability to be ON, PSPOT (t). PSPOT(t) is thus a measure of the instantaneous gene activity at a given position along the AP axis and at a given time within the nuclear cycle. We first assigned a value of 1 to a nucleus that expresses the MS2-MCP gene above a certain threshold (see methods) at any time during the nuclear cycle. PSPOT(t) is calculated by averaging over nuclei at a given position along the AP axis. We also use a cumulative PSPOT named PON, which is used only for embryo alignment and which is a cumulative statistic per nuclear cycle. PON indicates the probability for a nucleus at a given position to experience transcription during nuclear cycle. A value of 1 is assigned to each nucleus experiencing transcription during the cycle and PON is obtained by averaging over the nuclei at a given position. PSPOT(t) can decrease within a nuclear cycle if nuclei stop expressing, whereas PON cannot decrease.
As shown in Fig 5B–5D, the curves plotting the mean spot appearance PSPOT(t) as a function of position along the AP axis are similar when extracted either from the RNA FISH data (dashed line, [3]) or the MS2 movie snapshots (blue lines) with Hill coefficients varying from 4.5 (nc11) to ~ 7 at nc12 and nc13. Thus, the hb-MS2ΔZelda reporter is expressed at early cycles in an anterior domain with a boundary as steep as the boundary of the endogenous hb expression domain. These experiments show that the hb canonical promoter is sufficient to faithfully recapitulate the early zygotic expression of endogenous hb. This is in contradiction with our first interpretation [26] deduced from the expression of the original hb-MS2 misleading reporter with its high number of unexpected Zelda binding sites. Nevertheless, the now almost perfect match between PSPOT(t) activity along the AP axis of the FISH data of endogenous hb expression and of the data extracted from movies of the hb-MS2ΔZelda reporter at the end of the interphase, ensures that the analysis of this reporter expression dynamics can help understand how the step-like expression of hb arises.
The MS2 movies provide access to the transcription dynamics of each hb-MS2 ΔZelda single locus in all the visible nuclei of developing embryos. Key features are extracted from the time traces of the MS2-GFP spots (Fig 6A), choosing as the origin for time the onset of interphase for that particular nucleus (see details in S2 Text and S4 Fig). These features include: i) the initiation time (tinit) which measures the time period from the onset of interphase to the first detection of the MS2 transcription signal at the locus, ii) the time period during which the locus is activated (tactive) and iii) the time period at the end of interphase during which the locus is turned off (tend). From the time traces (Fig 6A), the integral activity (ΣI) integrates the area under the trace and provides a relative measure of the total amount of mRNA produced. The average mRNA production rate (μI) is calculated by dividing ΣI by tactive. Features are obtained from 5 (nc11), 8 (nc12) and 4 (nc13) embryos. Embryos were aligned spatially fixing the origin of the axis at boundary position (PON) at nc12 and the origin of time was calculated for each nuclei as the origin of the respective cycle (see S2 Text and S4 Fig). As shown in Fig 6, several of these features exhibit different behaviors depending of the position along the AP axis. Notably, tinit and ΣI appear more variable when loci are located close to the boundary than in the most anterior part (Fig 6B–6G). Similarly, in a region of about 10% EL at the boundary, the mRNA production rate drops from the constant value reached in the anterior region to 0 at the boundary (Fig 6H–6J). Thus, the dynamics of the transcription process at the hb-MS2ΔZelda reporter exhibit two distinct behaviors: in the anterior of the embryo, time trace features are stable likely reflecting a maximum PolII loading rate and saturated levels of Bicoid; in a region of ~ 10% EL anterior to the expression boundary, time trace features are fluctuating reflecting limiting amounts of Bicoid. The pattern steepness corresponds to a Hill coefficient of ~7 to 8 of the regulation function (see S3 Text). The MS2 live imaging gives us the opportunity to really decipher the dynamics of the hb expression compared to the FISH that was just giving us PSPOT(t) for one arbitrary time point (probability for hb loci to be active at a given position along the AP axis and at the time of embryo fixation).
The temporal dynamics of the hb pattern establishment at the scale of the whole embryo (along the AP axis) were extracted from the MS2 movies: the probability for the locus to be ON as a function of the position along the AP axis and time in the cycle (PSPOT(t)) can be visualized in S4 Movie and is plotted in the form of kymographs on pulled embryos as described above (Fig 7A–7C). At each cycle, spots can appear as early as ~ 150 s after mitosis. Given the interruption of transcription during mitosis, this limit of 150 s corresponds to the period required to re-establish transcription during the interphase and likely includes genome de-condensation, the time it takes for the Bicoid protein to be imported in the nucleus after mitosis, and the time it takes for the MS2 system to produce a signal that is above background. After this period, expression rapidly turns on in the anterior to reach the plateau value where all nuclei express the reporter (PSPOT(t) ~ 1). The time period to reach the plateau is shorter at the most anterior position in the field of view imposed by the movie recording (~ - 25% EL) and increases with the distance to the anterior pole (Fig 7D–7F) consistent with a dose-dependent activation process of hb by Bicoid [4, 8]. Then, the expression pattern is steadily established, which includes fixing the boundary position and steepness (see S4 Text for measurement details). This stable state lasts for a period of time which varies with the length of the cycle until hb expression rapidly gets switched off simultaneously at all positions along the AP axis: at 400 s at nc11 (Fig 7A & 7G), 500 s at nc12 (Fig 7B & 7G) and 750 s at nc13 (Fig 7C & 7G). Importantly, the dynamics of the boundary positioning is the same at the three nuclear cycles considered (Fig 7G) and the steady state of boundary positioning is reached rapidly: ~330 s after the onset of the cycle corresponding to ~ 180 s after the first hints of transcription at this position (Fig 7G). Dynamics of pattern steepness and average spot intensity also exhibit similar behaviors (see S5 Text and S9 Fig, respectively). Thus, expression of the hb-MS2ΔZelda reporter allows us to directly observe position-dependent transcriptional activation, which is consistent with Bicoid dose-dependent transcriptional activation. Also, it demonstrates that the steady-state of positional measurement is reached in no more than 3 min at each cycle with very similar dynamics between cycles. Once the steady state is reached, the hb boundary is fixed around the position ~ -5% EL.
To better understand how the stable response downstream of Bicoid is achieved, we build a stochastic model of hb expression regulation by the Bicoid transcription factor (TF), coupled with a stochastic transcription initiation process assuming random arrival of RNA polymerases when the gene is activated. The mechanism of hb expression regulation through the cooperative binding of multiple TFs to the promoter was originally proposed in 1989 to explain how the shallow Bicoid gradient could give rise to an expression pattern with a steep boundary [8, 9]. It was subsequently proposed [36] that within an equilibrium binding model, a pattern steepness quantified by a Hill coefficient of N requires a promoter with at least N TF binding sites. Here, we consider a model with N TF binding sites, where the promoter state Pi is decribed by the probability of having i TF bound at a given time (Fig 8A). Transitions between the states occur via binding and unbinding events of TF to the sites. We assume an equilibrium binding model, in which all transitions are reversible and the binding sites are identical. As a result the state of the promoter is described solely by the number of bound TF, and not their position on the cis regulatory array. We assume that gene expression is activated only when all N binding sites are bound by TF. In the regime of parameters corresponding to a high pattern steepness, this “all-or-nothing” assumption is shown to have very limited impact on the pattern dynamics (see [29] and our companion paper, [37]). Following this activation, RNA polymerases can arrive to the promoter in a Poisson process and initiate the downstream transcription initiation process. Our simplified model does not account for a non constant polymerase arrival rate [38].
We investigate the hb pattern dynamics by solving a stochastic time dependent master equation (Eq 7 in S6 Text) and considering binding rates that vary as a function of TF concentration [TF]: at the boundary position, TF concentrations are lower than in the anterior region [5, 39] and it takes more time for a diffusing TF to reach binding sites at the promoter than in the anterior region where TF concentration is higher. The value of the binding rate constants is further limited by the value of the diffusion coefficient (see S6 Text, 1). In our model, the cooperativity between the binding of TFs is modeled implicitly by the value of the unbinding rates of the TF to the binding sites: higher cooperativity corresponds to higher stability of the promoter state with N bound TF, which is modeled through a lower unbinding rate. The unbinding rate is limited at the boundary so that PSPOT (t) = 1/2. Meanwhile, the unbinding rate constants are kept constant along the AP axis (independant of Bicoid concentration) but differ depending on the occupancy state of the promoter to account for copperativity. At time t = 0, all binding sites are free of TF.
Motivated by the 6 known Bcd binding sites on the hb promoter [8, 17], we first consider a model with 6 binding sites (N = 6). The binding and unbinding rates are chosen (Eq 5 in S6 Text) to achieve the closest steepness and establishment period to the hb pattern observed in the movies (Fig 7A–7C). The model fits are performed on data pulled from all embryos and nuclear cycles. The gene expression pattern dynamics, shown as the probability (PSPOT(t)) of a nucleus having an active locus (bright spot) as a function of position along the AP axis and as a function of time during the nuclear cycle, shows a good qualitative agreement between the model and the data. In Fig 8B, similarly to Fig 7A–7C, initially there is no expression since the TF are not bound to the promoter. Active transcription loci first appear near the anterior pole, where the activator concentration is the highest, and then transcription activation propagates quickly to the mid-anterior region. After a certain amount of time, the steep expression pattern becomes stable, with the boundary (dashed green line) located near the mid-embryo region where the hb boundary is located. However, using the canonical number of binding sites (N = 6) results in a steepness of the boundary lower than that observed in the experimental data (Fig 8C). Increasing the binding sites number from 6 to 7, 8, 9 or 10 allows us to match the observed boundary steepness to the experimental data (H~7). Yet, we have shown in a companion manuscript [37] that the steeper the patterns are, the longer the time needed for them to establish. Specifically, obtaining the observed Hill coefficient of H~7 in 3 minutes is not possible with 7 or 8 binding sites but requires 9 binding sites (Fig 8E). Increasing N beyond 9 did not result in a significantly better fit (see S6 Text, 5).
In this study we show that the removal of 36 putative binding sites for the transcription factor Zelda (unfortunately present in the sequence of the MS2 cassette) reveals a new temporal dynamics of the hb canonical promoter at the onset of zygotic transcription. Unlike our original hb-MS2 reporter [26], the new hb-MS2ΔZelda reporter faithfully reproduces the early zygotic expression of the endogenous hb observed with RNA FISH [3, 26]. It indicates that the 750 bp of the hb locus, including 300 bp of the proximal enhancer, the P2 promoter and the intron, are sufficient to reproduce the endogenous expression of the hb gene in the early nuclear cycles (11 to 13). The dynamics of establishment of the hb pattern at these early stages of development is thus properly captured by the new hb-MS2ΔZelda reporter.
These MS2 movies provide access to the hb pattern dynamics, which was not perceivable in previous in situ experiments on fixed embryos. hb expression first occurs in the anterior then proceeds to the boundary region. A difference in the activation time (initiation time) following mitosis is observed even within the anterior region, allowing us to visualize for the first time position-dependent activation of hb and thus likely dose-dependent activation by Bicoid. Analysis of the MS2 time traces indicates that the transcription process is more variable among nuclei at the boundary of the expression domain, where Bicoid concentration is low and probably limiting, than in the anterior where the concentration of Bicoid is high. Tailor-made analysis of the time traces allowed us to extract different kinetic parameters of promoter activity in these two regions and to demonstrate that transcription of the hb-MS2ΔZelda reporter is bursty (described by a two state model) in both regions [40]. This anti-correlation between the relative variability of mRNA production among nuclei and the Bicoid concentration (position along the AP axis) supports the idea that Bicoid interactions with the hb promoter are rate-limiting processes contributing to “bursty” transcription. Nevertheless, rate-limiting interaction of Bicoid with DNA is not the sole factor contributing to bursty transcription, as it is also observed in the anterior region with very high Bicoid concentration [40]. Thus, despite extremely fast transcription initiation imposed by the 5 min interphase and frequent mitoses, bursty transcription is clearly observed at the onset of zygotic transcription in fly embryos.
The MS2 movies indicate that the hb boundary is established within 3 minutes at each nuclear cycle with a very high steepness (H ~7). How the steepness and positioning of the boundary are reached so rapidly is unclear. Our data (Fig 6 and Fig 7) indicate that it takes more time for hb expression to reach steady-state levels in the boundary region and that Bicoid is thus likely to be a rate-limiting factor in the formation of the hb pattern around the boundary region. The time scale of interactions (i.e. binding and unbinding) of Bicoid with the hb promoter is critical in determining how quickly the accurate hb response is established. Therefore, the interactions between Bicoid molecules and the hb promoter need to be modeled explicitly, rather than implicitly. We propose an equilibrium model of transcription regulation via the binding/unbinding of transcription factors to the operator sites of the target promoter. This model can account for various types and degrees of cooperativity between the binding of TFs, from which a steep hb pattern can emerge [29].
Using this model, we show in a tandem paper that accommodating a steep gene expression pattern within the considered model requires very slow promoter dynamics and thus would result in a very long pattern formation time [37]. Considering this trade-off, the steepness and the pattern formation time observed from the movies are even more intriguing. The most relevant model with 6 binding sites, motivated by previous work [8], fits well the pattern dynamics but fails to reproduce the observed steepness of the boundary in such a short time period of 3 min. The failure of the 6 binding sites model to completely reproduce the experimental data indicates that the assumption that 6 binding sites for Bicoid are sufficient for the observed response is wrong and that additional mechanisms have to be included in the model to enhance the steepness of the boundary to the observed level. It was recently proposed that energy expenditure, encoded as non-equilibrium binding of the TF, can allow a Hill coefficient greater than the number of binding sites (up to 11 with 6 sites) [29]. Alternatively, different regulatory scenarios could play a role. First, we found that increasing the number of binding sites up to 9 allows a significantly better fit of the model with the data than with 6 binding sites. It is thus possible that the 750 bp of the hb gene that are sufficient to elicit the pattern contain more than 6 Bicoid binding sites. As recently pointed out, the importance of low affinity binding sites might be critical to confer specificity and robustness in expression [37] and a closer analysis of the hb regulatory sequence looking for potential weak binding sites for Bicoid might clarify this point. A second possibility is the involvement of other transcription factors distributed as gradients that could bind to the hb promoter and contribute to the increase in the steepness of the boundary. The Hb maternal protein, which is also expressed as an anterior to posterior gradient and able to bind to the endogenous hb promoter [38], contributes to the hb expression process by allowing expression at lower Bicoid concentration thresholds [12] and faster activation [15]. Although this has not yet been investigated, maternal Hb might also contribute to sharpen the hb boundary in a timely manner. Alternatively, maternal repressors expressed as gradients in the posterior region or downstream of the boundary could also contribute to the steepness of the hb boundary. Among potential candidates are Caudal, which is expressed as a posterior to anterior gradient but so far has been described as a transcriptional activator in fly embryos [41], or Capicua, a transcriptional repressor at work in the center of the AP axis, where the hb boundary forms [7].
From the movies, the dynamics of establishment (before reaching steady state) of the hb pattern appears to be invariant at the three nuclear cycles considered (nc11, nc12 and nc13). At these three nuclear cycles, it takes ~180 s from the detection of the first MS2-MCP spots to the establishment of the pattern near the mid-embryo position (Fig 7G). This invariance in the dynamics of establishment indicates that there are no dramatic changes in the regulation of the transcription process during the three cycles suggesting that Bicoid remains one of the main patterning factors of hb transcription dynamics at the boundary region in these stages of development. The other transcription factors involved in hb expression [15, 19], if any, would need to be stably maintained during the three cycles. As mentioned above, one possible factor is the Hb protein itself, which was shown to contribute to the expression of the hb expression pattern with both a maternal and a zygotic contribution [12, 15]. A likely hypothesis is that up to nc13, the zygotic Hb protein production is balanced by maternal Hb protein degradation, thus leading to a stabilized Hb gradient over the three considered cycles.
Previous observations of Bicoid diffusion in the nucleus space pointed out that it requires at least 25 minutes for a single Bicoid binding site to sense the Bicoid concentration with 10% accuracy, a level observed using FISH and protein staining experiments [5, 15, 39]. This estimation is based on the Berg-Purcell limit in the precision of concentration sensing of diffusing molecules via surface receptors [42–44]. It should be noted that the Berg-Purcell limit applies when the interactions between the binding sites and the TF are independent. This limit for the Bicoid/hb system results in a non-steep gene expression pattern (H ~ 1.31, see S2 Table). In-depth analysis of the model in the tandem paper [45] shows that increasing the pattern steepness slows down the switching rate between the gene’s active and inactive states, due to the required cooperativity between the binding sites [29]. Consequently, our fitted models (with N = 6 or N = 9 binding sites) result in much higher errors in the integrated readout when compared to the model of independent binding sites (S12 Fig), and require a significantly longer integration time to achieve a specific precision level. Thus, the answer to the precision of the hb pattern achieved in such a short interphase duration remains elusive. In the future, systematic studies using synthetic promoters with a varying number of Bicoid binding sites and quantitative analyses of promoter dynamics captured with the MS2-MCP system will help characterize not only the cooperativity of Bicoid binding sites but also the kinetics of the downstream processes once the gene is activated. Added to this, the Bicoid search time for its binding sites on hb [15, 39] also needs to be revisited. The employed value (~4 s) in our model is estimated assuming a 3D search process inside the nucleus space but if Bicoid can slide along DNA in search for the sites, the process can be ~100 times faster [46]. Such a possibility is compatible with fluorescence correlation spectroscopy measurements of Bcd-eGFP motion [15, 47] and the recent observed clustering of the Bicoid molecules across the embryo [48]. The advent of single molecule tracking methods [48, 49] represent a promising approach to further shed lights on the mechanism of this process.
Importantly, despite these very rapid and precise measurements of Bicoid concentration along the AP axis, the expression process itself shows great variability in the total amount of mRNA produced during the interphase per nucleus (δmRNA/<mRNA> ~150% in the boundary region) [40]. It is thus difficult to gauge to which degree errors in sensing Bicoid concentration contribute to the variability of the total mRNA produced at the boundary [40]. This also raises the question of understanding how precision in the downstream processes required for embryo segmentation is achieved at the scale of the whole embryo. If the embryo is capable of spatially averaging hb expression between nuclei at the same AP position (for example by the diffusion and nuclear export of its mRNA and nuclear import of its protein) [25], this will help the system distinguishing between the anterior and posterior region based on the hb transcription pattern alone. However, spatial averaging has a limit, due to the very short time available, the limited number of nuclei and the finite diffusion coefficient of the hb gene products [25]. It is likely that nuclei need to integrate hb gene expression over the nuclear cycle to reduce the noise in the hb readout. In nc11, the pattern collapses soon after reaching steady-state due to the next mitosis round. Therefore, any integrations of gene expression in nc11 and earlier (interphase duration shorter than ~400 s) are likely to lead to bias the pattern boundary to the anterior. Only starting at nc12 is the interphase duration long enough to reliably produce a steep pattern with the border around the mid-boundary position, based on the number of mRNA produced per nucleus (ΣI).
The original reporter hb-MS2 and the Nup-RFP and MCP-GFP transgenes, both inserted on the second chromosome, were from [26]. All transgenic stocks generated in this study were obtained by BestGene. The zld294FRT19A chromosome was a gift from C. Rushlow [32] and female germline clones were induced using heat-shock FLP recombinaison [50] with a OvoD1hsFLPFRT19A chromosome (# 23880, Bloomington). All stocks were raised at 25°C.
The hb-18kb-BAC spanning the hb locus was the BAC CH322 55J23 obtained from PACMAN BAC libraries [51]. The 24xMS2 cassette was inserted in the 5’UTR within the hb intron (5’MS2-18kb-BAC) or in the 3’UTR (3’MS2-18kb-BAC) using BAC recombineering [52] (details in S1 Text). The plasmid used to generate the new hb-MS2ΔZelda reporter was generated by replacing in the pCasPeR4-hb-MS2 construct from [26], the MS2 cassette by a new MS2 sequence synthetically generated by Genscript in which all putative Zelda binding sites (24xMS2-SL-ΔZelda) had been mutated. The sequence coding for the CFP in the pCasPeR4-hb-MS2 construct from [26] was also replaced by the sequence coding the iRFP (Addgene 45457) in which a unique putative Zelda binding site had been mutated (See S1 Text for more information).
Sequence-based binding-site cluster analysis was performed using the online software ClusterDraw2 v2.55 at line.bioinfolab.net/webgate/submit.cgi. PWM for Zelda has been generated from [34] using the following Zelda binding sites: CAGGTAG; CAGGTAA; TAGGTAG; CAGGTAC; CAGGTAT; TAGGTAA; CAGGCAG and CAGGCAA. All heptamers detected as being a potential Zelda binding sites have been mutated and the new sequence synthesized by Genscript. The sequence of the new MS2 ΔZelda cassette is given in S1 Text.
Embryo fixation and RNA in-situ hybridization were performed as describe in [3]. Briefly, RNA probes were generated using T7/T3 in-vitro transcription kit (Roche). hb RNAs were labelled with digoxigenin-tagged anti-sense probes detected with a sheep anti-dig primary antibody (1/1000 dilution, Roche) and donkey anti-sheep Alexa 568 secondary antibody (1/400 dilution, Invitrogen). MS2 containing RNAs were labelled with biotin-tagged anti-sense probes detected with a mouse anti-bio primary antibody (1/400 dilution, Roche) and chicken anti-mouse Alexa 488 secondary antibody (1/400 dilution, Invitrogen). Embryos were incubated 10 min with DAPI for DNA staining and 10 min in WGA-Alexa 633 (1/500 dilution, Molecular Probes) for nuclear envelop staining. Fixed embryos were mounted in VectaShield (Vector) and imaged in 3D (~20Z x 0.45μm) with a XY resolution of 3040*3040, 8bits per pixels, 0.09μm pixel size, 1 airy unit using a Zeiss LSM780 confocal microscope with a Zeiss 40x (1.4 NA) A-Plan objective. A full embryo 3D RAW image is composed on three 3D RAW Images that were stitched using Stitch Image Grid plugins from FIJI with 10% overlap and the linear blending fusion method. Image processing and analyzing were performed as describe [26]. Embryos of the proper stage in the nc11 interphase were selected according a threshold based on the nuclear area above 80 μm2 (corresponding to late interphase) as in [3]. The expression map of endogenous hb or MS2 transgenes have been manually false colored on FIJI and flatten on the nuclear channel.
Imaging conditions were comparable to those outlined in [26, 53]. Embryos were collected 1h after egg laying, dechorionated by hand, fixed on the cover slip using heptane-dissolved glue, and immersed in to halocarbon oil (VWR). Mounted embryos were imaged at ~ 25°C on a Zeiss LSM780 confocal microscope with a Zeiss 40x (1.4 NA) A-Plan objective. Image stacks of the lower cortical region of the embryo close to the middle of the AP axis (pixel size 0.2 μm, 0.54 μs pixel dwell time, 8 bits per pixels, confocal pinhole diameter 92 μm, distance between consecutive images in the stack 0.5 μm, ~1200x355 pxl, ~30 z-stacks) were collected continuously. The GFP and RFP proteins were excited with a small fraction of the power output of a 488nm and a 568nm laser, 1.2% and 2% respectively. Images were acquired using the ZEN software (Zeiss). For each embryo, a tiled image of the midsection of the whole embryo was obtained, by stitching 3 separate images, from which the position of the anterior and posterior poles could be inferred.
Live imaging processing was performed in MATLAB as in [26]. Following imaging, movies are checked manually to verify all the nuclei included in data analysis are fully imaged in their depth and incompletely imaged nuclei (mostly nuclei at the periphery of the imaging field) are excluded. Nuclei segmentation is performed in a semi-automatic manner using our own software [54]. Only nuclei that exist throughout the nuclear interphase are used for the analysis. The MS2 spot detection was performed in 3D using a thresholding method. An average filter was applied before thresholding on each Z of the processed time point for noise reduction. MS2 spots were detected by applying a threshold equal to ~2 fold above background signal and only the spots composed of at least 10 connected voxels were retained. The intensity of the 3D spot is calculated as the sum of the voxel values of each Z-stack. At the end of nc13, some MCP aggregates may cause false spot detections: they are less bright and their shape is more spread compare to the MS2-MCP spots. The aggregates are eliminated automatically by raising the spot detection threshold without affecting the detection of the MS2 spots. We manually checked each movie to ensure the correct spot detection. The data from a segmented movie indicates for each nucleus its segmentation profile, identifier number and the intensity trace of the detected spot over time.
During mitosis, nuclei divide in waves, usually from the embryo poles. Therefore, nuclei at the anterior pole may produce MCP-MS2 spots earlier due to either earlier chromatin decondensation or earlier reentrance of Bicoid into the nucleic space [6]. We correct for this by realigning all the intensity traces by choosing the origin for time for each trace when the two sibling nuclei are first separated (see S2 Text and S4 Fig)
The general model of transcription regulation through transcription factor (TF) binding/unbinding to the operator sites (OS) is a based on a graph-based linear framework [29, 55, 56]. We introduce a single time scale, which is the TF search time for a single operator site at the boundary tbind ~ 4s. The details of the model are described in S6 Text and S7 Text.
|
10.1371/journal.pntd.0005729 | Dengue prediction by the web: Tweets are a useful tool for estimating and forecasting Dengue at country and city level | Infectious diseases are a leading threat to public health. Accurate and timely monitoring of disease risk and progress can reduce their impact. Mentioning a disease in social networks is correlated with physician visits by patients, and can be used to estimate disease activity. Dengue is the fastest growing mosquito-borne viral disease, with an estimated annual incidence of 390 million infections, of which 96 million manifest clinically. Dengue burden is likely to increase in the future owing to trends toward increased urbanization, scarce water supplies and, possibly, environmental change. The epidemiological dynamic of Dengue is complex and difficult to predict, partly due to costly and slow surveillance systems.
In this study, we aimed to quantitatively assess the usefulness of data acquired by Twitter for the early detection and monitoring of Dengue epidemics, both at country and city level at a weekly basis. Here, we evaluated and demonstrated the potential of tweets modeling for Dengue estimation and forecast, in comparison with other available web-based data, Google Trends and Wikipedia access logs. Also, we studied the factors that might influence the goodness-of-fit of the model. We built a simple model based on tweets that was able to ‘nowcast’, i.e. estimate disease numbers in the same week, but also ‘forecast’ disease in future weeks. At the country level, tweets are strongly associated with Dengue cases, and can estimate present and future Dengue cases until 8 weeks in advance. At city level, tweets are also useful for estimating Dengue activity. Our model can be applied successfully to small and less developed cities, suggesting a robust construction, even though it may be influenced by the incidence of the disease, the activity of Twitter locally, and social factors, including human development index and internet access.
Tweets association with Dengue cases is valuable to assist traditional Dengue surveillance at real-time and low-cost. Tweets are able to successfully nowcast, i.e. estimate Dengue in the present week, but also forecast, i.e. predict Dengue at until 8 weeks in the future, both at country and city level with high estimation capacity.
| Dengue is a fast-growing mosquito-borne viral disease, with an estimated annual incidence of 390 million infections, of which 96 million manifest clinically. Dengue burden is likely to increase in the future. Mentioning a disease in social networks is correlated with physician visits by patients, and can be used to estimate disease activity. Traditional, biologically-focused monitoring techniques, based on laboratory diagnostics, are accurate but costly and slow. Alternative approaches for surveillance aim to capture health-seeking behavior at earlier stages of disease progression, specially capturing the asymptomatic and mild clinic manifestation population who do not seek medical care formally. Twitter data have potential application for Dengue surveillance, improving the estimation and prediction of the disease, in space and time, being a valuable and low-cost addition to assist traditional surveillance. We show that tweets are strongly associated with Dengue cases. Tweets are a useful tool for estimating and forecasting Dengue cases until 8 weeks in the future, both at country and city level, even in less developed areas.
| Infectious diseases are a leading threat to public health, economic stability, and other key social structures [1]. Efforts to mitigate these impacts depend on accurate and timely monitoring to measure the risk and incidence of the disease [2,3,4]. Early detection of disease activity and rapid responses can reduce the impact of diseases [5]. The interdisciplinary field of computational social science aims to quantify real-world social phenomena using large datasets known as ‘big data’, based on data from social networks, such as Twitter and Facebook, to describe behavioral patterns in novel contexts [6,7]. The relative frequency of mentioning a disease in certain social networks, as in Twitter and others, is highly correlated with patients visits to doctors, making it possible to accurately estimate disease activity in each region of a country, with a small reporting lag [3,5]. Also, web search query data from Google and Wikipedia are capable of tracking disease activity and are available in near real-time [2,8].
Twitter is a unique social media channel, since users inform and discuss, through their short 140-character messages or ‘tweets’, about the most diverse topics, including health conditions [3,9,10]. This free social networking service has more than 190 million users registered worldwide and processes about 55 million tweets per day, with the possibility of sentiment analysis and location selection [3,9,10,11]. According to the US Bureau (2014), 45% of the Brazilian households have access to internet and 48% of the population are users of social media products, with an average of 3 hours per day spent in this activity. In Brazil, the main social media of preference is Facebook (94%), followed by Google Plus (75%), Twitter (56%) and LinkedIn (54%) [12]. Important to notice that Twitter usage (86%), as of other social media products, is mostly via mobile phone, which in Brazil has 134% of coverage, meaning an average of 1.34 cell phone subscriptions per person [12].
Dengue is an important public health burden, likely to increase in the future due to increased urbanization, scarce water supplies and environmental change [13]. Dengue is ubiquitous throughout the tropics and a fast spreading viral mosquito-borne disease [14], with an incidence increase of 30 times during the last 50 years [15]. The World Health Organization (WHO) estimates 50 to 100 million new infections per year in 100 different countries, with half the world’s population, or 3.5 billion people at risk [15], but more recent studies estimate the total incidence to be 390 million Dengue infections per year, of which 96 million manifest clinically [14]. Currently, there is no specific antiviral treatment to reduce severe illness or an effective vaccine to induce strong protection from infection [16]. The epidemiology of Dengue in Brazil is characterized by increasing geographical spread, as well as the total incidence of reported cases [17]. The epidemiological dynamics of Dengue disease is complex, and difficult to predict, partly due to the weaknesses of passive surveillance systems [13,17]. The majority of infections are clinically non-specific, consequently, Dengue disease is often underdiagnosed [13], but these patients are also infectious to mosquitoes and contribute for the transmission of the disease [18].
Bureaucracy and lack of resources have interfered with timely detection and reporting of Dengue cases in many endemic countries, including Brazil, where reporting delay is estimated to be of 3 to 4 weeks [3, 19]. Traditional, laboratory and clinically based diagnostic techniques are accurate but costly and slow. Alternative approaches to surveillance aim to capture health-seeking behavior at earlier stages of disease progression, specially capturing those with mild clinic manifestation population who do not seek medical care formally [2,6]. Some studies indicate that digital media reports reflect national epidemiological trends, acting as proxy for surveillance to provide early warning and situation awareness of emerging infectious diseases and Dengue [4,20]. While traditional Dengue surveillance data suffer from substantial delay, web-based data can fill in the gap providing a near real-time source of information [8,21,22]. Previous studies have shown that Twitter is a real-time source of information on Dengue symptoms activity in a population, and shows strong correlation with the number of notified cases [3, 23]. Some advocate that Twitter-based surveillance efforts may provide an important and cost-effective supplement to traditional disease-surveillance systems [10,11]. Besides Twitter, web search query data, such as Google Trends, were also found to be capable of tracking Dengue activity. Proper combination of these two sources of information may provide timely information to public health officials and contribute to real-time predictive models [8, 23]
In this study, we aimed at investigating if tweets with personal indication of Dengue content could be integrated with clinical Dengue data to produce an accurate model for the early detection and monitoring of Dengue epidemics at country and city level. For comparison, we also report other available web-based data, Google Trends and Wikipedia. Also, we studied the factors that might influence the goodness-of-fit of the proposed model. We concluded that a simple model using tweets is able to successfully nowcast, i.e. estimate Dengue in the present week, and forecast, i.e. predict Dengue until 8 weeks in the future, both at country and city level with good estimation capacity. Our model can be applied successfully to smaller and less developed cities, even though it may be influenced by the incidence of the disease, the activity of Twitter locally, and social factors, including human development index and internet access.
We compared the time series of web-based data indicating Dengue activity with real observed Dengue cases in Brazil country level (Fig 1), between September, 2012 and October, 2016. Dengue cases occurred continuously with a high weekly variation, with a minimum of 694 cases per week. The highest incidence of Dengue was observed in the months of March and April of each year, reaching 106,558 cases per week (Fig 1). Tweets, Google Trends (GT) and Wikipedia access logs showed strong and positive association with the observed Dengue cases (Fig 1A). Tweets showed high variation, with an average of 1,213 tweets per week, ranging from 125 to 6,984 (Fig 1B). Tweets presented a high positive association with Dengue cases (r = 0.87, p<0.001), especially in 2013 and 2014 (Fig 1A and 1B). In the last trimester of 2015 (October to December), there was increased tweet activity not associated with Dengue.
The relative GT index was 17.51 on average, varying from 4 to 100 (Fig 1C). GT showed the stronger linear association with Dengue cases (r = 0.92, p<0.001), compared to tweets (Fig 1A). Wikipedia logs presented the smallest linear association (r = 0.71, p<0,01) with Dengue cases, but could also be considered high (Fig 1A). The values started at 517, and achieved 35,250 logs per week, with mean value of 6,481 (Fig 1D). Unfortunately, Wikipedia data was only available until December, 2015. There is important Wikipedia activity during non-epidemic periods not associated with real Dengue cases.
Tweets with Dengue content were used to estimate weekly Dengue cases occurrence at country level, in Brazil (Fig 2, Table 2). Our selected model (Table 2) has tweets as covariate, as well as a temporal structure to account for the seasonality and annual cyclic characteristics of this disease. We can observe that the model with tweets plus a temporal structure presented a better Dengue estimation capacity than a model with either variables alone (Table 2, S2 Fig). We compared the selected tweets model with models including also “Dengue cases” as covariate. Three weeks is the usual time period for data from Dengue cases to become available [19], therefore we decided to include Dengue with three weeks lag (t-3) as explanatory variable for Dengue from week t. The latter model presented the best fit to observed data, with high explained deviance, low AIC, and reduced mean relative error (Table 2, S2 Fig). Otherwise, here the model with tweets and temporal structure was selected for further analyses, because it has the estimate capacity very similar to the model with Dengue as covariate, but is easy to apply and is also useful at city level (Table 2).
Our selected model indicates that tweets are a positive predictor for Dengue cases (Fig 2A), with an almost linear effect until 2,000 tweets, that stabilizes above this value. As expected, the relationship between Dengue and tweets is influenced by the week of the year (Fig 2B), since disease transmission is highly seasonal (Fig 1). Estimated Dengue cases showed a good fit to the observed data (Fig 2C), presenting a mean relative error of 0.345, and 93.7% of deviance explained by the model (Table 2). In order to validate the estimation capacity of our model, we applied out-of-sample analysis with tweets data not previously used by our model for adjustment. Our model could successfully estimate Dengue cases in this scenario, with the capacity for explaining the deviance of Dengue of 93,2% (Fig 2C).
Dengue forecasting, i.e. the prediction of the number of Dengue cases occurring in future weeks (up to 8 weeks), was also investigated (Fig 3, Table 3). The quality of the forecast varies with the week of prediction, as we can observe by the deviance explained index and the mean relative error of the prediction in relation to observed cases (Table 3). We also showed that tweets are performing better in estimating Dengue cases in the present week, “nowcast”, since people may tweet about the disease during its occurrence. Forecasting was possible with an increasing error with the increase in forecast weeks, but good approximation to real disease occurrence, as indicated by fitted and observed lines in the time series for four different epidemic years (Fig 3, Table 3).
Tweets were obtained from 283 different cities distributed all over Brazil, including all 5 regions and 26 states (Fig 4, S1 Table). We observe that cities with higher Twitter activity are mostly clustered at the southeastern region of the country (Fig 4A). These cities overlap with the region with the highest incidence of Dengue cases (Fig 4B).
We also analyzed the contribution of tweets to estimate Dengue at city level. For the majority of cities, we observed a high positive linear association between Dengue cases and tweets, with 67% of them with association above 50% (S1 Table). Our tweets model (Table 2) could successfully fit and estimate Dengue cases in 199 cities of a total of 283, since some cities had too few data for model estimate convergence. Model goodness-of-fit was high for most cities, with Dengue deviance explained above 60% in 88% of cities analyzed (Table 4, S1 Table). Cities with high Dengue estimation quality of our model are distributed around the country but mostly concentrated at the southeastern region (Fig 5).
We selected cities in different regions of the country to further investigate and validate the model application for Dengue estimation at city level (Fig 6, S1 Table). We selected the following cities: Belo Horizonte (Fig 6A), Fortaleza (Fig 6B), Manaus (Fig 6C), Porto Alegre (Fig 6D), Rio de Janeiro (Fig 6E), and São Paulo (Fig 6F). In all cities, tweets successfully estimated Dengue cases, as shown by the approximation of observed Dengue cases and its predicted values by the model, and by the high values of deviance explained (ranging from 76.1% to 90.3%) (Fig 6, S1 Table). The model was also able to fit Dengue cases in cities with lower linear correlation indexes, as Fortaleza and Sao Paulo (Fig 6B and 6F, S1 Table). Important to notice that here we evaluated the Dengue estimation capacity of the same tweets model applied to country level, but each city would have improved results with models considering specific characteristics of each individual city.
Tweets with Dengue content and their association with Dengue cases may be influenced by different factors. We divided the 283 cities into two groups, according to the quality of their Dengue estimation by the model: high quality group included cities (161) with model explained deviance equal or higher than 60%, and low quality group cities (122) with model explained deviance smaller than 60%, or zero (model did not converge). Cities with high quality of Dengue estimation based on the tweets model have a higher population, more Dengue cases and tweets activity (Table 5). They also had higher human development indices: mean (IDHM), education (IDHME) and income (IDHMI). Only longevity index (IDML) was not different between groups (Table 5). As expected, cities with good fit by the model were those with high coverage of houses with access to a personal computer and internet (Table 5).
Otherwise, considering a linear regression association between the variables analyzed here (Table 5) and the Dengue estimate explained deviance by the model, we can observe that Dengue estimation capacity of tweets is strongly associated with Dengue incidence, but are not or weakly associated with population and development indexes (Fig 7).
In this study, we analyzed the potential of Twitter data for estimating and forecasting Dengue cases. Here we show that tweets are strongly associated with Dengue cases, and contribute not only for estimating, but also for forecasting Dengue activity up to 8 weeks in the future.
Tweets, Google Trends and Wikipedia access logs with Dengue content show a strong and positive association with officially registered Dengue cases in Brazil. However, during the last trimester of 2015, there was an important increase of tweets activity that was not associated with Dengue cases. This increase may be associated with an increase in Dengue tweeting activity that may have been caused by media news and the onset of the Zika epidemic in the country. The Zika virus is transmitted by the same vector as Dengue, the mosquito Aedes aegypti, and the disease's first symptoms are very similar, but serious complications include Guillain-Barré syndrome, and congenital infections can occur which may lead to microcephaly and maculopathy [33]. Zika, which was widely spread in the Pacific islands, was introduced in Brazil in 2014–2015 and caused a widespread epidemic in Latin America [33,34]. Google Trends, similar to Twitter, increased at the last trimester of 2015, indicating probably higher public concern with both diseases.
Twitter is a real-time source of information on Dengue symptoms activity in a population, and was shown to have strong correlation with the number of notified cases [21], however, that association may be stronger during the increasing and decreasing phases, than during the disease peaks. Twitter, as a social network, may indicate the need for the Dengue patient to notify the disease to colleagues, therefore being a good estimator of disease occurrence. Otherwise, the other web-based data available for Dengue and evaluated here, GT and Wikipedia, are based on search queries, which would indicate a potential interest or curiosity over the disease, being more subjected to marketing campaigns and confusion with other diseases.
Models built on the fraction of Google search volume for Dengue-related queries were previously shown to adequately estimate true Dengue activity in different seasons [8, 22]. Here we confirm the high association between Google Trends and Dengue disease, also useful for disease surveillance and prevention. Wikipedia data suffer from a variety of instabilities that need to be understood and compensated for [2]. Language as a location proxy can only be used in some cases, since it is impossible to be used at finer scale, or even to indicate exactly the country, an important limitation. Overall, our feeling is that all three sources of data are probably useful to estimate Dengue at country-wide level. However, amongst these three web-based data, tweets with personal experience provide a strong association with real disease with potential to be an important explanatory variable for Dengue estimation models both in country and city level.
The epidemiology of Dengue fever is highly seasonal, with multi-annual fluctuations, caused by the irregular circulation of its four serotypes, and the interplay between environmental drivers [35,36]. We built a simple model based on tweets together with a temporal structure that could successfully be used to estimate Dengue activity at country level, with 93.7% of explained deviance. The capacity of tweets to nowcast, i.e. predict the present events as they occur, may be already enough to provide a time advantage to understanding Dengue situation moment. Twitter was also useful in similar way for tracking and forecasting behavior in the influenza-like illness, as a measure of public interest or concern [11]. Dengue forecast was also possible using the model with tweets as covariate, with up to 8 weeks or 2 months of forecasting window. This result suggests that Twitter data can be used in the development of a proactive surveillance program and help health managers to better directed their resources for disease prevention.
One advantage of Twitter is that it can be geolocated at city level, which is a useful spatial resolution for surveillance. This feature strongly differentiates it from other available web-based data, such as Google Trends and Wikipedia [3,9]. While GT are available per state [22], the Wikipedia logs can only be aggregated per language [2]. Cities with higher tweets activity are those with higher Dengue incidence. Both Dengue occurrence and Twitter use are usually associated with cities with higher concentration of population or urbanization [3,14]. Similarly, GT and Dengue cases correlate better in states with higher Dengue incidence [22].
The Twitter data has also some limitations to be considered. Not everyone who submits a tweet with Dengue content is actually ill, but just interested or curious about it. Good surveillance will depend on a sufficient volume of interest to generate signals and compensate noise [8]. Therefore, a main challenge remains at areas with smaller population of Twitter users [4]. The tweets model performed better in areas with high Dengue incidence, but its performance was only weakly associated with population size and development index. This may suggest a robust model that can successfully be applied to smaller and less developed cities, which would improve the application effectiveness of the model as a surveillance tool. One advantage of including tweets into forecast models is to improve real-time estimations of Dengue incidence, overcoming difficulties of traditional Dengue surveillance systems that rely solely on case report data. Twitter captures information from individuals, especially at earlier stages of illness, who may search health information on the internet before or even instead of making medical visits, and publish this knowledge to seek help and comfort from friends. Tweets-based models may actually be even more useful in endemic regions of the world where the traditional surveillance system is too weak and slow to react to disease notification.
Here we show that the high Dengue estimation capacity of tweets model is influenced by human development indices and internet access. Important to observe that mean, education and income development indices which is associated with more houses with access to a personal computer and internet are also associated with tweets incidence. Otherwise, longevity development index is less associated with tweets incidence and activity, suggesting that young and adults may be the majority of users of this data. The accuracy of Google Trends was not found to be strongly influenced by socio-economic factors, particularly because it relies on internet searches, which may be robust enough to capture population-level disease dynamics [22]. Despite these social limitations, it is clear that tweets-based surveillance provides adequate citywide and countrywide Dengue estimates. Social factors, however, may limit the value of using tweets to examine epidemics within a city. At this stage, freely available tweet data are not sufficient to provide accurate determination of space within a specific city.
The capacity of tweets to estimate Dengue cases represents a valuable complement to assist traditional Dengue surveillance. A novel data source, like Twitter, could complement traditional surveillance at low-cost, since it is passive, free, and requires minimal resources to run [3,11]. These data can help reduce some of the many gaps that exist in Dengue surveillance methods, such as low sensitivity and accuracy, and timeliness [13,14,19]. Improving Dengue surveillance in a cost-effective way remains a major obstacle. In Brazil, the underreporting is about 50%, but can reach values as high as 90%, and the reporting delay is estimated to be approximately 3 to 4 weeks [19]. The main added benefit in monitoring social media behavior through tweets is the potential for early warning. Detecting and confirming results of prevention and control measures is possible at the interface between computer science, epidemiology, and medicine [4]. Our study therefore demonstrates that tweets are a web-based data that strongly associate with Dengue cases and have the potential to successfully estimate Dengue cases. Tweets are an easy to use, cost-effectiveness, useful and robust tool for estimating Dengue cases, both at country and city level, and for Dengue forecasting until 8 weeks in the future.
|
10.1371/journal.pgen.1000171 | WDR55 Is a Nucleolar Modulator of Ribosomal RNA Synthesis, Cell Cycle Progression, and Teleost Organ Development | The thymus is a vertebrate-specific organ where T lymphocytes are generated. Genetic programs that lead to thymus development are incompletely understood. We previously screened ethylnitrosourea-induced medaka mutants for recessive defects in thymus development. Here we report that one of those mutants is caused by a missense mutation in a gene encoding the previously uncharacterized protein WDR55 carrying the tryptophan-aspartate-repeat motif. We find that WDR55 is a novel nucleolar protein involved in the production of ribosomal RNA (rRNA). Defects in WDR55 cause aberrant accumulation of rRNA intermediates and cell cycle arrest. A mutation in WDR55 in zebrafish also leads to analogous defects in thymus development, whereas WDR55-null mice are lethal before implantation. These results indicate that WDR55 is a nuclear modulator of rRNA synthesis, cell cycle progression, and embryonic organogenesis including teleost thymus development.
| Medaka, Oryzias latipes, is a small freshwater fish that is useful for studies of forward and reverse genetics. The availability of various inbred strains is the distinct advantage of medaka over zebrafish, especially in studies of the immune system. The thymus is a primary immune organ that is unique to vertebrates and supports the generation of T lymphocytes. Defective thymus development tends to cause abnormal T lymphocyte development, resulting in immunodeficiency or autoimmunity. However, the molecular pathways underlying thymus development have not been fully uncovered. Here we report the positional cloning of a gene responsible for one of the thymus-defective medaka mutants. We find that the hkc phenotype is caused by a missense mutation in a gene encoding the previously uncharacterized protein WDR55. Our results indicate that WDR55 is a novel nucleolar modulator of rRNA biosynthesis, cell cycle progression, and vertebrate development of organs, including teleost thymus. These results not only provide evidence of the existence of a new mechanism of rRNA production but also demonstrate that the malfunction of WDR55 causes cell cycle arrest and developmental failure, including defective thymus development.
| The thymus is a lymphopoietic organ that is unique to vertebrates and supports the generation of T lymphocytes. It is generated from the budding of third pharyngeal pouch endoderm and its interaction with ventrally migrating neural crest cells [1],[2]. Lymphoid precursor cells derived from hematopoietic stem cells immigrate to thymus primordium where they differentiate into mature T lymphocytes carrying diverse yet self-tolerant recognition repertoire [3],[4]. Defective thymus development tends to cause abnormal T lymphocyte development, resulting in immunodeficiency or autoimmunity [5]–[8]. Studies of patients and animal models have enabled identification of several genes required for thymus development. Tbx1 is the gene responsible for DiGeorge syndrome, a condition characterized by cardiovascular, thymic, parathyroid, and craniofacial anomalies [9]–[11]. Foxn1 is the gene responsible for severe immunodeficiency of nude phenotype in mouse and human, due to the lack of functional thymus and hair formation [12],[13]. Use of genetically modified mouse strains has enabled further identification of genes involved in thymus development [14],[1]. However, the molecular pathways underlying thymus development have not been fully uncovered.
We previously established a collection of ethylnitrosourea-induced medaka mutants that exhibited recessive defects in thymus organogenesis [15],[16]. Medaka, Oryzias latipes, is a small freshwater fish that is useful for studies of forward and reverse genetics [17]. Like zebrafish Danio rerio, medaka is one of the smallest vertebrate species equipped with an adaptive immune system that includes the thymus, T lymphocytes, and T-cell-mediated cellular immune responses, such as allograft rejection [18],[19]. The small size of the genome (800 Mb in medaka vs. 1700 Mb in zebrafish), along with the availability of various genomic resources, including a completed genome sequence, bacterial artificial chromosome library, and radiation hybrid maps, makes medaka a useful species for genomic analysis and genetic experiments, including transgenesis and morpholino antisense oligonucleotide-mediated gene knockdown [20]–[23]. The availability of various inbred strains is the distinct advantage of medaka over zebrafish [17], especially in studies of the immune system, such as the development and function of T lymphocytes. By screening ethylnitrosourea-induced mutants that covered approximately 60% of medaka genome, we established 22 mutant lines that have defects in immature-lymphocyte-specific recombination activating gene 1 (rag1) expression in the thymus. These medaka mutants would complement the panel of mutations affecting thymus organogenesis in zebrafish [24]–[26], since different spectrum of mutant phenotypes has been identified in medaka from that in zebrafish due to divergent functional overlap of related genes [16].
We report herein the positional cloning of a gene responsible for one of the thymus-defective medaka mutants, hokecha (hkc), in which thymus primordium fails to accumulate lymphoid cells. We find that the hkc phenotype is caused by a missense mutation in a gene encoding previously uncharacterized protein WDR55 that carries the tryptophan-aspartate-repeat motif. We show that WDR55 modulates the nucleolar production of ribosomal RNA (rRNA) and hkc mutation causes a defect in the nucleolar localization of WDR55. The defect in WDR55 causes the accumulation of aberrant rRNA intermediates and cell cycle arrest. We also show that WDR55 mutation in zebrafish causes defective development of the thymus. Thus, the present results indicate that WDR55 is a novel nucleolar modulator of rRNA synthesis, cell cycle progression, and embryonic organogenesis, including teleost thymus development.
We previously established a medaka strain, hokecha (hkc), in which rag1 expression in the thymus was undetectable [15]. T lymphocyte development in embryonic thymus of wild-type (WT) medaka could be visualized by whole-mount in situ hybridization of immature-lymphocyte-specific rag1, lymphocyte-specific ikaros, and T-lymphocyte-specific T-cell receptor beta (tcrβ), whereas none of these genes were detectable in the thymus of hkc mutants (Figure 1A). Unlike the thymus of wild-type medaka, accumulation of hematoxylin-rich lymphoid cells was not detectable at the pharyngeal region in hkc mutants (shown below). Systemic T lymphocytes were also undetectable in hkc by T-lymphocyte-specific genes tcrβ, cd4, and lck in whole embryos (Figure 1B).
T lymphocyte development in the thymus is initiated upon the migration of lymphoid precursor cells into thymus primordium [3]. To examine whether the development of lymphoid precursor cells is affected in hkc, we analyzed early hematopoiesis and pre-thymic lymphopoiesis in hkc embryos. Early hematopoiesis in the lateral mesoderm [27], which was detected by measuring erythrocyte-specific gata1 and lymphocyte-specific ikaros expression, was unaltered in hkc mutant embryos at stage 21 (34 hours post fertilization (hpf)) (Figure 1C). In the course of the embryogenesis, hematopoiesis is relocated to intermediate cell mass and ventral wall of dorsal aorta, which is considered to correspond to the aorta-gonad-mesonephros region in mammals [28]-[30]. ikaros and gata1 expression in the intermediate cell mass remained unaltered in hkc embryos at stage 23 (41 hpf) (data not shown). In addition, normally shaped red blood cells were generated in the circulation of hkc mutants (Figure 1C). These results indicate that hematopoiesis and early lymphopoiesis are detectable in hkc.
We next examined whether the development of thymus primordium might be affected in hkc. To this end, we transplanted EGFP-expressing immature lymphocytes into hkc embryos. EGFP-expressing immature lymphoid cells isolated from transgenic medaka expressing EGFP under the control of medaka rag1 promoter [30] were injected into wild-type or hkc embryos via blood vessel and traced under a fluorescence microscope. We detected the migration of EGFP+ cells in the thymus of wild-type embryos and the remarkably reduced accumulation of EGFP+ cells in the thymus of hkc embryos (Figure 1D). These results indicate that the development of thymus primordium that attracts lymphoid precursor cells is defective in hkc. Nonetheless, it is possible that lymphoid precursor cells in hkc are also defective in colonization and/or development in the thymus.
Thymus primordium is generated through the interaction of third pharyngeal pouch endodermal cells with neural-crest-derived mesenchymal cells [1],[2]. The expression of pax9 and dlx2 that detect endodermal cells and neural-crest-derived cells, respectively, in pharyngeal pouch was slightly distorted but comparably detected in hkc embryos (Figure 1E). Moreover, the expression of thymic epithelial cell specific foxn1 was detectable in hkc mutants (Figure 1E). On the other hand, pharyngeal arches in hkc were short and abnormally shaped (Figure 1E). We detected the thin seventh pharyngeal arch that we previously failed to detect [15] in hkc (Figure 1E).
These results indicate that hkc mutant medaka is unable to develop functional thymus primordium that is colonized by lymphoid precursor cells.
Positional cloning was carried out to identify the mutation responsible for the hkc phenotype. Linkage analysis mapped hkc gene within the 23 kb region on scaffold 567 (covered by a single BAC clone Md0218G12) of linkage group (LG) 18 (Figure 2A). According to gene prediction by Genscan, this region contained two genes, WDR55 (EST clone MF01SSB013N12) and an unnamed transcript that contained presumptive 303 bp coding region (Figure 2A). We found that hkc allele carried a guanine (G) to adenine (A) point mutation in the coding region of WDR55, whereas no mutations were found in the 303 bp transcript (Figure 2B). The predicted open reading frame of WDR55 encodes a 400 amino acid protein containing six tryptophan-aspartate-repeat (WDR) motifs. The deduced amino acid sequence was 58%, 59%, and 66% identical to human, mouse, and zebrafish WDR55, respectively (Figure 2C). BLAST search found no other WDR55-like loci in the genome of medaka, zebrafish, mouse, and human (data not shown).
The WDR motif is shared among various proteins involved in signal transduction, cell cycle control, and transcriptional regulation [31]. The point mutation in hkc caused a glycine to arginine substitution at the 112th amino acid residue that was projected from β sheets in the second WDR motif, according to structural prediction by Smith et al. (1999) [32] (Figures 2B, 2C, and 2D). Injection of wild-type WDR55 mRNA into homozygous hkc embryos rescued hkc phenotypes including defective rag1 expression in the thymus, whereas injection of the same amount of hkc WDR55 mRNA failed to rescue hkc phenotypes (Figure 2E). In contrast, more than half of wild-type medaka embryos that were administered morpholino antisense oligonucleotide to block the splicing of WDR55 and reduce spliced WDR55 mRNA (Figure 2F) phenocopied defective thymus development and small eye size found in hkc mutants (Figure 2F). Another WDR55-specific morpholino that was designed to hybridize a start-codon-containing sequence and to block translation of WDR55 mRNA also caused defective thymus development in wild-type medaka (Figure 2G). These results indicate that the G to A point mutation in WDR55 gene is responsible for the thymus-defective phenotype of hkc mutants.
Because the function of WDR55 was previously unknown, we first examined the intracellular localization of medaka WDR55 tagged with EGFP and expressed in human 293T cells. We found that EGFP fused with wild-type WDR55 was chiefly condensed in the nucleolus, as determined by the merged localization with co-transfected t-HcRed1-fibrillarin (Figure 3A). In contrast, EGFP fused with hkc-mutant WDR55 was excluded from the nucleolus (Figure 3A). Similar results showing the nucleolar accumulation of EGFP fused with wild-type WDR55 and the exclusion of EGFP fused with hkc-mutant WDR55 from the nucleolus were obtained upon transfection into mouse NIH3T3 cells (data not shown). Medaka somatic cells expressing EGFP fused with WDR55, but not hkc-mutant WDR55, showed intra-nuclear dot-like localization of EGFP (Figure 3A). Antibody detection of mouse WDR55 in NIH3T3 cells further showed that endogenous WDR55 was detectable in the nucleolus as well as the cytoplasm (Figure 3B). Endogenous WDR55 detected in the nucleolus co-localized with fibrillarin, which was enriched in the dense fibrillar component of the nucleolus [33], rather than with B23 nucleophosmin, which was enriched in the peripheral granular component of the nucleolus (Figure 3B). These results indicate that WDR55 is a nucleolar protein enriched in the dense fibrillar component and hkc mutation of WDR55 perturbs its nucleolar localization.
The dense fibrillar component in the nucleolus is the place where early processes of ribosome biosynthesis take place [34]. In order to examine the possible involvement of WDR55 in ribosome biosynthesis, NIH3T3 cells were transfected with siRNA specific for mouse WDR55 and examined for rRNA processing. siRNA transfection specifically reduced WDR55 expression in NIH3T3 cells (Figure 3C). We found that incompletely processed rRNA precursors as detected by hybridization with a 5.8S rRNA probe more strongly accumulated in WDR55-siRNA-transfected NIH3T3 cells than in control-siRNA-transfected or untransfected NIH3T3 cells (asterisks in Figure 3D). However, mature 5.8S, 18S, and 28S rRNAs were produced in WDR55-siRNA-transfected cells (Figure 3D). There were no significant differences in overall ribosome profiles between control and WDR55-siRNA-transfected NIH3T3 cells (data not shown). These results indicate that incompletely processed rRNA intermediates are accumulated by defective WDR55 expression.
It was previously shown that similar to WDR55, such WDR-motif-carrying proteins as WDR12 and Bop1 are localized in the nucleolus where they regulate rRNA processing [35]. However, unlike WDR12 and Pes1, WDR55 was not co-immunoprecipitated with WDR12, Bop1, or Pes1 (Figure 3E), suggesting that WDR55 is a novel modulator of rRNA production without physical association with the PeBoW complex that contains Pes1, Bop1, and WDR12.
At 2 days after WDR55-siRNA transfection into NIH3T3 cells, the expression of p21Waf1/Cip1, a gene that is regulated by p53 and controls cell cycle [36],[37], was significantly increased, while that of WDR55 mRNA was significantly decreased compared to those in control-siRNA-transfected cells (Figure 3F). Accordingly, the frequency of cells in S phase was markedly decreased in WDR55-siRNA-transfected cells (Figure 3G), indicating that the defective expression of WDR55 results in cell cycle arrest at G1 phase. Two other WDR55-siRNAs that reduced WDR55 mRNA expression less strongly than that used in Figures 3C, D, F, G increased p21 expression and decreased the number of S-phase cells less strongly (data not shown). These results demonstrate that WDR55 modulates nucleolar rRNA production and the defective expression of WDR55 affects p53 activation and cell cycle progression.
Northern blot hybridization of total RNA from 7 days post-fertilization (dpf) whole medaka embryos showed that incompletely processed rRNA intermediates accumulated in hkc mutants but not their siblings or wild-type medaka embryos (Figure 4A–F). Hybridization using probe C that was designed within the 5.8S rRNA sequence showed that rRNA processing intermediates ‘a’, ‘b’, and ‘c’ were more strongly detectable in hkc mutants than in siblings or wild-type medaka embryos (Figure 4A). Intermediates ‘a’ and ‘b’ but not ‘c’ accumulated in hkc mutants were detectable by probe B that was designed within the internal transcribed spacer 1 (ITS1) sequence between 18S and 5.8S rRNA sequences (Figure 4B), suggesting that intermediates ‘a’ and ‘b’ presumably corresponded to ITS1-containing intermediates with and without 18S rRNA, respectively, and intermediate ‘c’ presumably contained 5.8S rRNA, ITS2, and 28S rRNA sequences without ITS1 sequence (Figure 4C). Indeed, probe D that was designed within the ITS2 sequence between 5.8S and 28S rRNA sequences visualized the accumulation of intermediates ‘a’, ‘b’, and ‘c’ (Figure 4D). Nonetheless, hkc mutants produced mature 5.8S, 18S, and 28S rRNAs (probes C, A, and E, respectively; Figures 4D and E). These results indicate that similar to WDR55-siRNA-transfected NIH3T3 cells, hkc mutation of WDR55 affects rRNA processing and induces the accumulation of incompletely processed rRNA intermediates in vivo.
The expression of p53 and p21Waf1/Cip1 in whole embryos at 7 dpf was significantly higher in hkc mutants than in wild-type medaka (Figure 4F), suggesting that WDR55 mutation in hkc causes p53 activation and results in developmental defects. However, we did not detect restoration of rag1-expressing cells in the thymus or normal-sized eyes at 7 dpf in hkc mutants that also lacked functional p53 by Y186X truncation [38] (Figure 4G). It was previously suggested that the expression of p53 family molecules, such as p63 and p73, could be elevated in the absence of p53 and the elevated p63 and/or p73 might compensate for the loss of p53 [39],[40]. Indeed, we found that p53Y186X/Y186X mutant medaka embryos showed significantly elevated expression of p63 and p73 (Figure 4H). Thus, it was possible that defective thymus development in WDR55hkc/hkc p53Y186X/Y186X double mutants might be signaled via p63 and/or p73. These results suggest that hkc mutation causes the accumulation of incompletely processed rRNA intermediates and activates p53 family molecules.
We then addressed how hkc mutation resulted in defects in thymus development rather than systemic failure of the development at much earlier stages. To do so, we initially examined WDR55 expression in adult and embryonic medaka. We found that the expression of WDR55 was detectable in every organ, including the thymus, of adult medaka by quantitative RT-PCR analysis (Figure 5A). Among adult medaka tissues, WDR55 expression was most prominent in reproductive organs, such as testis and ovary (Figure 5A). During embryogenesis, WDR55 expression was stronger in early embryonic stages than in late ones (Figure 5B) and widespread in embryonic body, as detected by whole-mount in situ hybridization (Figure 5C). These results indicate that WDR55 is expressed ubiquitously in medaka and is not specific to the thymus, suggesting that the defects in hkc might not be limited to the thymus.
In fact, we found that in addition to the thymus, the spleen was absent in 6-dpf and 7-dpf hkc mutants (Figure 5D). Similar to the head and the eyes, the liver and the gall bladder were smaller in hkc mutants than in wild-type medaka (Figure 5D). Pharyngeal arches were abnormally shaped and lower jaws were malformed in hkc mutants (Figure 1E). On the other hand, hematopoiesis and development of the gills and the gut, as well as body axis formation, were not disturbed at 7 dpf. Nevertheless, hkc mutants were lethal between 8 and 10 dpf. These results indicate that hkc mutation of WDR55 causes systemic and lethal failure of medaka development by 10 dpf, and several organs including the thymus are more severely affected than other organs during embryogenesis of hkc mutants before lethality.
Then, we examined how the development in hkc mutants was not arrested at much earlier stages but could be sustained until 10 dpf. It is known that zygotic transcription in medaka begins around mid-blastula transition at 8.25 hours post-fertilization (hpf) or stage 11 [41], and early development at least before this transition is regulated by maternally inherited mRNA and proteins [42]. Indeed, embryonic transcription of WDR55, as revealed by measurement of paternal allele-specific mRNA in heterozygous embryos, was first detectable in early blastula at 6.5 hpf or stage 10 (Figure 5E), whereas WDR55 mRNA specific for maternally inherited allele was detectable until late blastula at stage 11 (Figure 5F). On the other hand, measurement of the decay of WDR55-EGFP fusion protein in cycloheximide-treated 293T cells indicated that the half-life of WDR55-EGFP fusion protein was substantially longer than 3 days and estimated to be 7 to 10 days in the cells (Figure 5G). These results suggest that maternally inherited WDR55 mRNA and proteins are present in medaka embryos and normal WDR55 proteins derived from female parents may support the embryogenesis of hkc mutants for up to 10 days.
We further wanted to address how hkc embryos exhibited the defects in several organs such as the thymus more severely than in other organs. To do so, we examined the status of cell proliferation in medaka embryonic tissues. Many large proliferating cells detected by bromodeoxyuridine (BrdU) labeling were found in the thymus of wild-type medaka at 7 dpf (Figure 5H). Many cells in the eyes were also proliferating in 3 dpf medaka embryos (Figure 5H). However, these proliferating cells in the thymus or the eyes were barely detectable in stage-matched hkc mutants (Figure 5H). In addition to the thymus and the eyes, we detected active proliferation in the intestine and the gills, as visualized by BrdU incorporation (data not shown). Together, these results suggest that cells generating several organs including the thymus and the eyes undergo massive cell cycle progression, so that the cells that form these organs of hkc mutant embryos would rapidly consume maternally inherited normal WDR55 proteins, thereby exhibiting severe defects in organogenesis. Nonetheless, it is unclear whether all of the tissues affected in hkc mutants exhibit rapid proliferation.
Zebrafish strain hi2786B was previously established as one of the lethal mutants caused by random retroviral insertion, and the retrovirus in hi2786B allele was found to be inserted within zebrafish WDR55 locus [43]. We found that the retrovirus in hi2786B allele was inserted in the WDR55 coding region before the first WDR motif (Figure 6A) and the 3′ region of WDR55 open reading frame was not transcribed in hi2786B embryos (Figure 6B), suggesting that hi2786B is a WDR55-null mutant. Until lethality around 10 to 12 dpf, overall body formation including head and tail appeared intact in hi2786B mutants (Figure 6C). However, similar to hkc mutants in medaka, thymus size was remarkably reduced and the pharyngeal arches were malformed in hi2786B mutants (Figure 6C). Also, similar to hkc mutants, the eyes of hi2786B mutants were small (Figure 6C). Swim bladder in hi2786B mutants was small as well (Figure 6C). These results indicate that WDR55 deficiency in hkc medaka and hi2786B zebrafish results in similar defects in development, including defective formation of the thymus.
WDR55 in mice was also expressed systemically, and WDR55-deficient mice previously established by targeted mutation were found to die before E9.5 [44]. We examined earlier development of WDR55-deficient mice and found that WDR55-deficient embryos disappeared as early as E3.5 (Figure 7A), indicating that WDR55 deficiency causes early arrest of mouse development before implantation. WDR55+/− progenies were derived from either mating WDR55+/− females with wild-type males or mating wild-type females with WDR55+/− males, although the number of heterozygote progenies obtained from these crosses was reduced (Figure 7B). Thus, spermatogenesis and oogenesis were not severely arrested without WDR55. Haploinsufficiency in WDR55+/− mice did not cause defects in development including thymus formation during embryogenesis (Figure 7C). These results indicate that WDR55 deficiency in mouse causes developmental arrest before implantation.
The present study revealed that a medaka hkc mutation that affects the development of functional thymus primordium and causes lethality by 10 dpf is due to a missense point mutation in the gene encoding WDR55, a novel protein. We found that WDR55 is a nucleolar protein ubiquitously expressed in various organs, and hkc mutation of WDR55 perturbs its nucleolar localization and causes the accumulation of incompletely processed rRNA intermediates. We also found that the defective expression of WDR55 affects cell cycle progression and hkc mutation activates the p53 pathway. A zebrafish WDR55 mutant showed similar developmental defects, including the lack of thymus formation and lethality by 12 dpf, whereas WDR55 deficiency in mouse caused much earlier developmental arrest before implantation. These results indicate that WDR55 modulates rRNA production, cell cycle progression, and vertebrate development, including teleost thymus organogenesis.
Recombinant gene mapping and DNA sequencing as well as mRNA-mediated rescue of hkc embryos and morpholino antisense oligonucleotide mediated phenocopy in wild-type embryos identified a point mutation in WDR55 that is responsible for defective thymus formation in hkc mutants. WDR55 is a member of a large family of proteins that contain 4–16 tryptophan-aspartate-repeat (WDR) motifs, which consist of 40-60 amino acids with glycine-histidine dipeptide at the 11th to 24th residues from the N terminus and tryptophan-aspartate dipeptide at the C terminus [31]. Crystal structure analysis of G-protein β subunit, a WDR-motif-containing protein, has indicated that a WDR motif forms one blade of β propeller [31],[32],[45]. The WDR motif is implicated in protein-protein interaction, and WDR-motif-containing proteins have a variety of functions, including signal transduction, cell cycle regulation, and RNA synthesis [31].
The present study describes the function of a previously uncharacterized WDR-motif-containing protein, WDR55. Fluorescence localization of EGFP fusion protein containing medaka WDR55 sequence as well as antibody detection of endogenously expressed mouse WDR55 in the cells indicates that WDR55 proteins are localized in the nucleolus and the cytoplasm. Interestingly, hkc mutation that causes arginine substitution of glycine, which is presumably localized in the protruded loop between the second and third β sheets in the second propeller [31], causes the exclusion of WDR55 from the nucleolus. It is unclear whether this mutation affects nucleolar localization of WDR55 by specifically altering the capability of WDR55 for nucleolar localization or disrupting the overall structure of the protein. Nevertheless, aberrant localization of WDR55 by hkc mutation suggests that nucleolar localization is pivotal for WDR55 to exert its function. Within the nucleolus, WDR55 is enriched in the dense fibrillar component where early processes of rRNA biosynthesis take place [34].
Our results show that both hkc mutation of medaka WDR55 in vivo and siRNA-mediated reduction of mouse WDR55 expression in cell culture affect rRNA processing in the nucleolus. WDR55 reduction does not severely impair the production of 18S, 5.8S, and 28S rRNAs; rather it induces excessive accumulation of rRNA processing intermediates. Previous studies have shown that many proteins are involved in rRNA processing [34]. Among them, several proteins containing WDR motifs are known to be involved in rRNA synthesis in the nucleolus. For example, Rsa4p is involved in rRNA processing and transport of large ribosomal subunits [46], and WDR12 associated with another WDR-motif-containing protein, Bop1, along with Pes1 is involved in the processing of 5.8S/28S rRNA [35]. It is interesting to note the similarity of WDR55 to this PeBoW complex consisting of Pes1, Bop1, and WDR12 in terms of structures sharing WDR motifs and functions in 5.8S/28S rRNA processing. However, the results of co-immunoprecipitation experiments do not support the possibility that WDR55 is an additional member of the PeBoW complex. Thus, WDR55 is a novel modulator of rRNA synthesis in the nucleolus. It is unclear whether WDR55 is involved directly in rRNA processing, indirectly in the clearance of rRNA processing intermediates, or in rRNA processing/synthesis via other mechanisms. Nonetheless, it is possible that WDR55 may participate in linking rRNA production to cell cycle regulation.
We found that siRNA-mediated reduction of WDR55 expression in a mouse cell line causes p53 activation and cell cycle arrest, and that hkc mutation of medaka WDR55 in vivo analogously causes p53 activation and developmental defects. It is well known that nucleolar accumulation of rRNA intermediates activates the p53 pathway [47]–[49]. Accordingly, dominant-negative WDR12 accumulates rRNA processing intermediates and arrests cell cycle through p53 activation [34]. Also, a zebrafish pes1 mutant exhibits small eyes and developmental failure in multiple organs [50]. Perhaps through a similar mechanism of nucleolar stress, either siRNA-mediated reduction or hkc mutation of WDR55 accumulates rRNA intermediates that may activate p53 family molecules and thereby lead to cell cycle arrest.
It is unclear whether developmental arrest in hkc mutants is indeed mediated by p53 activation, since our results show that p53 deficiency does not rescue developmental defects in hkc mutants. Since we detected significant increases in the expression of p63 and p73 in p53-deficient mutant medaka, it is possible that the elevated p63 and/or p73 may compensate the loss of p53 in p53-deficient medaka. Thus, it is possible that nucleolar stress in hkc mutants activates p53 family molecules that relay signals for cell cycle arrest and developmental defects.
WDR55 is ubiquitously expressed in various organs in medaka and mouse. We showed that hkc mutation of WDR55 causes lethal failure of medaka development by 10 dpf and several organs including the thymus are severely affected during embryogenesis of hkc mutants before lethality. Our results support the possibility that maternally inherited WDR55 mRNA and proteins may support the embryogenesis of hkc mutants. Since cells generating several organs including the thymus are shown to undergo massive cell cycle progression, it is possible that the cells that form these organs of hkc mutant embryos may rapidly consume maternally inherited normal WDR55 proteins, thereby exhibiting arrest in cell cycle progression and severe defects including failure in thymus organogenesis.
However, it is also possible that WDR55 may be somehow associated with molecules that are specifically expressed in several organs including the thymus, and the defect in WDR55 may exert cell-type-specific defects. Such a possibility was suggested in zebrafish mutant one-eyed pinhead (oep), where a ubiquitously expressed permissive EGF-related ligand Nodal co-receptor causes cell-type-specific abnormalities including cyclopia [51]. Biochemical analysis of WDR55 functions, especially in regard to the mechanisms modulating rRNA biosynthesis and the cell-type-specific susceptibility to the defect in WDR55, is awaited to prove this hypothesis.
Finally, this study shows that zebrafish hi2786B mutant that carries an insertional null mutation in WDR55 locus exhibits developmental defects similar to hkc mutant in medaka, including defective formation of the thymus. The similarity of defects in medaka and zebrafish supports the possibility that the missense hkc mutation of WDR55 causes a severe deficiency of functional WDR55 similar to a null mutation, rather than retaining the partial functions of WDR55. On the other hand, a null mutation of WDR55 in mouse causes much earlier developmental arrest before implantation. The difference in developmental defects caused by WDR55 mutation between mouse and the two teleost species may be linked to the difference in the role of maternally inherited mRNA and proteins during embryogenesis. Maternal-to-zygotic transition of gene expression occurs as early as the 2-cell stage in mouse [52], whereas zygotic transcription in oviparous vertebrates occurs during blastula stage [41]. Our data further show that the half-life of WDR55 proteins synthesized in the cells is 7 to 10 days. Thus, it is possible that maternally inherited mRNA and proteins may contribute to the embryogenesis for substantially longer periods in oviparous vertebrate species including teleosts than in mammals. This difference between mouse and teleost should be carefully considered when genetic analysis using teleost species is implicated to understand mammalian development and human biology with medical goals.
In conclusion, we found that WDR55 is a novel modulator of nucleolar production of rRNA and the deficiency in WDR55 causes cell cycle arrest and developmental defects, including teleost thymus organogenesis. Further analysis of WDR55 function should lead to a better understanding of rRNA biosynthesis and vertebrate development.
Medaka cab strain and cab-derived hkc mutant line were previously described [15]. Medaka kaga strain, which has a highly discrete genome sequence compared to cab strain, was used to map hkc mutation. p53-deficient mutant medaka line p53Y186X [38] and rag1-EGFP transgenic medaka line [30] were also used. Developmental stage was designated as described [53]. Where indicated, zebrafish hi2786B mutant strain [43] and WDR55-deficient mouse strain [44] were used. Animal experiments were performed with consent from the Animal Experimentation Committee of the University of Tokushima.
To identify the linkage group of hkc mutation, bulked segregant analysis was carried out on DNA isolated from 48 hkc embryos and 48 wild-type siblings derived from an hkc/cab×kaga mapping cross. Genetic mapping and chromosome walking on linkage group 18 were performed essentially as described by Geisler (2002) [54], using restriction fragment length polymorphism markers between cab and kaga strains. 407 hkc embryos were analyzed using markers described in MLBase (http://mbase.bioweb.ne.jp/dclust/ml_base.html) and additional markers that we identified. Scaffolds of medaka shotgun sequences were searched at Medaka Genome Project server (http://dolphin.lab.nig.ac.jp/medaka/). BAC library of medaka genomic DNA was previously described (Matsuda et al., 2001) [20]. Gene structure was predicted using Genscan (http://genes.mit.edu/GENSCAN.html).
50 µM of morpholino antisense oligonucleotides in 0.3x Danieu's solution containing 0.1% rhodamine-dextran was injected into fertilized cab eggs at 1-cell stage. The sequence of the morpholino was as follows: WDR55-splicing-inhibiting morpholino, 5′-AGA CTC CGT GTT CCT GAC CTT CAG-3′; and WDR55-translation-inhibiting morpholino, 5′-CCG CCA TGT TTG TTT GGT GAT TTT C-3′. Total RNA was extracted from the embryos at stage 25, and RT-PCR was carried out to confirm inhibition of mRNA splicing. Primers used for this RT-PCR were 5′-GGG CTA AAG CTG TTT AGC GT-3′ and 5′-GCC TCT CCC TTC CTC ATG TC-3′. Morphants were fixed at 5 or 6 dpf for in situ hybridization. The same amount of control morpholino (five-nucleotide substitution from the sequence specific for an unrelated gene TC53327) [30] did not affect the phenotype.
cDNA fragments containing the entire coding region of WDR55 derived from either wild-type or hkc embryos were amplified using primers 5′-GCA GCT GTT CAG CGC AGA AG-3′ and 5′-AAC ACA ACT TTC CTG TCC AA-3′, and cloned into pCRII vector (Invitrogen). EcoRI fragments containing entire insert sequences were subcloned to pCSII+ expression vector [55]. 3′ ends of inserts were cut with NotI, and cDNA was transcribed using mMESSAGEmMACHINE Kit (Applied Biosystems). 10 ng/µl of WDR55 mRNA was injected into 1-cell-stage fertilized eggs derived from mating of hkc/+ heterozygotes. 10 ng/µl of EGFP mRNA was co-injected as internal control to confirm successful injection and translation.
Adult thymocytes (5×102) of rag1-EGFP transgenic medaka were labeled with CellTracker Orange CMTMR (Molecular Probes) and injected into sinus venosus of dechorionated embryos at 5 dpf. Thymic regions of recipients were observed under a laser scanning microscope at 1 day after the transplantation.
Probes for detecting medaka rag1, ikaros, tcrb, dlx2, pax9, gata1, and foxn1 were described previously [15],[30]. Sense and antisense probes for WDR55 were produced from WDR55 cDNA as described above and were labeled with digoxigenin using DIG RNA Labeling Kit (Roche). Plasmid containing zebrafish rag1 [56] was a kind gift from Dr. C. E. Willett. Whole-mount in situ hybridization was carried out as described [15]. Alcian blue staining of cartilage structures was carried out as described [57].
Medaka embryos were soaked in 1 mg/ml BrdU in 0.03% sea salt water for 1.5 hours. Frozen sections (10 µm) were stained with anti-BrdU antibody and hematoxylin using BrdU In-Situ Detection Kit (BD Biosciences Pharmingen). Zebrafish larvae at 6 dpf were embedded in OCT compound (Sakura Finetek) and frozen. Five-micrometer sections were stained with hematoxylin and eosin. Cryosectioning and immunohistochemical staining of mouse fetal thymus were previously described [58].
siRNA specific for mouse WDR55 or a control siRNA (Invitrogen) was transfected into NIH3T3 cells using the protocol supplied by Invitrogen. At 44 hours after transfection, cells were either harvested for biochemical analysis or pulsed with 10 µM BrdU for 30 minutes, followed by staining with anti-BrdU antibody and 7-AAD (BD Biosciences Pharmingen). Three kinds of WDR55-specific siRNAs with different sequences gave similar results of varying degrees, so that the results from only one siRNA that demonstrated the strongest effects are shown.
cDNA fragments of WDR55 containing the entire open reading frame derived from either wild-type or hkc mutants were cloned into pEGFP-C1 (BD Biosciences Pharmingen). These constructs, pEGFP-WDR55WT (wild-type) and pEGFP-WDR55MT (hkc mutant), expressed fusion proteins that attached EGFP to the N-terminus of WDR55. Plasmid containing t-HcRed1-fibrillarin [59] was kindly provided by Dr. K. A. Lukyanov. pEGFP-WDR55WT, pEGFP-WDR55MT, or pEGFP-C1 was co-transfected with t-HcRed1-fibrillarin into 293T cells. At 30 hrs after transfection, fluorescence signals were detected with TCS SP2 laser scanning microscope (Leica).
In order to obtain the antibody specific for mouse WDR55, rabbits were immunized with a protein conjugated with a synthetic peptide of mouse WDR55 (Ac-CSSGHDQRLKFWDMTQLR-amide). NIH3T3 cells fixed with 4% PFA and permeabilized with 0.1% Triton-X were incubated with 1/300 dilution of anti-WDR55 antibody and 1/300 dilution of mouse anti-fibrillarin antibody (Abcam) for 1 hr. FITC-labeled goat anti-rabbit IgG antibody (Molecular Probes) and AlexaFluor633-labeled goat anti-mouse IgG antibody (Molecular Probes) were used for fluorescence visualization of antibody binding with a laser scanning microscope.
Total RNAs were extracted with Isogen (Wako Chemical) and cDNAs were synthesized using SuperscriptIII first strand synthesis system (Invitrogen). Quantitative RT-PCR was performed with SYBR premix ExTaq (Takara) and iCycler iQ Real Time PCR System (Bio-Rad). Amplified signals were confirmed to be single bands over gel electrophoresis, and normalized to the signals of medaka cytoplasmic actin, zebrafish β-actin, or mouse GAPDH. The primers used were as follows: medaka WDR55, 5′-GAC AGA TCC TCC AGA AAC GAA C-3′ and 5′-CAG GGT CCC TCT GTC ATC TC-3′; medaka lck, 5′-CGA ACA CTG CAA CTG TCC AA-3′ and 5′-ACA AGC TCC TTC AGC GAG TT-3′; medaka p63, 5′-CCA CGC TCA GAA CAA CGT GA-3′ and 5′-GAT CTG AAT GGG GCA CGT CT-3′; medaka p73, 5′-CAA TCC CCT CCA ACA CCG ATT-3′ and 5′-TCG TGA TTG GGG CAT CGT TTG-3′; zebrafish WDR55, 5′-AAA GAG CTC TGG TCA TCA GG-3′ and 5′-TAT CCC AAA CCT TCA GCG TT-3′; mouse WDR55, 5′-TCC ATC CGA CTC GAG ATC TG-3′ and 5′-GCC ATG TCG GCA ATG TAC TC-3′; and mouse GAPDH, 5′-CCG GTG CTG AGT ATG TCG TG-3′ and 5′-CAG TCT TCT GGG TGG CAG TG-3′. Other primers for medaka genes were described previously [15],[30],[37],[60]. Zebrafish β-actin primers and mouse p21 primers were as described by Mathavan et al. (2005) [61] and Boley et al. (2002) [62], respectively. For the detection of maternal and embryonic WDR55 expression, RT-PCR products were incubated with BstNI. BstNI cuts cab-derived but not kaga-derived WDR55 sequence. Primers used were 5′-GAC AGA TCC TCC AGA AAC GAA C-3′ and 5′-GCC GTC TCT TGA TGT TGA AGA C-3′.
Total RNA was separated by electrophoresis in either 1% agarose gel containing MOPS or 6% polyacrylamide gel containing 8 M urea, and blotted onto positively charged nylon membrane (Biodyne Plus, Pall). After UV crosslinking, blotted total RNA was stained with methylene blue. The sequences of locked nucleic acid (LNA) probes conjugated with digoxigenin (DIG) at 3′ terminus are as follows: medaka and mouse 5.8S (probe used in Figure 3C and probe C in Figures 4A and E), 5′-tTC tTC aTC gAC gCA cGA gC-3′; medaka 18S, 5′-tAC tCC cCC cGG aAC cCA aA -3′ (probe A in Figure 4D); medaka ITS1, 5′-GtG CtG CtT CgC CaC GtT Cg-3′ (probe B in Figure 4B); and medaka ITS2, 5′-GaG CgG GgA AcA CcG AtT Ga-3′ (probe D in Figure 4D) (Greiner Bio-One). Small letters indicate LNAs. Medaka genomic DNA fragment containing 28S sequence (probe E in Figure 4D) was amplified with 5′-GAT TCC CAC TGT CCC TAC CT-3′ and 5′-AGA TCA AGC GAG CTT TTG CC-3′ primers. This DNA fragment was cloned in pCRII vector (Invitrogen), cut with XhoI, and transcribed using SP6 polymerase for DIG-labeled antisense probe (Roche DIG RNA Labeling Kit). Hybridization was carried out at 68°C and signals were detected with Gene Images CDP-Star Detection Kit (Amersham BioSciences).
Lysates of siRNA-transfected NIH3T3 in lysis buffer (50 mM Tris-HCl pH 8.0, 137 mM NaCl, 1% NP-40, 0.5% deoxycholate, and 0.1% SDS) with protease inhibitors were electrophoresed over SDS-PAGE and transferred to Immobilon-P membranes (Millipore). Membranes were incubated with 1/300 dilution of anti-WDR55 antibody or 1/300 dilution of rabbit anti-calnexin antibody (Santa Cruz), followed by horseradish-peroxidase-conjugated goat anti-rabbit IgG antibody (Santa Cruz), and visualized using ECL Plus Western Blotting Detection System (Amersham). Immunoprecipitation of U2OS cell lysates with anti-Pes1, anti-Bop1, and anti-WDR12 antibodies (Ascenion) was carried out as described [35].
Database accession numbers from DDBJ/GenBank/EMBL for the genes identified in this study were as follows: medaka WDR55, AB372859; medaka lck, AB372860; medaka p63, AB372861; and medaka p73, AB372862.
|
10.1371/journal.pgen.1005004 | Recent Selective Sweeps in North American Drosophila melanogaster Show Signatures of Soft Sweeps | Adaptation from standing genetic variation or recurrent de novo mutation in large populations should commonly generate soft rather than hard selective sweeps. In contrast to a hard selective sweep, in which a single adaptive haplotype rises to high population frequency, in a soft selective sweep multiple adaptive haplotypes sweep through the population simultaneously, producing distinct patterns of genetic variation in the vicinity of the adaptive site. Current statistical methods were expressly designed to detect hard sweeps and most lack power to detect soft sweeps. This is particularly unfortunate for the study of adaptation in species such as Drosophila melanogaster, where all three confirmed cases of recent adaptation resulted in soft selective sweeps and where there is evidence that the effective population size relevant for recent and strong adaptation is large enough to generate soft sweeps even when adaptation requires mutation at a specific single site at a locus. Here, we develop a statistical test based on a measure of haplotype homozygosity (H12) that is capable of detecting both hard and soft sweeps with similar power. We use H12 to identify multiple genomic regions that have undergone recent and strong adaptation in a large population sample of fully sequenced Drosophila melanogaster strains from the Drosophila Genetic Reference Panel (DGRP). Visual inspection of the top 50 candidates reveals that in all cases multiple haplotypes are present at high frequencies, consistent with signatures of soft sweeps. We further develop a second haplotype homozygosity statistic (H2/H1) that, in combination with H12, is capable of differentiating hard from soft sweeps. Surprisingly, we find that the H12 and H2/H1 values for all top 50 peaks are much more easily generated by soft rather than hard sweeps. We discuss the implications of these results for the study of adaptation in Drosophila and in species with large census population sizes.
| Evolutionary adaptation is a process in which beneficial mutations increase in frequency in response to selective pressures. If these mutations were previously rare or absent from the population, adaptation should generate a characteristic signature in the genetic diversity around the adaptive locus, known as a selective sweep. Such selective sweeps can be distinguished into hard selective sweeps, where only a single adaptive mutation rises in frequency, or soft selective sweeps, where multiple adaptive mutations at the same locus sweep through the population simultaneously. Here we design a new statistical method that can identify both hard and soft sweeps in population genomic data and apply this method to a Drosophila melanogaster population genomic dataset consisting of 145 sequenced strains collected in North Carolina. We find that selective sweeps were abundant in the recent history of this population. Interestingly, we also find that practically all of the strongest and most recent sweeps show patterns that are more consistent with soft rather than hard sweeps. We discuss the implications of these findings for the discovery and quantification of adaptation from population genomic data in Drosophila and other species with large population sizes.
| The ability to identify genomic loci subject to recent positive selection is essential for our efforts to uncover the genetic basis of phenotypic evolution and to understand the overall role of adaptation in molecular evolution. The fruit fly Drosophila melanogaster is one of the classic model organisms for studying the molecular bases and signatures of adaptation. Recent studies have provided evidence for pervasive molecular adaptation in this species, suggesting that approximately 50% of the amino acid changing substitutions, and similarly large proportions of non-coding substitutions, were adaptive [1,2,3,4,5,6,7,8,9]. There is also evidence that at least some of these adaptive events were driven by strong positive selection (~1% or larger), depleting levels of genetic variation on scales of tens of thousands of base pairs in length [10,11].
If adaptation in D. melanogaster is indeed common and often driven by strong selection, it should be possible to detect genomic signatures of recent and strong adaptation [12,13,14]. Three cases of recent and strong adaptation in D. melanogaster are well documented and can inform our intuitions about the expected genomic signatures of such adaptive events. First, resistance to the most commonly used pesticides, carbamates and organophosphates, is known to be largely due to three point mutations at highly conserved sites in the gene Ace, which encodes the neuronal enzyme Acetylcholinesterase [15,16,17]. Second, resistance to DDT evolved via a series of adaptive events that included insertion of an Accord transposon in the 5’ regulatory region of the gene Cyp6g1, duplication of the locus, and additional transposable element insertions into the locus [18,19]. Finally, increased resistance to infection by the sigma virus, as well as resistance to certain organophosphates, has been associated with a transposable element insertion in the protein-coding region of the gene CHKov1 [20,21].
In-depth population genetic studies [17,19,21] of adaptation at these loci revealed that in all three cases adaptation failed to produce classic hard selective sweeps, but instead generated patterns compatible with soft sweeps. In a hard selective sweep, a single adaptive haplotype rises in frequency and removes genetic diversity in the vicinity of the adaptive locus [22,23,24]. In contrast, in a soft sweep multiple adaptive alleles present in the population as standing genetic variation (SGV) or entering as multiple de novo adaptive mutations increase in frequency virtually simultaneously bringing multiple haplotypes to high frequency [25,26,27,28,29]. In the cases of Ace and Cyp6g1, soft sweeps involved multiple de novo mutations [17,19,21] that arose after the introduction of pesticides, whereas in the case of CHKov1, a soft sweep arose in out-of-African populations from standing genetic variation (SGV) [17,19,21] present at low frequencies in the ancestral African population [20,21].
Unfortunately, most scans for selective sweeps in population genomic data have been designed to detect hard selective sweeps (although see [30]) and focus on such signatures as a dip in neutral diversity around the selected site [22,24,31], an excess of low or high-frequency alleles in the frequency spectrum of polymorphisms surrounding the selected site (i.e. Tajima’s D, Fay and Wu’s H, and Sweepfinder) [32,33,34,35,36], the presence of a single common haplotype [37], or the observation of a long and unusually frequent haplotype (iHS) [36,38,39,40]. In a soft sweep, however, multiple haplotypes linked to the selected locus can rise to high frequency and levels of diversity and allele frequency spectra should therefore be perturbed to a lesser extent than in a hard sweep. As a result, methods based on the levels and frequency distributions of neutral diversity have low power to detect soft sweeps [13,28,41,42].
Some genomic signatures do have power to detect both hard and soft sweeps. In particular, linkage disequilibrium (LD) measured between pairs of sites or as haplotype homozygosity should be elevated in both hard and soft sweeps. This expectation holds for hard sweeps and for soft sweeps that are not too soft, that is soft sweeps that have such a large number of independent haplotypes bearing adaptive alleles that linkage disequilibrium is no longer elevated beyond neutral expectations [41,43].
Given that none of the described cases of adaptation at Ace, Cyp6g1, and CHKov1 produced hard sweeps, it is possible that additional cases of recent selective sweeps in D. melanogaster remain to be discovered. Here we develop a statistical test based on modified haplotype homozygosity for detecting both hard and soft selective sweeps in population genomic data. We apply this test in a genome-wide scan in a North American population of D. melanogaster using the Drosophila Genetic Reference Panel (DGRP) data set [44], consisting of 162 fully sequenced isogenic strains from a North Carolina population. Our scan recovers the three known soft sweeps at Ace, Cyp6g1, and CHKov1, and identifies a large number of additional recent and strong selective sweeps. We develop an additional haplotype homozygosity statistic that can distinguish hard from soft sweeps and argue that the haplotype frequency spectra at the top 50 candidate sweeps are best explained by soft selective sweeps.
In this paper, we develop a set of new statistics for the detection and characterization of positive selection based on measurements of haplotype homozygosity in a predefined window. Our reasoning in developing these statistics is that haplotype homozygosity, defined as a sum of squares of the frequencies of identical haplotypes in a window, should be a sensitive statistic for the detection of both hard and soft sweeps, as long as the window is large enough that neutral demographic processes are unlikely to elevate haplotype homozygosity by chance [41,43]. At the same time, the window must not be so large that even strong sweeps can no longer generate frequent haplotypes spanning the whole window.
In order to determine an appropriate window length for the measurement of haplotype homozygosity in the DGRP data set, we first assessed the length scale of linkage disequilibrium decay expected in the DGRP data under a range of neutral demographic models for North American D. melanogaster. This length scale should roughly correspond to the window size over which we are unlikely to observe substantial haplotype structure by chance. We considered six demographic models (Fig. 1). The first demographic model is an admixture model of the North American D. melanogaster population proposed by Duchen et al. [45]. In this model, the North American population was co-founded by flies from Africa and Europe 3.05×10–4 Ne generations ago (where Ne ≈ 5x106). The second model is a modified admixture model, also proposed by Duchen et al. [45], in which the founding European population underwent a bottleneck before the admixture event (see S1 Table for complete parameterizations of both admixture models). The third model has a constant effective population size of Ne = 106 [46], which we considered for its simplicity, computational feasibility and, as we will argue below, its conservativeness for the purposes of detecting selective sweeps using our approach in the DGRP data. The fourth model is a constant Ne = 2.7x106 demographic model fit to Watterson’s θW estimated from short intron autosomal polymorphism data from the DGRP dataset (Methods). Finally, we fit a family of out-of-Africa bottleneck models to short intron regions in the DGRP data set using DaDi [47] (S2 Table) (Methods). The two bottleneck models we ultimately used are a severe but short bottleneck model (NB = 0.002, TB = 0.0002) and a shallow but long bottleneck model (NB = 0.4, TB = 0.0560), both of which fit the data equally well among a range of other inferred bottleneck models (see S1 Fig. for parameterization). All models except for the constant Ne = 106 model fit the DGRP short intron data in terms of the number of segregating sites (S) and pairwise nucleotide diversity (π) (S3 Table).
We compared the decay in pair-wise LD in the DGRP data at distances from a few base pairs to 10 kb with the expectations under each of the six demographic models using parameters relevant for our subsequent analysis of the DGRP data (Fig. 2). Specifically, we matched the sample depth of the DGRP data set (145 strains after quality control) and assumed a mutation rate (μ) of 10–9 events/bp per generation [48] and a recombination rate (ρ) of 5×10–7 centimorgans/bp (cM/bp) [49]. In the DGRP data analysis below, we exclude regions with a low recombination rate (ρ < 5x10–7 cM/bp). The use of ρ = 5x10–7 cM/bp should therefore generate higher LD in simulations than in the DGRP data and thus should be conservative for the purposes of defining the expected length scale of LD decay.
Fig. 2 shows that LD in the DGRP data is elevated beyond neutral expectations at all length scales (consistent with the observations in [50]), and dramatically so at the 10 kb length scale. The elevation in LD observed in the data is indicative of either linked positive selection driving haplotypes to high frequency, a lack of fit of current demographic models to the data, or both. Simulations under the most realistic demographic model, admixture [45], have the fastest decay in LD (S2 Fig.). This is likely because admixture models with two bottlenecks that are fit to diversity statistics generate more haplotypes compared to single bottleneck models, since the same haplotype is unlikely to be sampled independently in both bottlenecked ancestral populations. In contrast, LD under the constant Ne = 106 demographic scenario decays slower than in any other demographic scenario, as expected given that this model has the smallest effective population size.
Fig. 2 suggests that windows of 10 kb are large enough that neutral demography is unlikely to generate high values of LD and elevate haplotype homozygosity by chance, and should thus prevent a high rate of false positives. At the same time, the use of 10 kb windows for the measurement of haplotype homozygosity should still allow us to detect many reasonably strong sweeps, including the known cases of recent adaptation. The footprint of a hard selective sweep extends over approximately s/[log(Nes)ρ] basepairs, where s is the selection strength, Ne the population size, and ρ the recombination rate [22,23,51]. Sweeps with a selection coefficient of s = 0.05% or greater are thus likely to generate sweeps that span 10 kb windows in areas with recombination rate of 5×10–7 cM/bp. As the recombination rate increases, only selective sweeps with s > 0.05% should be observed in the 10 kb windows. Genomic analyses have suggested that adaptation in Drosophila is likely associated with a range of selection strengths, including values of ~1% [7,8,10] or greater as observed at Ace, Cyp6g1, and CHKov1. Our use of 10 kb windows in the rest of the analysis should thus bias the analysis toward detecting the cases of strongest adaptation in Drosophila.
We investigated haplotype spectra in simulations of neutral demography and both hard and soft selective sweeps arising from de novo mutations as well as SGV. For all haplotype spectra and homozygosity analyses in this paper we use windows of 400 SNPs, corresponding roughly to 10 kb in the DGRP data (Fig. 2). Haplotypes within a 400 SNP window are grouped together if they are identical at all SNPs in the window. We fixed the number of SNPs in a window to eliminate variability in the haplotype spectra due to varying numbers of SNPs.
The lower SNP density of the constant Ne = 106 model (S3 Table) effectively increases the size of the analysis window in terms of the number of base pairs when defining the windows in terms of the number of SNPs. Thus, the constant Ne = 106 model should reduce the rate of false positives because the recombination rate under this model is artificially increased. We therefore use the constant Ne = 106 model for the subsequent simulations of neutrality and selective sweeps.
To visualize sample haplotype frequency spectra, we simulated incomplete and complete sweeps with frequencies of the adaptive mutation (PF) at 0.5 or 1 at the time when selection ceased. (Note that below we will investigate a large number of scenarios, focusing on the effects of varying selection strength and the decay of sweep signatures with time). The number of independent haplotypes that rise in frequency simultaneously in soft sweeps—we call this “softness” of a sweep—should increase either (i) when the rate of mutation to de novo adaptive alleles at a locus becomes higher and multiple alleles arise and establish after the onset of selection at a higher rate, or (ii) when adaptation uses SGV with previously neutral or deleterious alleles that are present at higher frequency at the onset of selection [27,29]. More specifically, for sweeps arising from multiple de novo mutations, Pennings and Hermisson [29] showed that the key population genetic parameter that determines the softness of the sweep is θA = 4NeμA, proportional to the product of Ne, the variance effective population size estimated over the period relevant for adaptation [14,52], and μA, the mutation rate toward adaptive alleles at a locus per individual per generation [14]. The mutation-limited regime with hard sweeps corresponds to θA << 1, whereas θA > 1 specifies the non-mutation-limited regime with primarily soft sweeps. As θA becomes larger, the sweeps become softer as more haplotypes increase in frequency simultaneously [29]. In the case of sweeps arising from SGV, the softness of a sweep is governed by the starting partial frequency of the adaptive allele in the population prior to the onset of selection. For any given rate of recombination, adaptive alleles starting at a higher frequency at the onset of selection should be older and should thus be present on more distinct haplotypes and give rise to softer sweeps [27].
As can be seen in Fig. 3, most haplotypes in neutral demographic scenarios are unique in our 400 SNP windows, whereas selective sweeps can generate multiple haplotypes at substantial frequencies. Our plot of the haplotype frequency spectra and the expected numbers of adaptive haplotypes show that sweeps arising from de novo mutations become soft with multiple frequent haplotypes in the sample when θA ≥ 1. Sweeps from SGV become soft when the starting partial frequency of the adaptive allele prior to the onset of selection is ≥ 10–4 (100 alleles in the population). In both cases, sweeps become monotonically softer as θA increases or, respectively, the starting partial frequency of the adaptive allele becomes higher. These results conform to the expectations derived in [29].
The increase of haplotype population frequencies in both hard and soft sweeps can be captured using haplotype homozygosity [30,39,41]. If pi is the frequency of the ith most common haplotype in a sample, and n is the number of observed haplotypes, then haplotype homozygosity is defined as H1 = Σi = 1, …n pi2. We can expect H1 to be particularly high for hard sweeps, with only one adaptive haplotype at high frequency in the sample (Fig. 4A). Thus, H1 is an intuitive candidate for a test of neutrality versus hard sweeps, where the test rejects neutrality for high values of H1. A test based on H1 may also have acceptable power to detect soft sweeps in which only a few haplotypes in the population are present at high frequency. However, as sweeps become softer and the number of sweeping haplotypes increases, the relative contribution of individual haplotypes towards the overall H1 value decreases, and the power of a test based on H1 is expected to decrease.
To have a better ability to detect hard and soft sweeps using homozygosity statistics, we developed a modified homozygosity statistic, H12 = (p1 + p2)2 + Σi>2 pi2 = H1 + 2p1p2, in which the frequencies of the first and the second most common haplotype are combined into a single frequency (Fig. 4B). A statistical test based on H12 is expected to be more powerful in detecting soft sweeps than H1 because it combines frequencies of two similarly abundant haplotypes into a single frequency, whereas for hard sweeps the combination of the frequencies of the first and second most abundant haplotypes should not change haplotype homozygosity substantially [53]. We also considered a third test statistic, H123, which combines frequencies of the three most prevalent haplotypes in a sample into a single haplotype and then computes homozygosity. We will primarily employ H12 in subsequent analyses but will consider the effects of using H1 and H123 briefly as well.
To assess the ability of H12 to detect sweeps of varying softness and to distinguish positive selection from neutrality, we measured H12 in simulated sweeps arising from both de novo mutations and SGV while varying s, PF, and the time since the end of the sweep, TE, measured in units of 4Ne generations in order to model the decay of a sweep through recombination and mutation events over time. We first investigate the behavior of H12 under different selective regimes and then investigate its power in comparison with the popular haplotype statistic iHS.
Fig. 5A shows that for complete and incomplete sweeps with s = 0.01 and TE = 0, H12 monotonically decreases as a function of θA over the interval from 10–2 to 102. When θA ≤ 0.5, many sweeps are hard and H12 values are high. When θA ≈ 1, and practically all sweeps are soft, but not yet extremely soft, H12 retains much of its power. However, for θA > 10, where sweeps are extremely soft, H12 decreases substantially. Similarly, H12 is maximized when the starting frequency of the allele is 10–6 (one copy of the allele in the population generating hard sweeps from SGV) and becomes very small as the frequency of the adaptive allele increases beyond >10-3 (>1000 copies of the allele in the population) (Fig. 5B). Therefore, H12 has reasonable power to detect soft sweeps in samples of hundreds of haplotypes, as long as they are not extremely soft, but remains somewhat biased in favor of detecting hard sweeps.
H12 also increases as the ending partial frequency of the adaptive allele after selection ceased (PF) increases from 0.5 to 1 (Fig. 5A and 5B) and as the selection strength increases from 0.001 to 0.1 (Fig. 5C and 5D). We observe that sweeps arising from SGV with low selection coefficients have lower H12 values (Fig. 5D). This is most likely because such weak sweeps are effectively harder: as more of the haplotypes fail to establish, fewer haplotypes end up sweeping in the population leading to higher values of haplotype homozygosity. Fig. 5E and 5F further show that incomplete and complete sweeps decay with time due to recombination and mutation events, resulting in monotonically decreasing values of H12 with time. Overall this analysis demonstrates that H12 has most power to detect recent sweeps driven by strong selection.
We also assessed the ability of H12 to detect selective sweeps as compared to H1 and H123 by calculating the values of H1, H12, and H123 for sweeps generated under the parameters s = 0.01, TE = 0 and PF = 0.5. H12 consistently, albeit modestly, increases the homozygosity for younger soft sweeps as compared to H1 (S3 Fig.). The increase in homozygosity using H123 is marginal relative to homozygosity levels achieved by H12, so we chose not to use this statistic in our study.
Finally, we compared the abilities of H12 and iHS (integrated haplotype score), a haplotype-based statistic designed to detect incomplete hard sweeps [39,40], to detect both hard and soft sweeps. We created receiving operator characteristic (ROC) curves [54], which plot the true positive rate (TPR) of correctly rejecting neutrality in favor of a sweep (hard or soft) given that a sweep has occurred versus the false positive rate (FPR) of inferring a selective sweep, when in fact a sweep has not occurred.
In our simulations of selective sweeps we used θA = 0.01 as a proxy for scenarios generating almost exclusively hard sweeps, and θA = 10 as a proxy for scenarios generating almost exclusively soft sweeps. We chose θA = 10 for soft sweeps because this is the highest θA value with which H12 can still detect sweeps before substantially losing power given our window size of 400 SNPs and sample size of 145. Note that for soft sweeps with a lower value of θA the power of H12 should be higher. We modeled incomplete sweeps with PF = 0.1, 0.5, and 0.9, with varying times since selection had ceased of TE = 0, 0.001, and 0.01 in units of 4Ne generations. We simulated sweeps under three selection coefficients, s = 0.001, 0.01, and 0.1.
Fig. 6 and S4 Fig. show that the tests based on H12 and iHS have similar power for the detection of hard sweeps, although in the case of old and strong hard sweeps (TE = 0.01, s ≥ 0.01) iHS performs slightly better than H12. On the other hand, H12 substantially outperforms iHS in detecting soft sweeps and has high power when selection is sufficiently strong and the sweeps are sufficiently young. As sweeps become very old, neither statistic can detect them well, as expected.
We applied the H12 statistic to DGRP data in sliding windows of 400 SNPs with the centers of each window iterated by 50 SNPs. To classify haplotypes within each analysis window, we assigned the 400 SNP haplotypes into groups according to exact sequence identity. If a haplotype with missing data matched multiple haplotypes at all genotyped sites in the analysis window, then the haplotype was randomly assigned to one of these groups (Methods).
To assess whether the observed H12 values in the DGRP data along the four autosomal arms are unusually high as compared to neutral expectations, we estimated the expected distribution of H12 values under each of the six neutral demographic models. Fig. 7 shows that genome-wide H12 values in DGRP data are substantially elevated as compared to expectations under any of the six neutral demographic models. In addition, there is a long tail of outlier H12 values in the DGRP data suggestive of recent strong selective sweeps.
To identify regions of the genome with H12 values significantly higher than expected under neutrality, we calculated critical values (H12o) under each of the six neutral models based on a 1-per-genome false discovery rate (FDR) criterion. Our test rejects neutrality in favor of a selective sweep when H12 > H12o (Methods and S1 Text). The critical H12o values under all neutral demographic models are similar to the median H12 value observed in the DGRP data (Table 1), consistent with the observations of elevated genome-wide haplotype homozygosity and much slower decay in LD at the scale of 10 kb in the DGRP data compared to all neutral expectations (Fig. 2). We focused on the constant Ne = 106 model because it yields a relatively conservative H12o value (Table 1) and preserves the most long-range, pair-wise LD in simulations (Fig. 2).
For our genomic scan we chose to use the 1-per-genome FDR value calculated under the constant Ne = 106 model with a recombination rate of 5×10–7 cM/bp. Note that most H12o values are similar to the genome-wide median H12 value of 0.0155.
In order to call individual sweeps, we first identified all windows with H12 > H12o in the DGRP data set under the constant Ne = 106 model. We then grouped together consecutive windows as belonging to the same ‘peak’ if the H12 values in all of the grouped windows were above H12o for a given model and recombination rate (Methods). We then chose the window with the highest H12 value among all windows in a peak and used this H12 value to represent the entire peak.
We focused on the top 50 peaks with empirically most extreme H12 values, hypothesized to correspond to the strongest and/or most recent selective events (Fig. 8A). The windows with the highest H12 values for each of the top 50 peaks are highlighted in Fig. 8A. The highest H12 values for the top 50 peaks are in the tail of the distribution of H12 values in the DGRP data (Fig. 7) and thus are outliers both compared to the neutral expectations under all six demographic models and the empirical genomic distribution of H12 values. We observed peaks that have H12 values higher than H12o on all chromosomes, but found that there are significantly fewer peaks on 3L (2 peaks) than the approximately 13 out of 50 top peaks expected when assuming a uniform distribution of the top 50 peaks genome-wide (p = 0.00016, two-sided binomial test, Bonferroni corrected).
The three peaks with the highest observed H12 values correspond to the three known cases of positive selection in D. melanogaster at the genes Ace, Cyp6g1, and CHKov1 [17,19,21], confirming that the H12 scan is capable of identifying previously known cases of adaptation. In S4 Table, we list all genes that overlap with any of the top 50 peaks. Fig. 9A and S5 Fig. show the haplotype frequency spectra observed at the top 50 peaks. In contrast, Fig. 9B shows the frequency spectra observed under the six demographic models with the corresponding critical H12o values.
We performed several tests to ensure the robustness of the H12 peaks to potential artifacts (S1 Text). We first tested for associations of H12 peaks with inversions in the sample, but did not find any (S1 Text, S5 Table). In addition, we reran the scan in three different data sets of the same population and confirmed that unaccounted population substructure and variability in sequencing quality do not confound our results (S1 Text, S7 Fig.). We also sub-sampled the DGRP data set to 40 strains ten times and plotted the resulting distributions of H12 values. We found that in all subsamples there is an elevation in haplotype homozygosity relative to neutral demographic scenarios, suggesting that the elevation in haplotype homozygosity values is driven by the whole sample and not a particular subset of individuals (S8 Fig.). Finally, to ensure that haplotype homozygosity is not elevated by family structure, we excluded all related individuals and reran the scan, again recovering the majority of our top peaks (S1 Text, S7 Fig.).
We scanned chromosome 3R using H1 and H123 as our test statistics in order to determine the impact of our choice of grouping the two most frequent haplotypes together in our H12 test statistic on the location of the identified peaks (S9 Fig.). We found that the locations of the identified peaks are similar with all three statistics, but that some smaller peaks that cannot be easily identified with H1 are clearly identified with H12 and H123, as expected.
We applied the iHS statistic as described in Voight et al. 2006 [40] to all SNPs in the DGRP data to determine the concordance in the sweep candidates identified by iHS and H12 (Methods). Briefly, we searched for 100 kb windows that have an unusually large number of SNPs with standardized iHS values (|iHS|) > 2. The positive controls Ace, Cyp6g1, and CHKov1 are located within the 95 top 10% iHS 100 kb windows (Fig. 8B), validating this approach.
To determine how often a candidate region identified in the H12 scan is identified in the iHS scan and vice versa, we overlapped the top 50 H12 peaks with the 95 top 10% iHS 100Kb windows. We defined an overlap as the non-empty intersection of the two genomic regions defining the boundaries of a peak in the H12 scan and the non-overlapping 100Kb windows used to calculate enrichment of |iHS| values. We found that 18 H12 peaks overlap 28 |iHS| 100Kb enrichment windows. In contrast, fewer than 5 H12 peaks are expected to overlap approximately 7 iHS 100Kb windows by chance (Methods). The concordance between the two scans confirms that many of the peaks identified in the two scans are likely true selective sweeps and also suggests that the two approaches are not entirely redundant.
Our analysis of H12 haplotype homozygosity and the decay in long range LD in DGRP data suggests that extreme outliers in the H12 DGRP scan are in locations of the genome that may have experienced recent and strong selective sweeps. The visual inspection of the haplotype spectra of the top 10 peaks in Fig. 9A and the remaining 40 peaks in S5 Fig. reveals that they contain many haplotypes at substantial frequency. These spectra do not appear similar to those generated by hard sweeps in Fig. 3 or extreme outliers under neutrality in Fig. 9B, but instead visually resemble incomplete soft sweeps with s = 0.01 and PF = 0.5 either from de novo mutations with θA between 1 and 20 or from SGV starting at partial frequencies of 5x10–5 to 5x10–4 prior to the onset of selection (Fig. 3). The sweeps also appear to become softer as H12 decreases, consistent with our expectation that H12 should lose power for softer sweeps.
In order to gain intuition about whether the haplotype spectra for the top 50 peaks can be more easily generated either by hard or soft sweeps under various evolutionary scenarios, we developed a new haplotype homozygosity statistic, H2/H1, where H2 = Σi>1 pi2 = H1—p12 is haplotype homozygosity calculated using all but the most frequent haplotype (Fig. 4C). We expect H2 to be lower for hard sweeps than for soft sweeps because in a hard sweep only one adaptive haplotype is expected to be at very high frequency [53]. The exclusion of the most common haplotype should therefore reduce haplotype homozygosity precipitously. As sweeps get softer, however, multiple haplotypes start appearing at high frequency in the population and the exclusion of the most frequent haplotype should not decrease the haplotype homozygosity to the same extent. Conversely H1, the homozygosity calculated using all haplotypes, is expected to be higher for a hard sweep than for a soft sweep as we described above. The ratio H2/H1 between the two measures should thus increase monotonically as a sweep becomes softer, thereby offering a summary statistic that, in combination with H12, can be used to test whether the observed haplotype patterns are more likely to be generated by hard or soft sweeps. Note that we intend H2/H1 to be measured near the center of the sweep where H12 is the highest. Otherwise, when H2/H1 is estimated further away from the sweep center, mutation and recombination events will decay the haplotype signature and hard and soft sweep signatures can become indistinguishable.
To assess the behavior of H2/H1 as a function of the softness of a sweep, we measured H2/H1 in simulated sweeps of varying softness arising from de novo mutations and SGV with various s, PF, and TE values. Fig. 10 shows that H2/H1 has low values for sweeps with θA ≤ 0.5 or when the starting partial frequency of the adaptive allele prior to the onset of selection is <10–5, i.e., when sweeps are mainly hard. As a sweep becomes softer, H2/H1 values approach one because no single haplotype dominates the haplotype spectrum. In the case of sweeps arising from de novo mutations, H2/H1 values are similar for partial (PF = 0.5) and complete sweeps (PF = 1) and for sweeps of varying strengths (s = 0.001, 0.01, 0.1). However, in the case of sweeps arising from SGV, sweeps with higher selection strengths do have higher H2/H1 values, reflecting the hardening of sweeps for smaller s values as we discussed previously (Fig. 5D). Both sweeps from de novo mutations and SGV have higher H2/H1 values for older sweeps, reflecting the decay of the haplotype frequency spectrum over time.
While hard sweeps and neutrality cannot easily generate both high H12 and H2/H1 values, soft sweeps can do both. In Fig. 11 we assess the range of H12 and H2/H1 values expected under hard and soft sweeps. To compare the likelihood of a hard versus soft sweep generating a particular pair of H12 and H2/H1 values, we calculated Bayes factors: BF = P(H12obs, H2obs /H1obs |Soft Sweep)/P(H12obs, H2obs /H1obs |Hard Sweep). We approximated BFs using an approximate Bayesian computation (ABC) approach under which the nuisance parameters—selection coefficient (s), partial frequency of the adaptive allele after selection has ceased (PF), and age (TE)—are integrated out by drawing them from uniform prior distributions: s ~ U[0,1], PF ~ U[0,1], and TE ~ U[0,0.001]×4Ne. We stated the hard and soft sweep scenarios as point hypotheses in terms of the θA value generating the data. Specifically, we assumed that hard sweeps are generated under θA = 0.01. For soft sweeps, we generated sweeps of varying softness by using θA values of 5, 10, and 50. Note that hard and soft sweeps can also be simulated from SGV with various starting frequencies of the beneficial allele, but for the purposes of generating hard sweeps with a single sweeping haplotype versus soft sweeps with multiple sweeping haplotypes, simulations from SGV or de novo mutations are mostly equivalent.
The panels in Fig. 11 show BFs calculated under several evolutionary scenarios for a grid of H12 and H2/H1 values. All panels in Fig. 11 show that hard sweeps are common when H2/H1 values are low for most H12 values tested. For very low H12 (<0.05) values, when sweeps display low haplotype homozygosity to begin with and are difficult to detect with H12, both hard and soft sweeps are likely for a wide range of H2/H1 values. Soft sweeps are common for any high H2/H1 values conditional on H12 being sufficiently high when simulating soft sweeps with θA = 10 and 5 (Fig. 11A and 11B). However, soft sweeps generated with θA = 50 are too soft to produce high H12 values, confirming our results in Fig. 5. As a consequence hard sweeps are common for high H12 values regardless of H2/H1 values under this scenario (Fig. 11C). In Fig. 11A, 11D and 11E, the recombination rate is varied, and a comparison of these panels show that the recombination rate has little impact on the space where hard sweeps can be expected to be more likely. Fig. 11F shows that simulations under admixture increase support for soft sweeps in regions of the space already in support of soft sweeps generated under the constant Ne = 106 demographic scenario (Fig. 11A–E). Fig. 10 shows that there is clearly a dependency between H12 and H2/H1 and that both values need to be taken into account when determining the softness of a peak. In particular, H2/H1 is most informative when applied to regions of the genome with the highest H12 values.
Overlaid on all panels in Fig. 11 are the H12 and H2/H1 values at the top 50 peaks. Note that in almost all cases, the top 50 peaks have H12 and H2/H1 values that are easiest explained by soft sweeps. In order to more explicitly test each candidate sweep for its compatibility with a hard and soft sweep model, we generated hard sweeps with θA = 0.01 and soft sweeps with a maximum a posteriori θA value (θAMAP), i.e., our best estimate of the softness for a particular peak. We used an ABC method to infer the θAMAP for each peak by sampling the posterior distribution of θA conditional on the observed values H12obs and H2obs /H1obs from a candidate sweep (S1 Text). All θAMAP values inferred for the top 50 peaks were significantly greater than 1 with the smallest being 6.8 (S10 Fig.), suggesting that soft sweeps would be commonly generated under any of the θAMAP values estimated (Fig. 3). We used recombination rates estimated for each peak [49] and simulated the data under the constant population size model with Ne = 106 for computational feasibility. Among our top 50 peaks, we found strong evidence in support of soft sweeps in all 50 cases (BF > 10), very strong evidence in 47 cases (BF > 30), and almost decisive evidence (BF > 98) in 44 cases (S3 Table). Taken together, these results provide evidence that soft sweeps most easily explain the signatures of multiple haplotypes at high frequency observed at the top 50 H12 peaks.
In this study, we found compelling evidence for a substantial number of recent and strong selective sweeps in the North Carolina population of D. melanogaster and further found that practically all these events appear to display signatures of soft rather than hard sweeps. To detect sweeps, we used our new haplotype statistic, H12, which measures haplotype homozygosity after combining the frequencies of the two most abundant haplotypes into a single frequency in windows of 400 SNPs (~10 kb in the DGRP data).
We chose to use windows defined by a constant number of SNPs rather than windows of constant physical or genetic length in order to simplify the statistical analysis. This is because windows of constant physical or genetic length tend to have varying SNP density, and therefore also varying distributions of haplotypes even under neutrality. Our choice of a fixed number of SNPs avoids this source of noise, but it raises the question of whether the H12 peaks simply define regions that have particularly low recombination rates or high SNP densities, and thus short windows in terms of the number base pairs or genetic map length. We made sure to avoid the first pitfall by analyzing only windows with reasonably high recombination rates (ρ ≥ 5x10–7 cM/bp, 82% of the genome) and by using conservative thresholds for the significance cutoffs. We also confirmed that the analysis windows with the highest H12 values in our top 50 peaks do not have shorter windows in terms of base pairs than on average (S11 Fig.). We were further concerned that our choice of using windows with a fixed number of SNPs would bias us against detecting complete hard sweeps. However, our simulations showed that this was not the case (Fig. 5).
We fully acknowledge that the result of applying the haplotype statistics developed in this manuscript to the North Carolina population may be idiosyncratic to the particular demographic structure of this one population. However, H12 in the DGRP data is substantially elevated compared to the expectation under any of the tested neutral demographic models, including both published admixture models [45] and the bottleneck models we fit to the DGRP short intron SNP data. In fact, the median value of H12 in the genome lies in the tails of distributions of H12 values generated from > 105 simulations for each neutral demographic scenario. Similarly, pairwise LD in DGRP data decays much more slowly than expected under neutrality (Fig. 2). These patterns can be due either to (i) pervasive and strong positive selection that drives long haplotypes to high frequency in the population, (ii) misspecification of the demographic model, or (iii) both. Although background selection (BGS) is pervasive in D. melanogaster [55] and strongly impacts levels of polymorphism, it is unlikely to be responsible for high levels of haplotype homozygosity [56,57].
Both selective and neutral demographic explanations of the elevated LD need to be investigated further. It will be important to determine whether current estimates of the rate and strength of adaptation in D. melanogaster are consistent with the elevated levels of haplotype homozygosity and LD in general, even under simple demographic models. Alternatively, an unusually high rate of adaptation in the recent past might be required to explain the signatures we observe in the data. Likewise, it is possible that some demographic model of the North Carolina population, which is yet to be specified, can account for the observed LD patterns. Both extensive forward simulations and additional studies of LD and haplotype homozygosity patterns in other populations will be important to resolve these issues.
Importantly, however, the top fifty H12 peaks we focused on in this study are outliers not only under all tested demographic models, but also relative to the empirical genome wide H12 distribution. The top three peaks correspond to the well-known cases of soft selective sweeps arising from de novo mutations and SGV at the loci Ace, Cyp6g1, and CHKov1 [17,19,21] as described in the Introduction. The recovery of these positive controls further validates that our method can identify sweeps arising from both de novo mutations and SGV and is robust to misspecifications of demographic models.
In order to confirm the robustness of the H12 peaks, we ran iHS [40] on the DGRP data and recovered 18 of the top 50 peaks, including the three positive controls, demonstrating the validity of both methods and that the two methods are not entirely redundant (Fig. 8B). We also failed to detect any correlation between H12 peaks and inversions in the genome. We tested for any unaccounted substructure in the data confounding our results by rerunning the scan in several data sets, including one where all related individuals were excluded. In all cases, we found that our top peaks remained unchanged and that haplotype homozygosity was consistently elevated in the data relative to neutral demographic simulations (S1 Text). We are thus confident that the top H12 peaks are true outliers and likely indicate recent and strong selective events in the North Carolina population of D. melanogaster.
To assess whether the top 50 peaks can be more easily generated by hard versus soft sweeps, we developed a second statistic, H2/H1, which is a ratio of haplotype homozygosities calculated without (H2) and with (H1) the most frequent haplotype in a sample. We demonstrate that this statistic has a monotonically increasing relationship with the softness of a sweep (Fig. 10), in contrast to H12, which has a monotonically decreasing relationship with the softness of a sweep.
H2/H1 and H12 together are informative in determining the softness of a sweep. Specifically, hard sweeps can generate high values of H12 in a window centered on the adaptive site but cannot simultaneously generate high H2/H1 values in the same window. However, soft sweeps can generate both high H12 and H2/H1 values in such a window. Note that in order to differentiate hard and soft sweeps with reasonable power, H2/H1 can only be applied in cases where H12 values are already high and there is strong evidence for a sweep. Indeed, as can be seen in all evolutionary scenarios presented in Fig. 11, when H12 is high and H2/H1 is low, hard sweeps are common, and when both H12 and H2/H1 are high, soft sweeps are common. However, when H12 is low, i.e. when there is little evidence for a sweep to begin with, either because the sweep was driven by weak selection or happened a long time ago, a wider range of H2/H1 values are compatible with hard sweeps. This demonstrates that H2/H1 can be used only in windows with very high H12 values. In most cases this should not unduly restrict the analysis as all robustly identified sweeps must have high H12 values given the difficulties of correctly specifying demographic models for any population.
The visual inspection (Fig. 9 and S5 Fig.) and the Bayesian analysis of the H12 and H2/H1 values suggest that all top 50 H12 peaks were driven by soft sweeps. Note that we simulated hard and soft sweeps for the Bayesian analysis under the constant Ne = 106 demographic model for computational feasibility and to make our analysis conservative for the purposes of rejecting the hard sweep scenario. This is because the lower SNP density in the Ne = 106 model (S3 Table), as compared to DGRP data, effectively increases the analysis window size in terms of base pairs, and by extension, also increases the number of recombination events each window experiences. Thus, hard sweeps should look “softer” under this choice of demographic model [53]. Even still, soft sweeps and not hard sweeps seem to more easily explain the signatures at our top 50 peaks.
If soft sweeps are indeed common in D. melanogaster, then adaptation must commonly act on SGV at low enough frequencies to generate high H12 values or involve multiple de novo adaptive mutations entering the population simultaneously. The SGV scenario is clearly plausible, particularly if much adaptation in out-of-Africa populations of D. melanogaster utilized variants that are rare in Africa. We do, however, expect that many adaptive events will involve SGV at higher frequencies and such adaptive events will generate sweeps that are too soft to be detectable using the H12 statistic. Similarly, θA values much larger than 10 will also generate sweeps too soft to be detected by H12. Curiously, this upper bound of θA is consistent with the median θA inferred from our top 50 peaks, ~12.8 (S10 Fig.). This coincidence suggests that we might still be missing many sweeps that are too soft for detection using H12.
Is it plausible that some of the sweeps were generated by de novo mutation? The answer must be clearly yes given that two of three known cases of recent adaptation, at Ace and Cyp6g1, were generated by de novo mutation. In order for this to be possible, the total population scaled adaptive mutation rate (θA) must be on the order of one or even larger [27,29]. The commonly assumed value of Ne = 106 for the effective population size in D. melanogaster and mutation rate per base pair (~10–9 bp/generation [48]) implies θA values of approximately 1%, assuming that adaptation at a given locus relies on mutation at a single nucleotide. One reason why θA can be commonly greater than 0.01 is that many mutations at a locus can be adaptive, for instance if adaptation relies on gene loss and any stop codon or indel is equally adaptive. In this case, all such adaptive mutations at a locus will combine to generate a soft sweep.
In addition, the population size relevant for recent adaptation might be much closer to the census population size at the time of adaptation and thus can be much larger than the commonly assumed value of Ne = 106 for the effective population size in D. melanogaster. We favor this explanation of a much larger effective population size of D. melanogaster relevant for recent and strong adaptation for two reasons. First, it is unlikely that every single case of recent and strong adaptation was driven by a situation where the adaptive mutation rate at a locus was a hundred times higher than a mutation rate at a single site. Second, in the case of adaptation at Ace, adaptation was driven by three point mutations, and the soft sweeps at Ace are incompatible with the relevant population size being on the order of 106 [17]. The relevant population size for recent and strong adaptation in D. melanogaster should be thus more than 100-fold than 106. Note that the relevant population size here is that of the D. melanogaster population as a whole and not just the North Carolina DGRP population. A likely possibility is that we observe signatures of multiple local hard sweeps arising within sub-demes of the North American Drosophila population or in the ancestral European and African populations prior to admixture, that combine to generate signatures of soft sweeps [58].
Nevertheless, it is quite puzzling that we were unable to detect any hard sweeps. One possibility is that hard sweeps do exist but are driven by weaker selection than we can detect in our scan. Indeed, Wilson et al. [52] argued that sweeps driven by weak selection could become hard even when they occur in populations of large size. This is because such sweeps take a long enough time to increase in frequency allowing rare but sharp bottlenecks to eliminate all but the highest frequency adaptive allele. It is also possible that hard sweeps were common in the past and degraded over time, while recent adaptation from de novo or rare variants produced primarily soft sweeps. While it is possible that hard sweeps correspond to the weaker and older selection events that we lack the power to identify, it is reassuring that our method is biased toward discovering the strongest, most recent, and thus most consequential adaptive events in the genome.
The abundance of signatures of soft sweeps in D. melanogaster has important implications for the design of methods used to quantify adaptation. Some methods may work equally well whether adaptation proceeds via hard or soft sweeps. For instance, estimates of the rate of adaptive fixation derived from McDonald-Kreitman tests [59] are not expected to be affected strongly because these estimates depend on the rate of fixation of adaptive mutations and not on the haplotype patterns of diversity that these adaptive fixations generate in their wake. Tests based on the prediction that regions of higher functional divergence should harbor less neutral diversity [10,11,60] are generally consistent with recurrent hard and soft sweeps, as both scenarios are expected to increase levels of genetic draft, and thus reduce neutral diversity in regions of frequent and recurrent adaptation. Note that soft sweeps generate less of a reduction in neutral diversity. As a consequence, such methods might underestimate the rate of adaptation. However, methods that quantify adaptation based on a specific functional form of the dependence between the level of functional divergence and neutral diversity may lead to different conclusions under hard and soft sweeps [10]. Finally, methods that rely on the specific signatures of hard sweeps, such as the presence of a single frequent haplotype [39,40], sharp local dips in diversity [22], or specific allele frequency spectra expected during the recovery after the sweep might often fail to identify soft sweeps [35]. Hence, such methods might give us an incomplete picture of adaptation. Moreover, such methods might erroneously conclude that certain genomic regions lacked recent selective sweeps, which can be problematic for demographic studies that rely on neutral polymorphism data unaffected by linked selection.
Our statistical test based on H12 to identify both hard and soft sweeps and our test based on H12 and H2/H1 to distinguish signatures of hard versus soft sweeps can be applied in all species in which genome-scale polymorphism data are available. The current implementation requires phased data but the method can easily be extended to unphased data as well by focusing on the frequencies of homozygous genotypes. Our method requires a sufficiently deep population sample for the precise measurement of haplotype frequencies, which is essential for determining whether a haplotype is unusually frequent in the sample. For example, in our DGRP scan, the majority of the 50 highest H12 peaks had a combined frequency of the two most common haplotypes below 30%, while only the top three peaks had a combined frequency of approximately 45%. Determination of whether a sweep is hard or soft should be particularly sensitive to the depth of the population sample. Finally, in order to determine whether an observed H12 value is sufficiently high enough to suggest that a sweep has occurred in the first place, reliable estimates of recombination rates are needed. We encourage the use of an empirical outlier approach to identify sweep candidates, especially because it is often difficult to accurately infer appropriate demographic models.
Our results provide evidence that signatures of soft selective sweeps were abundant in recent evolution of D. melanogaster. Soft sweep signatures may be common in many additional organisms with high census population sizes, including plants, marine invertebrates, insects, microorganisms, and even modern humans when considering very recent evolution in the population as a whole. Indeed, the list of known soft sweeps is large, phylogenetically diverse, and is constantly growing [14]. A comprehensive understanding of adaptation therefore must account for the possibility that soft selective sweeps are a frequent and possibly dominant mode of adaptation in nature.
Population samples under selection and neutrality were simulated with the coalescent simulator MSMS [61]. We simulated samples of size 145 to match the sample depth of the DGRP data and always assumed a neutral mutation rate of 10–9 events/bp/gen [48].
MSMS can simulate selective sweeps both from de novo mutations and SGV. We simulated sweeps of varying softness arising from de novo mutations by specifying the population parameter θA = 4NeμA at the adaptive site. We simulated sweeps arising from SGV by specifying the initial frequency of the adaptive allele in the population at the onset of positive selection. The adaptive site was always placed in the center of the locus. We assumed co-dominance, whereby a homozygous individual bearing two copies of the advantageous allele has twice the fitness advantage of a heterozygote. To simulate incomplete sweeps we specified the ending partial frequency of the adaptive allele after selection has ceased. To simulate sweeps of different ages, we conditioned on the ending time of selection (TE) prior to sampling.
When simulating selection with the admixture demographic model, it was unfortunately not possible in MSMS to condition on TE. For this demographic scenario, we instead conditioned on the start time of selection in the past and the starting partial frequency of the adaptive allele prior to the onset of selection, with selection continued until the time of sampling. In doing so, we assumed a uniform prior distribution of the start time of selection, U[0 to 3.05×10–4Ne] generations, with the upper bound specifying the time of the admixture event.
We simulated loci of length 105 bp for sweep simulations with s < 0.1 and 106 bp for sweep simulations with s = 0.1. For neutral simulations, we simulated loci of length 105 bp. We assumed a constant effective population size of Ne = 106 and a recombination rate of 5×10–7 cM/bp, reflecting the cutoff used in the DGRP analysis.
Our statistics H12 and H2/H1 were estimated over windows of size 400 SNPs centered on the adaptive site. Simulated samples that yielded fewer than 400 SNPs were discarded. For the comparison with iHS, we calculated iHS values for the SNP immediately to the right of the selected allele, and determined the size of the region by cut-off points at which iHS levels decayed to values observed under neutrality. In some simulation runs under the extreme scenario with s = 0.1 and TE = 0, iHS had not yet decayed to neutral levels at the edges of the simulated sweep. However, this should have only minor impact on the ROC curves.
The DGRP data set generated by Mackay et al. (2012) [44] consists of the fully sequenced genomes of 192 inbred D. melanogaster lines collected from Raleigh, North Carolina. Reference genomes are available only for 162 lines. Of these 162 lines, we filtered out a further 10% of the lines with the highest number of heterozygous sites in their genomes, possibly reflecting incomplete inbreeding. The IDs of these strains are: 49, 85, 101, 109, 136, 153, 237, 309, 317, 325, 338, 352, 377, 386, 426, 563, and 802. Any remaining residual heterozygosity in the data was treated as missing data. Our final data set consisted of 145 strains.
We measured linkage disequilibrium (LD) in DGRP data and in simulations of neutral demographic scenarios in samples of size 145. Simulations were performed assuming a neutral mutation rate of 10–9 events/bp/gen and a recombination rate of 5x10–9 cM/bp. LD was measured using the R2 statistic in sliding windows of 10 kb iterated by 50 bps. LD was measured between the first SNP in the window with an allele frequency between 0.05 and 0.95 and the rest of the SNPs in the window with allele frequencies between 0.05 and 0.95. If any SNP had missing data, the individuals with the missing data were excluded from the LD calculation. At least 4 individuals without missing data at both SNPs were required to compute LD, otherwise the SNP pair was discarded. LD plots were smoothed by averaging LD values binned in non-overlapping 20 bp windows until a distance of 300 bps. After that, LD values were averaged in bins of 150 bp non-overlapping windows.
We scanned the genome using sliding windows of 400 SNPs with intervals of 50 SNPs between window centers and calculated H12 in each window. If two haplotypes differed only at sites with missing data, we clustered these haplotypes together. If multiple haplotypes matched a haplotype with missing data, we clustered the haplotype with missing data at random with equal probability with one of the other matching haplotypes. We treated heterozygous sites in the data as sites with missing data (“N”).
To identify regions with unexpectedly high values of H12 under neutrality, we calculated the expected distribution of H12 values under the admixture, admixture and bottleneck, constant Ne = 106, constant Ne = 2.7x106, severe short bottleneck, and shallow long bottleneck demographic scenarios specified in Fig. 1. For each scenario, we simulated ten times the number of independent analysis windows (approximately 1.3x105 simulations) observed on chromosomes 2L, 2R, 3L, and 3R using three different recombination rates: 10–7 cM/bp, 5×10–7 cM/bp, and 10–6 cM/bp. All simulations were conducted with locus lengths of 105 basepairs. We assigned a 1-per-genome FDR level to be the 10th highest H12 value in each scenario.
Consecutive windows with H12 values that are above the 1-per-genome-FDR level were assigned to the same peak by the following algorithm: first, we identified the analysis window with the highest H12 value along a chromosome above the 1-per-genome-FDR with a recombination rate greater than 5×10–7 cM/bp. We then grouped together all consecutive windows with H12 values that lie above the cutoff and assigned all these windows to the same peak. After identifying a peak, we chose the highest H12 value among all windows in the peak to represent the H12 value of the entire peak. We repeated this procedure for the remaining windows until all analysis windows were accounted for.
We scanned the DGRP data using a custom implementation of the iHS statistic written by Sandeep Venkataram and Yuan Zhu. iHS was calculated for every SNP with a minor allele frequency (MAF) of at least 0.05 without polarization. Any strain with missing data in the region of extended haplotype homozygosity for a particular SNP was discarded in the computation of iHS. All iHS values were standardized by the mean and variance of iHS values calculated at all SNPs sharing a similar MAF (within ± 0.05). As described in Voight et al. [40], we calculated the enrichment of SNPs with standardized iHS values > 2 in non-overlapping 100 Kb windows.
To determine the number of top H12 peaks that should overlap the top |iHS| enrichment regions by chance, we calculated the expected fraction of the genome that should overlap the top candidates in both scans. The top 50 H12 peaks cover a total of 7,166,386 bps of the genome, or, 7.42% of the genome. Similarly, the top 95 |iHS| enrichment windows with |iHS| > 2 cover 9,500,000 bps of the genome, or 9.83% of the genome. Thus, only 0.73% of the genome should overlap both the top H12 peaks and top |iHS| enrichment windows by chance. Multiplying this percentage with the total number of bps in the DGRP data set (96,595,864) and normalizing by the total area of the genome covered by the top 50 H12 peaks and top 95 |iHS| enrichment regions, only ~10% of the fraction of the genome covered by H12 peaks should overlap ~7.4% of the fraction of the genome covered by |iHS| enrichment regions. Assuming a uniform distribution of H12 peaks in the region of the genome covered by H12 peaks, approximately 5 H12 peaks should overlap approximately 7 |iHS| enrichment regions by chance.
We fit six simple bottleneck models to DGRP data using a diffusion approximation approach as implemented by the program DaDi [47]. DaDi calculates a log-likelihood of the fit of a model based on an observed site frequency spectrum (SFS).
We estimated the SFS for presumably neutral SNPs in the DGRP using segregating sites in short introns [62]. Specifically, we used every site in a short intron of length less than 86 bps, with 16 bps removed from the intron start and 6 bps removed from the intron end [63]. We projected the SFS for our data set down to 130 chromosomes (after excluding the top 10% of strains with missing data), resulting in 42,679 SNPs out of a total of 738,024 bps.
We specified a constant population size model as well as six bottleneck models with the sizes of the bottlenecks ranging from 0.2% to 40% of the ancestral population size. Using DaDi [47], we inferred three free parameters: the bottleneck time (TB), final population size (NF), and the final population time (TF) (S1 Fig. and S2 Table). All six bottleneck models produced approximately the same log likelihood values and estimates of NF and TF. Further, the estimates of S and π obtained from simulated data matched the estimates obtained from the observed short intron data (S3 Table). Note that the estimate of TB is proportional to NB, reflecting the difficulty in distinguishing short and deep bottlenecks from long and shallow bottlenecks. We inferred Ne = 2,657,111 (≈2.7x106) for the constant population size model, assuming a mutation rate of 10–9/bp/generation.
To infer θAMAP values for the top 50 peaks (S1 Text), we assumed uniform distributions for all model parameters in our ABC procedure: The adaptive mutation rate (θA) took values on [0,100], the selection coefficient s on [0,1], the ending partial frequency of the adaptive allele after selection has ceased (PF) on [0,1], and the age of the sweep (TE) on [0,0.001]×4Ne. We assigned a recombination rate to each peak according to the estimates from Comeron et al. (2012) [49] for the specific locus. For the ABC procedure, we binned recombination rates into 5 equally spaced bins. Then, for each peak, we simulated the recombination rate from a uniform distribution over the particular bin its recombination rate fell in. The recombination rate intervals defining the 5 bins were: [5.42*10–7, 1.61*10–6), [1.61*10–6, 2.68*10–6), [2.68*10–6, 3.74*10–6), [3.74*10–6, 4.81*10–6), [4.81*10–6, 5.88*10–6) in units of cM/bp. We assumed a demographic model with constant Ne = 106 and a non-adaptive mutation rate of 10–9 bp/gen in our simulations.
For each peak, we sampled an approximate posterior distribution of θA by finding 1000 parameter values that generated sweeps with H12 and H2/H1 values within 10% of the observed values H12obs and H2obs /H1obs for the particular peak. We calculated the lower and upper 95% credible interval bounds for θA using the 2.5th and 97.5th percentiles of the posterior sample. On each posterior sample, we applied a Gaussian smoothing kernel density estimation and obtained the maximum a posteriori estimate θAMAP for each peak.
We used the same procedure for obtaining approximate posterior distributions of θA and θAMAP estimates under the admixture model. In this case, instead of sampling the time when selection ceased, we sampled the time of the onset of selection with uniform prior distribution: U[0, 3.05×10–4]×Ne, where 3.05×10–4Ne generations is the time of the admixture event. The prior distributions for all other parameters were the same as for the constant Ne = 106 model.
We used an ABC approach to calculate Bayes factors for a range of H12 and H2/H1 values. We simulated hard sweeps with θA = 0.01 and soft sweeps with θA = 5, 10, 50, or the θAMAP inferred for a particular peak, depending on the scenario being tested. In the constant Ne = 106 models shown in Fig. 11A–E, selection coefficients, partial frequencies of the adaptive allele after selection has ceased, and sweep ages were drawn from uniform distributions as follows: s ~ U[0,1], TE ~ U[0, 104]×4Ne, PF ~ U[0,1]. For the admixture model in Fig. 11F, the age of the onset of selection was sampled from a uniform distribution: U[0, 3.05×10–4]Ne generations, where 3.05×10–4Ne generations corresponds to the time of the admixture event.
We calculated Bayes factors by taking the ratio of the number of data sets simulated with H12 and H2/H1 values with a Euclidean distance < 0.1 from the observed values H12obs and H2obs /H1obs for each set of 106 simulated data sets under soft versus hard sweeps (105 data sets were generated for explicitly testing each peak with θAMAP). We calculated the Euclidean distance as follows: di = [(H12obs—H12i)2 /Var(H12) + (H2obs/H1obs—H2i/H1i)2 /Var(H2/H1)]1/2, where Var(H12) and Var(H2/H1) are the estimated variances of the statistics H12 and H2/H1 calculated using all simulated data sets.
|
10.1371/journal.pgen.1007206 | Deep sequencing of HBV pre-S region reveals high heterogeneity of HBV genotypes and associations of word pattern frequencies with HCC | Hepatitis B virus (HBV) infection is a common problem in the world, especially in China. More than 60–80% of hepatocellular carcinoma (HCC) cases can be attributed to HBV infection in high HBV prevalent regions. Although traditional Sanger sequencing has been extensively used to investigate HBV sequences, NGS is becoming more commonly used. Further, it is unknown whether word pattern frequencies of HBV reads by Next Generation Sequencing (NGS) can be used to investigate HBV genotypes and predict HCC status. In this study, we used NGS to sequence the pre-S region of the HBV sequence of 94 HCC patients and 45 chronic HBV (CHB) infected individuals. Word pattern frequencies among the sequence data of all individuals were calculated and compared using the Manhattan distance. The individuals were grouped using principal coordinate analysis (PCoA) and hierarchical clustering. Word pattern frequencies were also used to build prediction models for HCC status using both K-nearest neighbors (KNN) and support vector machine (SVM). We showed the extremely high power of analyzing HBV sequences using word patterns. Our key findings include that the first principal coordinate of the PCoA analysis was highly associated with the fraction of genotype B (or C) sequences and the second principal coordinate was significantly associated with the probability of having HCC. Hierarchical clustering first groups the individuals according to their major genotypes followed by their HCC status. Using cross-validation, high area under the receiver operational characteristic curve (AUC) of around 0.88 for KNN and 0.92 for SVM were obtained. In the independent data set of 46 HCC patients and 31 CHB individuals, a good AUC score of 0.77 was obtained using SVM. It was further shown that 3000 reads for each individual can yield stable prediction results for SVM. Thus, another key finding is that word patterns can be used to predict HCC status with high accuracy. Therefore, our study shows clearly that word pattern frequencies of HBV sequences contain much information about the composition of different HBV genotypes and the HCC status of an individual.
| HBV infection can lead to many liver complications including hepatocellular carcinoma (HCC), one of the most common liver cancers in China. High-throughput sequencing technologies have recently been used to study the genotype sequence compositions of HBV infected individuals and to distinguish chronic HBV (CHB) infection from HCC. We used NGS to sequence the pre-S region of a large number of CHB and HCC individuals and designed novel word pattern based approaches to analyze the data. We have several surprising key findings. First, most HBV infected individuals contained mixtures of genotypes B and C sequences. Second, multi-dimensional scaling (MDS) analysis of the data showed that the first principal coordinate was closely associated with the fraction of genotype B (or C) sequences and the second principal coordinate was highly associated with the probability of HCC. Third, we also designed K-nearest neighbors (KNN) and support vector machine (SVM) based classifiers for CHB and HCC with high prediction accuracy. The results were validated in an independent data set.
| The hepatitis B virus (HBV) is a DNA virus infecting around 257 million people worldwide (http://www.who.int/mediacentre/factsheets/fs204/en/) and can cause liver diseases and hepatocellular carcinoma (HCC), one of the most common types of liver cancer [1, 2]. About 500,000 HBV patients die each year worldwide from HBV related complications and about 10% of the HBV infected individuals will have HCC during their life time [3]. However, the understanding of the differences of HBV compositions based on next generation sequencing (NGS) technologies between chronic hepatitis B (CHB) and HBV related HCC is limited.
The HBV sequences are currently divided into 10 HBV genotypes, A to J, with genome wide differences of 8%, and 35 subgenotypes using genome wide differences of 4% [3–5]. HBV genotypes have been shown to be associated with geographical locations [6, 7]. In China, the most common genotypes are B and C [8, 9]. Besides, some individuals can be infected by viruses of multiple genotypes and there can be some recombinations among the different genotypes. Different genotypes have varied effects on disease severity, course and likelihood of complications, response to treatment and possibly vaccination [10, 11]. It has been shown that genotype C is associated with more disease complications and higher chance of HCC transition than genotype B [12].
Due to the high mutation rate of the HBV and the possibility of multiple HBV infections, there are high inter- and intra- patient HBV geneticdiversities. Previous studies revealed that basal core promoter (BCP) A1762T/G1764A mutations were strongly associated with the occurrence of HCC [13–16]. Truncated large surface proteins due to deletions in the pre-S gene were observed to accumulate in the endoplasmic reticulum (ER), resulting in ER stress and hepatocarcinogenesis [17, 18]. It was also shown that some pre-S deletions or mutations were risk factors for the development of liver cirrhosis and HCC [19–22]. Meta-analysis studies indicated that pre-S deletion mutations and BCP double mutations were associated with HCC risk [13, 23–25]. Several studies have found that combination of mutations in the HBV genome could predict HCC occurrence more accurately than individual mutations [26–28].
Traditionally, only the dominant genotypes and haplotypes within the patients were investigated due to the technological limitations of Sanger sequencing that are usually time consuming and economically expensive to sequence a large number of sequences within individuals. With the development of high-throughput NGS technologies, it is now possible to investigate the HBV genetic diversity within individuals carefully and to develop more sophisticated and robust prediction models for predicting HCC.
In this study, we aim to explore the diversity of HBV pre-S sequences within HCC and CHB patients, to identify their differences, and to establish prediction models for HCC with machine learning methods based on word pattern frequencies. In detail, we first carried out a large scale HBV pre-S region study of 94 HCC patients and 45 chronic HBV infected individuals. The heterogeneity of HBV composition and the HBV genotype fraction in individuals were investigated. We used a novel alignment-free method based on word pattern frequencies to cluster the individuals and investigated the cluster distributions of HCC patients and CHB individuals. We further applied K-nearest neighbors (KNN) and support vector machine (SVM) approaches to predict HCC status based on word counts and the predictive model was validated using an independent data set consisting of 46 HCC patients and 31 CHB individuals. The key novelties of this study are the use of word patterns for the analysis of HBV sequences to cluster HBV infected individuals and to predict HCC status. Our study clearly showed the surprising high power of word patterns for clustering HBV genotypes and predicting HCC status.
We genotyped each sequence in the NGS data using STAR [49] and calculated the fraction of genotypes B and C sequences for every individual as described in the “Materials and methods” section. The fraction of recombinants in 95% of the individuals (132/139) was less than 5% and most of the reads were of genotype B or C (S1 Supplementary material S1 Fig). Therefore, we ignored the recombinant reads and the reads of other genotypes and concentrated on the reads of genotype B or C in all the individuals. The histograms of the fractions of genotype B sequences among the 94 HCC patients and 45 CHB individuals are given in Fig 1(A). It can be seen from the figure that most individuals have both genotypes B and C sequences for both HCC and CHB individuals. The fraction of genotype B sequences among HCC patients has a tendency to be lower than that for the CHB individuals, consistent with previous observations that genotype C individuals are more likely to have HCC than genotype B individuals [29]. About 70% of the HCC patients have genotype B fraction less than 30% and only about 50% of the CHB patients have genotype B fraction less than 30%. While about 37% of the CHB individuals have genotype B fraction at least 70%, only about 5% of the HCC patients have genotype B fraction at least 70%.
Based on our data, we further investigated the relationship between having HCC and the fraction of genotype B in an individual. It can be shown that the probability of having HCC for given genotype B fraction increases with the ratio of fraction of individuals having the given genotype B fraction among HCC patients over that of CHB patients. Therefore, we binned both the HCC and CHB individuals according to the genotype B fraction. For each bin, we calculated the fractions HCC and CHB individuals and then calculated their ratio as shown in Fig 1(B). When the number of occurrences in a bin was small, the estimated fraction was not reliable. Thus, we required that the fractions for both HCC and CHB in each bin to be at least 5%. If either the HCC fraction or the CHB fraction in an interval was smaller than 5%, we merged it with the later intervals until both fractions were above 5%. Therefore, we merged the bins 0.3~0.4, 0.4~0.5, and 0.5~0.6 into one bin when we calculated the ratio of the fractions. Similarly, we merged the bins 0.7~0.8, 0.8~0.9 and 0.9~1.0 to form another bin. As we can see from Fig 1B that this fraction is higher than 1.0 when the fraction of genotype B sequences is less than 0.6, while it is much less than 1 when the fraction of genotype B sequences is above 0.6.
To see how genotyping method would affect the results, we also used another genotyping program, jpHMM [30], to genotype the reads. The histogram of the fraction of recombinant reads for the 139 individuals is shown in FigS2a in the S1 Supplementary material. The fraction of genotype B using jpHMM is highly associated with that based on STAR (Pearson correlation coefficient = 0.9968 and p-value = 1.0e-151) as shown in FigS2b) in the S1 Supplementary material. FigS3 in S1 Supplementary material shows a similar figure as Fig 1 when jpHMM was used for genotyping. Again we see that the probability of having HCC increases with the fraction of genotype C sequences based on jpHMM.
Based on the word pattern frequencies of the NGS reads from the HBV pre-S region for the individuals, we used Manhattan distance to calculate the dissimilarity between any pair of individuals. We then used principal coordinate analysis (PCoA) to project the individuals onto two-dimensional Euclidean space. Fig 2A and 2B show the PCoA results for the 94 HCC patients and 45 CHB individuals using word length k = 6 and k = 8, respectively.
To see the relationship between the PCoA results and the fraction of genotype B or C in the NGS data of the HBV pre-S sequences, we colored the points corresponding to the individuals according to the fractions of B and C genotypes with red indicating 100% genotype B and blue indicating 100% genotype C with intermediate color in between based on the STAR genotyping results. We also downloaded the HBV genotypes B and C reference sequences from NCBI (accession number of genotype B: D00329, AB073846, AB602818; genotype C: X04615, AY123041, AB014381) and used the pre-S region to serve as references. We counted the occurrences of word patterns of these sequences, calculated their dissimilarity with the 139 samples, and plotted the 141 samples in the PCoA figure. We have several observations from Fig 2. First, the fraction of genotype B sequences in each individual is highly associated with the values of the first principal coordinate. From left to right of the figures, the fraction of genotype B sequences increases with the first coordinate. To see this pattern more clearly, we plotted Fig 2C and 2D that show the relationship between the first principal coordinate and the fraction of genotype B using k = 6 and k = 8, respectively. The Pearson correlation coefficient (PCC) between the fraction of genotype B sequences and the first principal coordinate is as high as 0.97 when k = 6 and k = 8. Second, the HCC tumor samples are distributed more broadly on the PCoA plots and are more diverse than the CHB individuals. The second principal coordinate seems to be associated with the HCC status with high second PCoA coordinate indicating high probability of HCC. Although the second principal coordinates for most of the CHB individuals are at similar levels as for the reference genotypes B and C sequences, many HCC samples have much higher second principal coordinate. To see the pattern more clearly, we divided the second coordinate into 5 bins: < −0.15; −0.15~−0.1; −0.1~−0.05; −0.05~0; > 0. In each bin, we calculated the fractions of CHB and HCC individuals in the bin. We also calculated their ratio and plot the relationship between the ratio and the second coordinate in Fig 2E and 2F. It can be seen that when the second coordinate is smaller than -0.1, the fraction of CHB individuals dominates and with the increase of second coordinate, the fraction of HCC individuals increases. When the second coordinate is bigger than 0, there are no CHB individuals. On the other hand, some of the HCC patients and CHB individuals mix together in the principal coordinate plots and there is no clear separation for HCC patients and CHB individuals. The above conclusions are consistent for both k = 6 and k = 8.
Fig 2 shows that the first principal coordinate is highly associated with the fractions of genotype B(C) when intuitively choosing k = 6 and k = 8. Therefore, we chose the word length k to maximize the correlation. Table 1 shows the Pearson and Spearman correlations between the first principal coordinate and the fraction of genotype B sequences for word length k ranging from k = 2 to k = 8.
Both the Spearman and the Pearson correlation coefficients increase with word length k. When k ≥ 6, the PCC becomes stable. Note that for k = 6 the correlation is already very high and considering computational efficiency, we use k = 6 to show our results on the training data in the rest of the paper.
In addition to the PCoA plots, we also grouped the individuals using hierarchical clustering with UPGMA (Un-weighted Pair Group Method with Arithmetic Mean) to calculate the distance between two clusters. We used the distance matrix calculated from Manhattan distance with k = 6 and input it into the software Mega (http://www.megasoftware.net/). Fig 3 shows the clustering results and the genotypes are analyzed using STAR. The corresponding results using jpHMM are given in S1 Supplementary material Fig S5. The individuals are generally divided into two main clusters. Cluster I contains 44 individuals, 38 of them with dominant genotype B and cluster II contains 95 individuals, 94 of them with dominant genotype C. The overlaps between the two clusters and groups of individuals with genotypes B or C are given in Table 2. The clusters are significantly associated with the dominant individual genotypes (p-value = 2.2e-16, χ2-test). Six individuals (HCC1, HCC13, HCC83, HCC84, HCC88, and HCC102) out of 101 (76HCC+25CHB) with dominant genotype C belong to cluster I. Their corresponding fractions of genotype B are 0.49, 0.49, 0.18, 0.27, 0.14, 0.29, respectively. On the other hand, only one individual (CHB60) out of 38 (18HCC+20CHB) with dominant genotype B belong to the second cluster and its fraction of genotype B is 0.59. We can see that the mis-clustered individuals are highly mixed, and their secondary genotypes also have relatively high fraction. The normalized fractions of genotypes B and C sequences of all individuals using STAR and jpHMM are given in the S1 Table.
Within cluster I, there is a small sub-cluster Ia that is dominated by CHB individuals. On the other hand, the HCC patients and CHB individuals are not clearly separated in cluster I. Within cluster II, a small cluster IIa is dominated by CHB individuals and the HCC patients are generally far away from this group. The results from the hierarchical clustering of the individuals are consistent with the observations based on PCoA results.
We noticed 11 CHB patients within the large cluster IIb that contains mostly HCC patients. Therefore, we checked the meta-data to see if these 11 individuals had high risk factors for HCC including liver cirrhosis, advanced age, male sex, etc. Six out of the 11 CHB patients in cluster IIb had meta-data available. Five patients (CHB46, CHB48, CHB50, CHB60, CHB91) are male and one is over 60. Patient CHB55 is female, who has liver cirrhosis and was over 60 years old. Thus, our meta-data do show that these patients have more risk factors.
We also colored the points in the PCoA plots corresponding to the individuals according to the fractions of B and C genotypes with red indicating 100% genotype B and blue indicating 100% genotype C with intermediate color in between based on the jpHMM genotyping results, and the corresponding figure is shown as FigS4 in the S1 Supplementary material. Similar observations as based on STAR genotyping were obtained. Table S2 in the S1 Supplementary material shows again that the first principal coordinate is highly associated with the fraction of B genotypes in an individual, consistent with the results using the STAR genotyping tool.
We used two methods, K-nearest neighbors and support vector machine (SVM), to predict HCC status based on the word pattern frequency vector of the HBV pre-S region of the samples.
The prediction results based on KNN are given in Table 3. It can be seen from the table that the cross validation results measured by AUC are roughly the same with different word length k and the AUCs center around 0.88. For the independent test data, the AUC increases slightly with the word length from 0.62 for k = 2 to 0.67 when k is between 6 and 8.
The AUC values of SVM using cross validation and testing set and corresponding parameter C using different word length k are shown in Table 4. We observe from the table that the prediction accuracy measured by AUC with cross-validation increases slightly with word length from 0.86 when k = 2 to 0.93 when k = 7. On the other hand, the AUC for the independent data decreases with word length from 0.77 when k = 3 to 0.70 when k = 8. When k = 2, the AUC is only 0.65. The good performance of the SVM model when k = 3 may be due to the relatively small number of learning samples such that the derived SVM model with small number of word patterns is more stable.
Several recent studies have clearly shown the advantage of NGS over traditional Sanger sequencing in detecting rare HBV sequence mutations [15] and for the prediction of anti-virus therapy response [31, 32]. In this study, we used high throughput sequencing to investigate composition of HBV sequences in a large number of both CHB and HCC individuals, to compare differences of genetic composition between them, and to predict HCC status using novel word pattern based approaches. Several interesting results were obtained. First, we showed that there was extensive heterogeneity of HBV composition among the individuals based on the NGS data. Almost all the individuals contain some marked fractions of both genotype B and genotype C HBV sequences in Chinese individuals infected with HBV. Previous studies have shown the existence of co-infection of different genotypes of HBV [33–35] and inter-genotype HBV co-infection is the prerequisite of HBV recombination incidence that have been reported broadly [36–38]. Our results highlight the importance of using NGS to study the distribution of different genotypes within individuals.
Second, we used a novel word pattern based approach to cluster the individual samples and investigated the cluster distributions of HCC patients and CHB individuals. Alignment-free sequence comparison based on word counts has been widely used in studying the relationships among sequences or NGS data as reviewed in [39, 40]. However, this approach has not been used for the analysis of HBV data. In this paper, we used alignment-free sequence comparison methods based on word counts to study the relationship among the individuals. We used a dissimilarity matrix based on Manhattan distance between the word frequencies of the NGS data to cluster all the individuals. We showed that there was a strong correlation between the clustering and the fractions of genotypes (B or C) of individuals. This observation was surprising and proved the effectiveness of the alignment-free method on classification based on sequence dissimilarity.
Third, since the second coordinate of PCoA was remarkably correlated with the probability of having HCC, we further applied K-nearest neighbors (KNN) and support vector machine (SVM) approaches to classify HCC or CHB individuals based on word counts. Using cross-validation, we achieved a high area under the receiver operational characteristic curve (AUC) of around 0.88 for KNN and 0.92 for SVM for word length from 4 to 8.
Fourth, we validated the prediction models on an independent set of 46 HCC patients and 31 CHB individuals. The AUC for the independent set was around 0.70 when word length is from 6 to 8 for SVM and 0.67 for KNN. Surprisingly, the AUC for SVM was 0.77 when word length is 3. The good result of k = 3 may be explained by the appropriate number of features compared with the number of individuals. The results showed the usefulness of our prediction models for separating HCC patients from CHB individuals. Numerous studies have revealed the divergence in pre-S region between CHB and HCC patients and deletions in pre-S was one of the most noticeable characteristic of HCC patients [41–44]. In addition, fewer studies also found that several nucleotide mutations were also associated with incidence of HCC [19, 45, 46]. Nevertheless, we have succeeded in the establishment of predictive model for HCC via the word pattern frequencies of the pre-S gene following the NGS. The superior performances in both the cross validation and independent cohort validation are also indicative of the advantages of NGS compared with Sanger sequencing.
Finally, we showed that the HCC status can be effectively predicted based on word pattern frequencies using support vector machine and that prediction accuracy increases with the number of reads and becomes stable at about 3000 reads per individual. To our knowledge, this is the first study focusing on the implication of the number of reads on model effectiveness trained on NGS data. With the development of NGS technology, investigators are interested in appropriate number of reads and our study provides guidelines for designing of NGS studies.
Despite these significant results, our study has several limitations. First, the numbers of HCC and CHB individuals, although large compared to previous studies, were still not very large and more individuals are needed to further confirm the applicability of our word pattern based method for investigating HBV infected individuals. Second, the AUC values for the independent test data using both KNN and SVM were much smaller than the corresponding mean AUC values for cross-validation. Potential explanations for the lower AUC value for the independent test data is that the independent samples may come from populations different from that in the training data. Potential experimental variations from the testing data may also decrease the prediction accuracy. Third, we concentrated on the HBV pre-S region in this study and other regions may have different properties. Further studies for other regions or even the whole genome are needed. Fourth, we investigated Chinese HCC and CHB individuals with dominant B and C genotypes. The applicability of our results to other ethnic groups or population samples needs to be further investigated.
In conclusion, our study showed the applicability of word pattern based methods to investigate the diversity of HBV sequences, to compare HBV communities among different individuals, and for the prediction of HCC status. Further studies are needed to extend the results to much larger genomic regions over large number of individuals.
HBV were divided into ten major genotypes A to J with the dominant genotype B or C in China. Merged pre-S region sequences were genotyped with HBV STAR software [49] that is one of the most widely used software tools for HBV genotyping [50–52]. It is based on a statistically defined, position-specific scoring model (PSSM) [53]. Even though our sequence reads are relatively short compared to the whole genome, it has been shown that any 300 bps sequence segment of the polymerase N-terminal domain containing pre-S is reliable for sequencing-based HBV genotyping [54]. STAR [49] uses all the known HBV sequences with known genotypes to construct a PSSM for each genotype A to H (I and J are not well understood) and then scores each read with respect to each genotype to have eight scores. We further transformed the scores into Z scores as in [49]. As recommended in [49], if the maximum score of a read was above 2.0, we predicted the genotype of the read as the one yielding the highest Z score. If the maximum score was below 2.0, STAR uses a slide window of 150bps to find the genotype for each window. We considered the reads with Z score below 2.0 and having windows with distinct genotypes as recombinant reads.
Consistent with the fact that the dominant HBV genotypes are B and C in China, over 95% of the reads are of the two genotypes or recombinants of B and C for all the samples with some small fractions of genotype A. The fraction of recombinant reads for 95% of the samples (132/139) was less than 5%, and only 3 samples had the fraction of recombinant reads above 20%. Therefore, we ignored the fractions of other genotypes and recombinant reads, normalized the fractions of B and C to sum to 1, and calculated the fraction of genotypes B and C, respectively, for each sample.
In addition to STAR, we also used another program jpHMM [30] for the identification of recombinant reads in NGS reads to see how different programs will affect our results. jpHMM uses a jumping hidden Markov model to identify recombinant reads between different genotypes. For each read, it identifies regions corresponding to a particular genotype. We defined a read to be a non-recombinant if a consecutive region of at least 400bps belongs to the same genotype while only at most 57bps belong to different genotypes. The details were given in the S1 Supplementary material section 2.
For each individual, we counted the number of occurrences of any word pattern of length k (also called k-tuples, k-mers, k-grams) in the NGS data. The relative frequency of a word of length k was its count divided by the total counts of all the words of length k for the individual. The distance between any pair of individuals was measured by the Manhattan distance between their corresponding frequency vectors. We constructed a distance matrix of all samples from the training set to see how the individuals cluster together. We chose the Manhattan distance because previous studies showed that it gave better clustering results than Euclidean distance for the clustering of genome sequences in many applications [55]. For different values of k, we used principal coordinate analysis (PCoA) to project the data onto two-dimensional space to see how the individuals group together. The basic idea of PCoA was to represent the data in the low dimensional space so that the distances between the samples in the low dimensional space are as close as possible to their true distances. In addition, we hierarchically clustered the individuals based on their word pattern frequencies. We used UPGMA to calculate the distance between any two clusters as the average of all the pairwise distances between the pairs of individuals from both clusters.
We investigated the optimal approaches for predicting HCC status from the word pattern frequencies. Based on the PCoA and hierarchical clustering results, it can be seen that if the word pattern frequency vector of an individual is similar to others having HCC status, the individual is more likely to have HCC. Therefore, we first used the K-nearest neighbors (KNN) algorithm to predict HCC status, where K is the number of neighbors used for prediction. In KNN, an individual is predicted as having HCC if the fraction of HCC individuals among the top K most similar individuals according to word pattern frequency is above a threshold. We also used supporting vector machine (SVM) to predict HCC status using word pattern frequencies as features. For SVM, we had several kernel functions and parameters to choose from. We used linear kernel only here because for most cases it can work well and it has only one parameter C. For the parameter C, we used cross validation within the training set to choose C yielding the highest AUC (area under the receiver operational characteristic curve) value and used the parameter to construct a model for predicting the testing set.
|
10.1371/journal.ppat.1002848 | A Candida Biofilm-Induced Pathway for Matrix Glucan Delivery: Implications for Drug Resistance | Extracellular polysaccharides are key constituents of the biofilm matrix of many microorganisms. One critical carbohydrate component of Candida albicans biofilms, β-1,3 glucan, has been linked to biofilm protection from antifungal agents. In this study, we identify three glucan modification enzymes that function to deliver glucan from the cell to the extracellular matrix. These enzymes include two predicted glucan transferases and an exo-glucanase, encoded by BGL2, PHR1, and XOG1, respectively. We show that the enzymes are crucial for both delivery of β-1,3 glucan to the biofilm matrix and for accumulation of mature matrix biomass. The enzymes do not appear to impact cell wall glucan content of biofilm cells, nor are they necessary for filamentation or biofilm formation. We demonstrate that mutants lacking these genes exhibit enhanced susceptibility to the commonly used antifungal, fluconazole, during biofilm growth only. Transcriptional analysis and biofilm phenotypes of strains with multiple mutations suggest that these enzymes act in a complementary fashion to distribute matrix downstream of the primary β-1,3 glucan synthase encoded by FKS1. Furthermore, our observations suggest that this matrix delivery pathway works independently from the C. albicans ZAP1 matrix formation regulatory pathway. These glucan modification enzymes appear to play a biofilm-specific role in mediating the delivery and organization of mature biofilm matrix. We propose that the discovery of inhibitors for these enzymes would provide promising anti-biofilm therapeutics.
| Biofilms are a community of microbes that grow attached to each other and adherent to a surface. One distinguishing feature of this form of growth is the presence of a surrounding extracellular matrix which is proposed to provide a structural scaffold and protection for biofilm cells. This later function contributes to the extreme resistance to anti-infective therapies, another innate characteristic of biofilms. One carbohydrate component of the matrix of Candida albicans, β-1, 3 glucan, has been linked to overall accumulation of matrix material and the antifungal drug resistance phenotype. Although the glucan synthase pathway has been implicated in glucan production, the delivery and incorporation of these carbohydrates into the matrix remains a mystery. The current investigation describes three gene products that serve a matrix delivery role. The functions of these gene products include glucanase and glucanosyltransferase activities. Mutants unable to produce these enzymes demonstrate reduced matrix glucan, decreased total matrix biomass accumulation, and enhanced susceptibility to antifungal drug therapy. The observations here offer insight into a novel pathway that contributes to biofilm maintenance. Enzymes in this biofilm-specific process may provide useful anti-biofilm drug targets.
| Candida spp. are an increasingly common cause of bloodstream infections in hospitalized patients [1], [2]. This rise in incidence is at least in part related to the organism's ability to produce biofilm infections on medical devices [3]. A biofilm is a community of microbes attached to a surface and encased in an extracellular matrix [4]–[6]. The biofilm lifestyle is a common form of growth in nature and the most common cause of infection in humans. The most troublesome characteristic of biofilms is that they are up to 1,000-fold more resistant to common antifungals than their planktonic counterparts, even without accumulation of specific drug-resistance genes [7]–[10]. This lack of effective therapy contributes to dismal outcomes for patients with invasive candidiasis, including death in up to 40% of patients. Delineating the mechanisms of biofilm formation and associated treatment resistance is one method of identifying optimal management strategies and therapeutics of this devastating infectious disease.
The focus of our investigations is the construction and configuration of the extracellular biofilm matrix, one of the properties that distinguishes biofilm from planktonic growth [11]. The function of matrix remains incompletely understood, but previous investigations have identified roles such as providing infrastructure for biofilm accumulation, controlling disaggregation, and granting protection from antimicrobial drugs and the host immune system [12]–[14]. Although the complete composition of the C. albicans biofilm matrix has yet to be fully elucidated, studies have identified the inclusion of carbohydrates, proteins, and nucleic acids components [11], [13], [15]. The goal of the present studies was to identify genes that control matrix delivery. We hypothesized that this process involves a biofilm-specific pathway composed of enzymes capable of modifying matrix carbohydrates. This hypothesis is based on two findings. First, one of the carbohydrates, β-1,3 glucan, has been linked to overall matrix production and drug resistance through glucan synthase gene FKS1 (common nomenclature for the gene GSC1) [16], [17]. Second, microarray analysis of in vivo rat catheter biofilms demonstrated transcript abundance of multiple glucan modification genes [18].
Here we use a candidate gene set to investigate the role of glucan matrix delivery. The gene set was selected to include glucan modification genes which demonstrated transcriptional upregulation in a rat venous catheter biofilm model. In addition, we included gene products which are known or hypothesized to utilize β-1,3 glucan as a substrate [19]–[25]. Many of the selected genes had been shown previously to function in planktonic cell wall synthesis and remodeling [23]–[30]. We constructed gene mutants and screened for biofilm formation, matrix delivery and antifungal drug susceptibility.
In the current studies we describe the role of three glucan modifying genes for glucan delivery and matrix incorporation. These gene products encode two glucanosyltranferases (BGL2, PHR1) and a glucanase (XOG1), respectively [22], [23], [25]–[29]. Each appears necessary for modification and delivery of carbohydrate to the mature biofilm matrix. Without delivery and accumulation of matrix glucan, the biofilms exhibit enhanced susceptibility to antifungal drugs. As the biofilm matrix is integral for biofilm maintenance and drug resistance, these delivery enzymes provide promising targets for anti-biofilm drug development.
We have previously described the presence of β-1,3 glucan in the biofilm matrix of C. albicans and identified the role of the glucan synthase pathway for production of this material [16], [17], [30]. The machinery needed for delivery of this matrix component from the cell to the matrix was, however, not known. We reasoned that proteins that act upon a glucan substrate might contribute to the delivery process. Results of an in vivo microarray analysis of a rat venous catheter biofilm demonstrated differential expression of 11 potential glucan modification genes [18]. A candidate gene set was constructed by combining these 11 genes with 4 additional genes selected from a search of the Candida genome database for putative glucan modifying function (glucanases, transferases, and glucosidases). A combination of homozygous deletion mutants were created for fourteen genes and a heterozygous mutant for one gene presumed to be essential (Table S1 in Text S1). Our initial experiments consisted of two screens. First, we examined overall biofilm growth in all strains. Each of the mutants produced mature in vitro biofilms similar to reference strains, with the exception the phr1−/− strain which exhibited a modest biofilm defect (75% cell burden compared to the reference strain). The phr−/− strain also demonstrated a modest defect in adhesion to a polystyrene substrate (67% relative to the reference strain). The mutant strains exhibited normal planktonic growth in YPD compared to the reference strain. Secondly, we measured the β-1,3 glucan concentrations in the matrix from mature in vitro biofilms using both the commercial limulus lysate assay (Glucatell) and a glucan ELISA. These assays identified three deletion mutants, bgl2−/−, xog1−/−, and phr1−/−, which produced up to 10-fold less matrix β-1,3 glucan than the reference biofilm (Table 1 and Figure 1A). The observations were confirmed in independent transformants for each gene (Table 1). Furthermore, complementation of the mutants with a single copy of each gene restored glucan matrix concentrations to reference strain levels.
The relevance of these glucan modification genes to in vivo biofilm growth is suggested by their transcriptional abundance in a rat venous catheter biofilm [18]. At 12 h of in vivo biofilm growth, microarray studies showed that transcription of BGL2 and PHR1 was upregulated. During mature biofilm growth (24 h), BGL2 and XOG1 transcripts were abundant. RT-PCR confirmed marked increases in expression during biofilm growth (Table 1). We asked if these glucan modification enzymes were functioning in conjunction with the previously described Zap1-regulated matrix production [31]. This zinc transcription factor is a negative regulator of biofilm matrix production, including matrix glucan production. Surprisingly, these glucan modification enzymes appear to function independently of Zap1. First, transcription of BGL2, XOG1, or PHR1 was not significantly altered in the zap1−/− mutant biofilm. Second, there were no significant changes in ZAP1 transcription in the glucan modification mutant biofilms (data not shown). These findings suggest that BGL2, XOG1, and PHR1 comprise a distinct biofilm matrix delivery pathway.
The mutants with reduced matrix glucan (bgl2−/−, xog1−/−, and phr1−/−) were evaluated for biofilm architecture, matrix appearance, and total matrix abundance by scanning electron microscopy of in vitro biofilms. These glucan modifying enzyme mutants were capable of biofilm formation, but exhibited diminished extracellular biofilm material (Figure 1B). The association between reduced glucan and total matrix biomass is similar to that described for mutants in the β-1,3 glucan synthesis pathway [17], [30]–[33].
Since β-1,3 glucan has been described as a matrix component, we considered the possibility that the glucan in the matrix may also impact biofilm persistence or resistance to disaggregation. To test the functional role of matrix glucan, and the impact of glucan matrix delivery, we examined biofilm cell disaggregation in the reference strain and this subset of glucan modifying enzyme mutants following exposure to low concentrations of β-1,3 glucanase. Previous studies in this model have shown that higher concentrations of this enzyme will disperse intact mature biofilms [16]. In the present investigation, exposure to a low concentration of β-1,3 glucanase resulted in disaggregation of approximately 25% of the reference biofilm (Figure 1C). However, the same glucanase incubation allowed dispersion of approximately 80–90% of the glucan modifying mutant biofilms. These observations argue that matrix β-1,3 glucan provides an adhesive function within the biofilm matrix. The disaggregation findings are also consistent with the matrix biochemical and imaging observations showing less matrix β-1,3 glucan and total matrix biomass.
A previously demonstrated link between matrix glucan and drug resistance led us to test the impact of these glucan modifying enzymes on this important biofilm phenotype [14], [16]. Each of the fifteen glucan modifying mutants in the candidate gene set was screened for susceptibility to the triazole, fluconazole, during in vitro biofilm growth (Table 1 and Table S1 in Text S1). The three glucan modifying mutants that delivered less matrix glucan exhibited enhanced susceptibility to fluconazole. Although the highest concentration of fluconazole resulted in no net change in cell burden in reference biofilms, this same drug exposure reduced the mutant biofilms by 35 to nearly 70% (Table 1 and Figure 2A). A dose dependent anti-biofilm effect was observed over the entire dose range examined (not shown). These findings were confirmed for the independent transformants (Table 1). Furthermore, the biofilm-associated antifungal resistance was restored in complemented strains (Figure 2A). Deletion of the three glucan modifying genes did not cause a significant change in planktonic antifungal drug susceptibility (Table 1), so we infer that this is a biofilm-specific phenotype.
As drugs from the echinocandin class target β-1,3 glucan synthesis, we further examined the impact of these select glucan modification mutants on biofilm susceptibility to a drug from this class. Each of strains (parent and the three mutants) demonstrated extensive susceptibility to low echinocandin concentrations (<0.03 µg/ml). No difference in drug activity was observed over the range of concentrations examined.
In order to determine the clinical relevance of these observations, we examined drug susceptibility using the in vivo rat central venous catheter biofilm model [34]. The impact of the fluconazole treatment was tested by measuring the viable burden of biofilm cells present following a twenty-four hour period of exposure to the drug instilled within the catheter lumen. Drug treatment produced minimal change in biofilm burden in the reference strain. In vivo study of the glucan modifying mutants recapitulated the observations from the in vitro model. The burden of catheter associated cells was reduced by 1.5 to more than 2 logs compared to the reference strain (Figure 2B).
Earlier studies suggest that the mechanistic basis underlying the glucan matrix associated resistance phenotype is due to sequestration of antifungal by the matrix material away from the drug's cellular target [14]. We tested the biofilm sequestration capacity of the reference strain and the subset of glucan modification mutants, bgl2−/−, xog1−/−, and phr1−/− (Figure 2C). Each of the mutants sequestered less radioactive fluconazole than the reference strain, with the greatest defect observed for the phr1−/− biofilm (nearly 4-fold). The mechanistic reason for differences among the glucan modifying strains is not clear and is clearly an interesting area for further inquiry. This finding further links the glucan modifying enzymes and matrix glucan deposition to biofilm drug resistance.
Understanding the function of this subset of glucan modifying enzymes, Bgl2, Xog1, and Phr1, in cell wall construction and maintenance remains incomplete. We hypothesized that the cell wall of mutant strains with reduced matrix β-1,3 glucan may exhibit similar glucan reductions in the cell wall. Previous studies in a phr1−/− mutant show altered cell wall glucan and chitin content during planktonic growth [29]. We were surprised to find similar cell wall β-1,3 and 1,6 glucan content among the biofilm cells of this subset of glucan modifying mutants and the reference strain (Figure 3A). These results support a model in which the individual modification enzymes are dispensable for cell wall glucan production during biofilm growth, but are required for delivery of glucan from the cell to the extracellular matrix. The difference between the PHR1 cell wall results in the planktonic and current biofilm studies further underscore a novel, biofilm specific role for this gene product.
Light microscopy and transmission electron microscopy (TEM) were used to inspect the mutant cell wall phenotypes. Light microscopy of the cells demonstrated a previously described abnormal hyphal morphology in the phr1−/− strain (data not shown) [35]. However, the other mutants appeared similar to the reference strain. By TEM, the yeast cell walls for each of the strains appeared quite similar in thickness and ultrastructural composition, consistent with the carbohydrate composition analyses (Figure 3B). The relative thickness of the cell wall of at least 50 images from each strain was quantified using ImageJ software. The average cell wall thickness for each strain was not significantly different from the reference strain (data not shown).
A parallel study of cell wall function was performed to assess the potential impact of the glucan modifying genes on the cell wall integrity pathway that has been shown to contribute to the biofilm formation and drug resistance mechanism [36], [37]. Susceptibilities to β -1,3 glucanase, hydrogen peroxide, SDS, and calcofluor white were similar among the bgl2−/−, xog1−/−, and the reference biofilms (Figure 3C). The phr1−/− strain exhibited hypersensitivity to β -1,3 glucanase and calcofluor white, and a relative resistance to SDS. The change in susceptibility to calcofluor white in these biofilm experiments is similar to that described for planktonic conditions [38]. These phenotypic screens suggest potential perturbation of the CWI pathway associated with PHR1 disruption, but we did not detect a similar signal for the other glucan modifying mutants.
The β-1,3 glucan synthase has been shown as necessary for β-1,3 glucan production and development of biofilm matrix [16], [39]. We theorized that one or more of the glucan modification enzymes acts upon the β-1,3 glucan product of the synthase enzyme in a tightly controlled glucan delivery and matrix maturation of pathway. To explore this hypothesis we examined transcript abundance of the glucan synthase, FKS1, in the glucan modifier mutants. We reasoned that reduced delivery of glucan to the matrix may signal the cell to produce additional β-1,3 glucan which would be marked by an increase in the FKS1 transcript. The FKS1 mRNA abundance results were consistent with the theory that matrix glucan levels influence the cell glucan production machinery. Transcript levels were elevated more than 1.5-fold in each of the modifier mutants (Figure 4A).
Additional testing of these relationships included a functional study of the impact of overexpression of the glucan modification genes, BGL2, XOG1, and PHR1 in the FKS1−/+ heterozygote. This strain produces less matrix glucan and exhibits a biofilm antifungal drug susceptible phenotype [32]. We theorized that if the glucan modifier enzymes act upon the glucan product of Fks1p for matrix delivery, then overexpression of the modifiers would not repair the drug susceptibility defect in the FKS1−/+ background. Indeed, the overexpression of the glucan modifiers did not restore the wildtype biofilm resistance phenotype (Figure 4B). In a complementary experiment, we also examined the impact of overexpression of FKS1 in the glucan modifier null−/− background. These manipulations restored the antifungal resistance phenotype to each of the modifier deletion mutants (Figure 4C).
One simple explanation for these observations is a model in which the glucan modification enzymes provide complementary activity. Studies in the last several years have taught us that redundancy in the biofilm formation process is a common theme for other important functions, such as adherence [40], [41]. A second interpretation of the findings is a paradigm in which the glucan synthesis and modification pathways are distinct with regard to the biofilm matrix resistance mechanism. The suggestion of complementary activity for matrix delivery and drug resistance was further investigated by overexpression analysis of the glucan modifier genes in companion deletion mutant backgrounds and double knockout strains. We successfully introduced a PHR1 overexpression allele into the bgl2−/− strain, and introduced a BGL2 overexpression allele into xog1−/− and phr1−/− strains. We were unable to successfully introduce a XOG1 overexpression allele in the bgl2−/− or phr1−/− strains. Similarly, we were unable to introduce the PHR1 overexpression allele in the xog1−/− background. Biofilm susceptibility testing demonstrated restoration of the drug resistance phenotype associated with overexpression of a companion glucan modifier in the glucan modification mutant background for all strains tested (Figure 4D). These results are similar to those observed for in the glucan synthase mutant. We infer that the findings suggest a complementary relationship among the glucan modifiers.
Additional examination of the association among the glucan modifiers included testing the impact of mutants in which two modifiers were disrupted. We were unable to construct the double knockout xog1−/−, phr1−/−, suggesting that loss of both of these genes may result in a non-viable mutant. For unclear reasons, the constructed double knockouts (xog1−/−, bgl2−/− and bgl2−/−, phr1−/−) demonstrated a growth defect in RPMI-MOPS under both biofilm and planktonic conditions, such that no biofilm could form in RPMI. They were, however, able to adhere to plastic and produce filaments in response to increased temperature when grown in YPD (Figure 5A). These double knockout strains also exhibited normal planktonic growth in YPD (Figure 5B). Therefore, we adapted the XTT biofilm drug susceptibility assay to include YPD media for comparison of double mutant and parent strains. In this assay, both of the double knockouts (bgl2−/−, phr1−/− and xog1−/−, bgl2−/−) demonstrate increased susceptibility to fluconazole when compared to their single modifier knockout parent strains (Figure 5C). While these strains produced relatively normal biofilms in the 96 well format, similar study with YPD in the larger format utilized for matrix composition analysis was insufficient in these strains. Thus, we were unable to reliably compare matrix glucan content. Although the observed RPMI growth defects and assay modification are limitations, the experiments suggest that deletion of two modification genes results in a further decline in matrix delivery. These findings support the theory that the modifiers act in parallel and can partially compensate for each other.
The extracellular matrix is critical for mature biofilm formation [42]. This material not only contributes to the adhesive nature of biofilm cells, but has been shown to protect the cells from antimicrobial agents and the host immune system as well [12], [30], [33], [43], [44]. Understanding the matrix components' production and delivery processes is one path for the development of effective biofilm therapies. A key constituent of the C. albicans matrix is β-1,3 glucan [16], [31]. Previous work identified an increase in cell wall glucan associated with biofilm growth [16]. Subsequent observations demonstrated the importance of the glucan synthase pathway for production of β-1,3 glucan in both the cell walls of biofilm cells and the extracellular matrix [32]. The predominant β-1,3 glucan synthase in C. albicans is encoded by FKS1 [45]. Both the MAP kinase pathway and the transcription factor ZAP1 have been identified as upstream components of the biofilm matrix production process [30], [31], [46]. However, the process of delivering glucan from the cell wall and the resulting mature biofilm matrix accumulation remained unknown. The present findings identify a novel role of several glucan modification enzymes for delivery of matrix glucan and other components to the cohesive extracellular matrix network.
The delivery enzymes from the current screen have been shown or suggested to act upon the β-1,3 glucan substrate. The function of each includes glucan hydrolysis and in some instances transfer and formation of new branch linkages. Previous studies in two unrelated bacterial pathogens, Pseudomonas aeruginosa and Streptococcus mutans, have demonstrated the importance of similar transferase enzymes for delivery of glucan to their biofilm matrices [43], [47]. Our glucan matrix and biofilm antifungal susceptibility screens point to a role for three genes, BGL2, XOG1, and PHR1. BGL2 and PHR1 encode glucanosyltransferases and XOG1 encodes a β-1,3 exoglucanase [22], [27], [29]. Each of these genes has been shown to play a role in cell wall remodeling and specifically glucan chain elongation and cross-linking during planktonic cell growth for both C. albicans and S. cerevisiae [23]–[30]. Interestingly, each of the enzymes shown to impact matrix glucan delivery did not appear to impact the quantity of cell wall ultrastructure or β-1,3 glucan concentration. This suggests that these enzymes function specifically for matrix delivery, distinct from the cell wall assembly pathway during biofilm growth. One exception is PHR1. Disruption of this gene appeared to alter the cell wall integrity pathway during biofilm growth, based upon enhanced susceptibility to cell wall stress by calcofluor white. This observation is similar to that described for planktonic conditions [29]–[30].
Previous investigations found elevated transcript levels of BGL2, XOG1, and PHR1 during in vivo biofilm growth compared to planktonic growth [18]. This biofilm associated upregulation is consistent with a role in a biofilm-specific function, such as matrix formation. The current studies identify a biofilm-specific pathway for these enzymes involving matrix delivery. One proposed mechanism is that the enzymes release and modify cell wall glucan for deposition in the extracellular space. An alternative explanation is that the enzymes act in the extracellular space, contributing to steric changes in glucan that are important for mature matrix organization and function. The enzymes Bgl2, Xog1, and Phr1 have been localized to the cell wall, supporting the hypothesis of cell wall activity. However, Bgl2 and Xog1 also contain secretion sequences providing feasibility for an extracellular function. Phr1 contains a GPI-linked tail, making it more likely to be tethered to the extracellular portion of biofilm cells. Candida biofilm proteomic analysis (our data not shown) identified Bgl2 Xog1, and Phr1 incorporated in the biofilm extracellular matrix consistent with an extracellular role. Enzyme isolation and further structural analysis of matrix components in parent and mutant strains may be an attractive strategy to differentiate between these matrix delivery functions.
The known glucan modification function of these enzymes intuitively supports a model whereby matrix delivery is downstream of the primary β-1,3 glucan synthase encoded by FKS1 (Figure 6). Transcriptional and functional analyses of our target gene overexpression strains support a pathway with partially redundant glucan modification enzymes that link to Fks1. We propose that the overexpression of glucan modifications enzymes is unable to compensate for disruption of FKS1 due to the lack of available glucan substrate. Overexpression of FKS1 partially restores the glucan mutant phenotype by over production of glucan substrate which is processed through parallel pathways of redundant glucan modification enzymes. Data derived from studying the double modifier mutants and the overexpression of modifiers in companion knockout strains also supports a degree of redundancy among the glucan modifiers. Furthermore, upregulation of the FKS1 transcript in each of the enzyme modifier knockouts suggests feedback signaling for the cell to produce additional β-1,3 glucan during biofilm growth in the absence of glucan matrix delivery associated with each glucan modifier mutant.
We considered the possibility that the glucan modification pathway was under control of Zap1, a transcription factor known to function in matrix production. Surprisingly, review of previously reported global expression analysis of ZAP1 did not identify altered expression of XOG1, PHR1, or BGL2. We confirmed the absence of differential mRNA abundance of these select glucan modification enzymes in the ZAP1 mutant (data not shown). Thus, this work has identified a novel matrix glucan delivery pathway that is distinct from the previously described matrix-inhibitory pathway controlled by ZAP1. Figure 6 shows a proposed model of the relationship of these matrix delivery enzymes to Fks1 and Zap1.
These studies show a novel biofilm matrix delivery pathway linked to the drug resistance phenotype. It is intriguing to consider the potential for drug target development designed specifically to identify enzyme inhibitor molecules. Because homologues are not present in the human genome, the likelihood of a safe pharmacologic anti-biofilm agent is promising. Additional work on the signaling and upstream genetic control of these enzymes promises to shed additional light on this important feature of biofilm formation.
All animal procedures were approved by the Institutional Animal Care and Use Committee at the University of Wisconsin according to the guidelines of the Animal Welfare Act, The Institute of Laboratory Animal Resources Guide for the Care and Use of Laboratory Animals, and Public Health Service Policy.
Strains were stored in 15% (vol/vol) glycerol stock at −80°C and maintained on yeast extract-peptone-dextrose (YPD) medium with uridine (1% yeast extract, 2% peptone, 2% dextrose, and 80 µg/ml uridine) prior to experiments. C. albicans transformants were selected on synthetic medium (2% dextrose, 6.7% yeast nitrogen base [YNB] with ammonium sulfate, and auxotrophic supplements), or on YPD plus clonNat (2% Bacto peptone, 2% dextrose, 1% yeast extract, and 400 µg/ml clonNat [Werner Bioagents]) or on YPD plus 70 µg/ml hygromycin B (PhytoTechnology Laboratories). Prior to biofilm experiments, C. albicans strains were grown at 30°C in YPD and biofilms were grown in RPMI 1640 buffered with morpholinepropanesulfonic acid (RPMI-MOPS).
The C. albicans strains used in these studies are listed in Table 1 and the genotypes in Table S2 in Text S1. Homozygous deletion strains were constructed from one of two parent strains, BWP17 or SN152. PCR product-directed gene deletion in the BWP17 background was performed as previously reported [48], [49]. Fusion PCR disruption cassettes were utilized to construct null strains in the SN152 background as previously described [50]. Complementation of mutant strains was performed using selection for arginine prototrophy as previously published [30], [51]. DNA cassettes of the entire gene as well as 1 kb up and downstream were amplified using PCR. The primers were designed to affix a BamHI site to the 5′ end of the DNA cassette and an AscI site to the 3′ end. Because XOG1 had a BamHI cutting site within the gene, it was complemented using two AscI sites instead. Digested PCR products were ligated into the E. coli plasmid pC23, which carries ampicillin resistance for selection and encodes the Candida dubliniensis Arg4. Plasmids were linearized using PmeI and transformed using the lithium acetate protocol. All genotypes were verified by colony PCR using corresponding detection primers. Primers are listed in Table S3 in Text S1.
Overexpression of genes, FKS1, XOG1, BGL2, and PHR1, was accomplished by replacing the endogenous promoter of one allele with the promoter of TDH3, using the plasmid pCJN542 containing the NAT1 – TDH3 gene cassette as described previously [52]. Primers were designed with homology to the plasmid as well as to the promoter region of the targeted gene. This homology allowed for the entire cassette produced from the plasmid (including the NAT1 gene and TDH3 promoter) to be inserted into the promoter region of the gene of interest using homologous recombination, resulting in the gene now being driven by the highly active TDH3 promoter. Transformants were selectively grown on YPD+clonNAT. All genotypes were verified by colony PCR.
Double deletion mutants were created in the SN152 background. The alleles for the first mutant were constructed by sequential replacement with the HygBR and NouR resistance markers, respectively [53]. The second gene was disrupted by replacement of auxotrophic genes as described above [50]. The mutant strains were confirmed by colony PCR. The strain xog1−/− : phr1−/− could not be created.
RNA was collected from biofilm cells grown in 6-well plates, as described below. RNA was purified using the RNeasy Minikit (Qiagen) and quantified using a NanoDrop spectrophotometer. TaqMan primer and probe sets designed using Primer Express (Applied Biosystems, Foster City, CA) for ACT1, FKS1, BGL2, XOG1, and PHR1 are shown in Table S4 in Text S1. The QuantiTect probe reverse transcription-PCR (RT-PCR) kit (Qiagen) was used in an iQ5 PCR detection system (Bio-Rad) with the following program: 50°C for 30 min, initial denaturation at 95°C for 15 min, and then 40 cycles of 94°C for 15 s and 60°C for 1 min. Reactions were performed in triplicate. The expression of each gene relative to that of ACT1 is presented. The quantitative data analysis was completed using the delta-delta CT method [54]. The comparative expression method generated data as transcript fold change normalized to a constitutive reference gene transcript (ACT1) and relative to the reference strain.
Biofilms were grown in 6-well or 96-well flat-bottom polystyrene plates as previously described [51], [55]. The C. albicans inoculum (106cells/ml) was prepared by growth in YPD with uridine overnight at 30°C, followed by dilution in RPMI-MOPS based on hemocytometer counts. For 6-well plates, 1 ml of culture was inoculated in each well. After a 60 min adherence period at 30°C, the non-adherent inoculum was removed and 1 ml of fresh medium (RPMIMOPS) was applied to each well. Plates were incubated at 37°C for 48 h on an orbital shaker set at 50 rpm. Medium was removed and fresh medium was added midway through the incubation period.
A jugular vein rat central venous catheter infection model was used for in vivo biofilm studies [34]. Candida strains were grown to late logarithmic phase in YPD at 30°C with orbital shaking at 200 rpm. Following a 24 h conditioning period after catheter placement, infection was achieved by intraluminal instillation of 500 µl of C. albicans (106cells/ml). After an adherence period of 6 h, the catheter volume was withdrawn and the catheter was flushed with heparinized saline. For drug treatment experiments, fluconazole (250 µg/ml) was instilled in the catheter after 24 h of biofilm growth. After a 24 h drug treatment period, the post treatment viable burden of Candida biofilm on the catheter surface was measured by viable plate counts on Sabouraud's dextrose agar (SDA) following removal of the biofilm by sonication and vortexing.
A tetrazolium salt XTT [2,3-bis-(2-methoxy-4-nitro-5-sulfophenyl)-2H-tetrazolium-5-carboxanilide inner salt] reduction assay was used to measure in vitro biofilm drug susceptibility [56], [57]. Biofilms were formed in the wells of 96-well microtiter plates, as described above. After a 6 h biofilm formation period, the biofilms were washed with phosphate-buffered saline (PBS) twice to remove non-adherent cells. Fresh RPMI-MOPS and drug dilutions were added, followed by additional periods of incubation (48 h). The antifungals studied included fluconazole at 4 to 1,000 µg/ml. Drug treatments were reapplied after 24 h, and plates were incubated for an additional 24 h. Following treatment with 90 µl XTT (0.75 mg/ml) and 10 µl phenazine methosulfate (320 µg/ml) for 30 min, absorbance at 492 nm was measured using an automated plate reader. The percent reduction in biofilm growth was calculated using the reduction in absorbance compared to that of controls with no antifungal treatment. Assays were performed in triplicate, and significant differences were measured by analysis of variance (ANOVA) with pairwise comparisons using the Holm-Sidak method.
The CLSI M27 A3 broth microdilution susceptibility method was used to examine the activities of fluconazole against planktonic C. albicans [58]. Endpoints were assessed after 24 h by visible turbidity.
Agents with various mechanisms of action known to impact cell integrity were tested [37], [59]. A 96-well XTT assay, as described above, was used for measurement of the biofilm response to stress-inducing agents. The concentration required for a 50% reduction in XTT absorbance (50% effective concentration [EC50]) was recorded as the endpoint. Assays were performed in triplicate. The following concentration ranges were tested: calcofluor white, 0.2 to 200 µg/ml; β 1,3 Glucanase, 0.625 to 5 units/ml; H2O2, 25–200 µM; and sodium dodecyl sulfate (SDS), 0.001 to 2%.
In vitro biofilms were grown on sterile coverslips (Thermanox) in sterile 12 well plates and coated with 10 µl of human NaEDTA plasma each, which were dried at 30°C. 40 µl of yeast in RPMI, counted and diluted as in the biofilm models described above, was added to each coverslip for 60 min at 30°C. The initial inoculum was then removed and the plates incubated in 1 ml RPMI+MOPS+5% NaEDTA human plasma for 20 h at 37°C and 50 rpm on an orbital shaker. Media was replaced with 1 ml of fixative (4% formaldehyde, 1% glutaraldehyde in PBS) and coverslips were incubated at 4°C for 24 hours. The coverslips were then washed with PBS and treated with 1% osmium tetroxide for 30 min at ambient temperature. After a series of alcohol washes (30 to 100%), final desiccation was performed by critical-point drying. Coverslips were mounted, palladium – gold coated, and imaged in a scanning electron microscope (SEM LEO 1530) at 3 kV. The images were assembled using Adobe Photoshop 7.0.1.
C. albicans biofilms were grown on 6-well polystyrene plates for 48 h as described above. Cells were prepared for transmission electron microscopy (TEM) as previously described [30]. Following fixation in 4% formaldehyde and 2% glutaraldehyde, cells were postfixed with 1% osmium tetroxide and 1% potassium ferricyanide, stained with 1% uranyl acetate, dehydrated in a graded series of ethanol concentrations, and embedded in Spurr's resin. Sections (70 nm) were cut, placed on copper grids, poststained with 8% uranyl acetate in 50% methanol and Reynolds' lead citrate, and analyzed by TEM (Philips CM 120). The total cell and cell wall areas of 50 reference and mutant biofilm cells were measured using NIH Image J (http://rsbweb.nih.gov/ij/). The percentages of the cell wall area, defined as the cell wall area divided by the total cellular area, were calculated. Student's t test was used to determine statistical significance of differences between strains.
Biofilms growing in 6-well plates for 48 h were washed with PBS and collected for cell wall carbohydrate analysis as previously described [16], [60]. Briefly, cells (5 mg dry cell weight) were washed with PBS and broken apart with glass beads. Isolated cell walls were alkali extracted for 60 min with 500 µl of 0.7 M NaOH at 75°C three times. The combined alkali-soluble supernatants were neutralized with 250 µl glacial acetic acid. Following neutralization, the alkali-insoluble pellet was digested with 100 U Zymolyase 20T (MP Biomedicals) at 37°C for 16 h. One half of the Zymolyase-soluble fraction was dialyzed (Slide-A-Lyzer dialysis cassette, 7,000-molecular-weight-cutoff [MWCO]; Pierce) to yield a β-1,6-glucan fraction. The β-1,3-glucan fraction was calculated as the difference between the total Zymolyase-soluble glucan and β-1,6-glucan fractions. The carbohydrate contents of each fraction were measured as hexoses by the phenol-sulfuric acid method and normalized for dry cell wall weight. ANOVA with pairwise comparisons (Holm-Sidak method) was used to determine statistical significance.
The matrix β-1,3 glucan content was measured using a Limulus lysate based assay, as previously described [16]. Matrix was collected from C. albicans biofilms growing in the wells of 6-well polystyrene plates for 48 h. Biofilms were dislodged using a sterile spatula, sonicated for 10 min, and centrifuged 3 times at 4,500×g for 20 min to separate cells from soluble matrix material. Samples were stored at −20°C, and glucan concentrations were determined using the Glucatell (1,3)-β-D-glucan detection reagent kit (Associates of Cape Cod, MA) per the manufacturer's directions. Glucan concentrations were normalized for comparison across strains based upon viable biofilm burden using the XTT assay described above.
Matrix β-1,3 glucan was also measured using an ELISA assay. Biofilm was grown for 48 hours in 5×850 cm2 roller bottles (Corning, Thermo-Fisher) at 37°C. Biofilms were harvested into H2O using a sterile spatula then sonicated at 42 kHz for 20 min to dislodge the matrix. Next, biofilms were centrifuged 3×4,000 rpm for 20 min to separate the cells from the soluble matrix. The supernatant was lyophilized, dialyzed in a 3 kDa dialysis membrane (Spectra, Thermo-Fisher), and re-lyophilized to a powder. One mg of powdered matrix, dissolved in 1 ml of PBS was used as the sample in the ELISA assay and laminarin was used as a standard. A range of 1–1000 ng/ml of laminarin was used for the ELISA standard curve. 200 µl of 1 mg/ml matrix for each strain was assayed in triplicate. Plates were incubated overnight at 4°C, followed by blocking with 1% BSA for 45 min at ambient temperature. A 1∶2000 dilution of anti- β-1,3-glucan (BioSupplies Inc, Australia) was used as the primary antibody and a 1∶10,000 dilution of goat anti-mouse IgG-Biotin labeled [Sigma, Saint Louis] was used as a second antibody. Avidin-Peroxidase (Sigma) was used for detection.
A radiolabeled fluconazole accumulation protocol was adapted for biofilm use as previously described [51], [61]. Biofilms were grown in 6-well plates, as detailed above. The biofilms were washed with sterile water twice. For stock solution preparation, radioactive [H3] fluconazole (Moravek Biochemicals; 50 µM, 0.001 mCi/ml in ethanol) was diluted 100-fold in water. The stock solution was then diluted 6-fold in RPMI-MOPS, and each biofilm well received a total of 600 µl of this medium to yield a total of 8.48×105 cpm of [H3] fluconazole. After incubation for 30 min at 37°C and orbital shaking at 50 rpm, unlabeled (cold) 20 µM fluconazole in RPMI-MOPS was added and biofilms were incubated for an additional 15 min. Biofilms were then washed twice with sterile water, dislodged with a spatula, and collected as intact biofilms for scintillation counting. The biofilms were then disrupted by vortexing and sonication to separate cells and matrix. Following centrifugation, cells were separated from the soluble matrix material. Cells were subsequently disrupted by bead beating, and the intracellular and cell wall portions were collected by centrifugation. The fractions were then suspended in ScintiSafe 30% LSC cocktail (Fisher Scientific) and counted in a Tri-Carb 2100TR liquid scintillation analyzer (Packard). ANOVA was used to determine statistical significance of differences among strains.
Biofilms were grown using the 96 well microtiter model described above for 24 hours. Then, 90 µl of fresh media and 90 µl of serial 2 fold dilutions of the β-1,3 glucanase (Zymolyase - 20T, MP Biomedicals) in 0.9% NaCl was added to each well, with concentrations ranging from 5 U/ml to 0.625 U/ml. Plates were incubated at 37°C for 24 hours, at which point the media was removed and the biofilms were washed gently in 100 µl of PBS to remove any non-adherent cells. The plates were read using the XTT assay as described above. For comparison, a duplicate set of plates was spun at 3,000 RPM for 5 minutes before the media was removed on the final day. These biofilms were read via the XTT assay immediately, without washing, thus quantifying all living cells in each well to show whether β-1,3 glucanase at the concentrations used causes cell disaggregation or lysis.
|
10.1371/journal.pntd.0003791 | Bacterial Factors Associated with Lethal Outcome of Enteropathogenic Escherichia coli Infection: Genomic Case-Control Studies | Typical enteropathogenic Escherichia coli (tEPEC) strains were associated with mortality in the Global Enteric Multicenter Study (GEMS). Genetic differences in tEPEC strains could underlie some of the variability in clinical outcome.
We produced draft genome sequences of all available tEPEC strains from GEMS lethal infections (LIs) and of closely matched EPEC strains from GEMS subjects with non-lethal symptomatic infections (NSIs) and asymptomatic infections (AIs) to identify gene clusters (potential protein encoding sequences sharing ≥90% nucleotide sequence identity) associated with lethality.
Among 14,412 gene clusters identified, the presence or absence of 392 was associated with clinical outcome. As expected, more gene clusters were associated with LI versus AI than LI versus NSI. The gene clusters more prevalent in strains from LI than those from NSI and AI included those encoding proteins involved in O-antigen biogenesis, while clusters encoding type 3 secretion effectors EspJ and OspB were among those more prevalent in strains from non-lethal infections. One gene cluster encoding a variant of an NleG ubiquitin ligase was associated with LI versus AI, while two other nleG clusters had the opposite association. Similar associations were found for two nleG gene clusters in an additional, larger sample of NSI and AI GEMS strains.
Particular genes are associated with lethal tEPEC infections. Further study of these factors holds potential to unravel the mechanisms underlying severe disease and to prevent adverse outcomes.
| Typical enteropathogenic E. coli (tEPEC) strains are associated with high mortality among infants with moderate-to-severe diarrhea, but most infants infected with tEPEC strains survive, and some have no symptoms. To investigate the bacterial factors associated with severe outcome, we determined the genomic sequences of 70 EPEC strains. Twenty four tEPEC strains came from children with lethal infections. The prevalence of each gene was compared to that in strains from 23 matched infants who had non-lethal symptomatic infection and to that in 23 matched infants who had asymptomatic infection. We identified 392 genes associated with outcome, some of which were more prevalent in strains from lethal infections, while others were less prevalent. The genes included several encoding potential virulence factors such as type 3 secreted effectors and enzymes involved in O-antigen synthesis. A PCR assay validated the association of groups of alleles encoding variants of the NleG ubiquitin ligase with clinical outcome. Further study of the factors associated with severe outcome could lead to novel diagnostic, therapeutic and prevention strategies.
| Nearly 70 years ago, bacterial strains now known as enteropathogenic Escherichia coli (EPEC) were first reported to cause neonatal and infant diarrhea with high case-fatality rates [1,2]. During the ensuing years, investigators made great strides in elucidating the molecular mechanisms and cell biology of EPEC infection. EPEC strains use a Type 3 Secretion (T3S) system to inject into host cells a variety of proteins that disrupt the actin cytoskeleton, block innate immune responses, and modulate apoptosis [3,4]. The ability to deliver these effector proteins is essential for disease in humans [5]. Infection of enterocytes leads to loss of microvilli and the formation of cup-like pedestals to which the bacteria intimately adhere, a process known as attaching and effacing [6,7]. Typical EPEC (tEPEC) strains also produce a type IV bundle-forming pilus (BFP) [8], the individual fibers of which intertwine and cause the bacteria to form aggregates that initially attach to cells. Genomic characterization of multiple EPEC strains reveals a complex evolutionary history, evidence of extensive horizontal transfer of critical virulence determinants, and thousands of genes that are not universal among EPEC [9].
Since initial reports, the global epidemiology of EPEC infections has evolved. Whereas in the past, tEPEC strains were reported to be a leading bacterial cause of neonatal diarrhea in numerous developing countries [10–13], their prevalence appears to have declined coincident with socioeconomic gains [14]. However, the recently published results of the Global Enteric Multicenter Study (GEMS), the most comprehensive investigation of the etiology of moderate-to-severe diarrhea (MSD) in childhood yet performed, shed new light on the virulence of tEPEC strains [15]. While tEPEC strains were not responsible for a large burden of disease in the four African and three Asian sites surveyed, they were associated with a significantly elevated risk of mortality in young infants with MSD.
To gain insight into bacterial factors that play a role in severe clinical outcomes of tEPEC infection, we conducted a unique, matched case-control genomic investigation of tEPEC strains from children enrolled in GEMS who had lethal infections (LIs), compared with children who had non-lethal symptomatic infections (NSIs) and asymptomatic infections (AIs).
Strains were cultured from the feces of infants or children who had acute MSD or who were recruited as healthy controls; tEPEC strains had been identified as described [16]. Individual colonies were shipped from countries of origin to Baltimore, Maryland, USA and verified by PCR, using primers listed in S1 Table as previously described [9] for bfpA and escV to identify tEPEC, and stxA1 and stxA2 to identify and exclude Shiga-toxin producing E. coli. PCR using the same conditions and primers also listed in S1 Table was used to detect nleG clusters in a larger collection of 63 NSI and 56 AI unmatched tEPEC and atypical EPEC (aEPEC, bfpA-/escV+/stx-) GEMS isolates.
To maximize the probability that observed differences between groups would be due to bacterial genetic factors, each available tEPEC LI strain was matched to EPEC strains from NSI and AI subjects on site and sex, and then proximity-matched using the Mahalanabis method [17] on a range of other variables found to be associated with LI status. Importantly, this matching is completely distinct from the original GEMS study because in the current study all subjects were culture-positive for EPEC. Details are provided in Supporting Information. The strength of these associations was weighed in the Mahalanabis procedure to derive a propensity score. Matches were sought to minimize the propensity score and identify the best possible match. When no match with a tEPEC strain was available, an aEPEC strain was selected by the same criteria.
Detailed genomic sequencing and assembly methods are available in Supporting Information. The average genome coverage was approximately 226-fold.
Genes were predicted in each of the 70 genomes and related genes were then grouped with uclust [18] into gene clusters based on the degree of similarity, using an nucleotide identity threshold of 90%. Following the clustering, a file was generated that contained a consensus sequence for each cluster. The consensus sequences were then translated and compared to each genome using TBLASTN [19] as described above. The maximum TBLASTN bit score value obtained for each cluster was used as the denominator to generate a ratio for the cluster compared to each genome. The level of similarity of protein-encoding genes was compared across all 70 genomes in this study using a large-scale BLAST score ratio (LS-BSR) analysis as previously described [20]. The presence and absence of each potential gene cluster in the 70 genomes was ascertained using established thresholds for the BSR analysis [20].
The Stata 12 statistical software package (Statacorp LP, College Station, TX) was used to perform the preparatory logistic regression for the Mahalanobis proximity matching, and the Stata module mahapick was used to perform the matching and calculate propensity scores. The significance of associations between particular gene clusters and clinical outcome in matched strain pairs was analyzed using McNemar’s exact test [21], which compares the number of discordant pairs in which the variable is present in the case and absent in the control to the number in which the variable is absent in the case and present in the control. First, all strain pairs were considered and subsequently, to reduce the risk of type I error, subgroups of only well-matched strain pairs and only tEPEC strain pairs were analyzed. Pearson’s chi-square test was used to test the significance of differences in prevalence of particular gene clusters between unmatched NSI and AI strains.
The research reported in this study was exempt from institutional review.
Thirty-three infants and children enrolled in GEMS who had had acute MSD and tEPEC identified in the stool did not survive 60 days [15]. tEPEC strains were available and verified from twenty four (73%) of these infants with LI (Fig 1). For each of these LI cases, the most closely matched case with NSI and control with AI, based on propensity score, who were infected with tEPEC (or aEPEC if no tEPEC strain was available), came from the same site and had the same gender, were selected (see Supporting Information for factors affecting propensity score, S2 Table for matching details). A single NSI strain was the closest match for two LI strains and a single AI strain was the closest match for two other LI strains (S2 Table). Thus, 70 strains in total were selected for genomic analysis (Fig 1). NSI patients and LI patients were similar with regard to relevant clinical variables (Table 1). As might be expected, AI controls were not as closely matched (Table 1), particularly for height-for-age Z-score. The only significant difference between the groups was discordance between LI-AI pairs regarding whether or not the bacteria were tEPEC (P = 0.031, McNemar’s exact test), due to the unavailability of tEPEC AI strains to serve as controls for some LI strains. From the distribution of propensity scores, 19 LI-NSI strain pairs and 16 LI-AI strain pairs with scores ≤ 3 were considered well-matched (see Supporting Information).
We used a large-scale BLAST score ratio (LS-BSR) analysis to examine the gene content of the 70 EPEC isolates. In the LS-BSR analysis, predicted genes from each genome are grouped together into clusters when they have ≥ 90% nucleotide identity (Sahl et al., 2013). Conversely, alleles of the same gene are scored as different gene clusters when they have <90% nucleotide identity. For the 70 genomes analyzed in this study, there were 14,412 gene clusters identified. Of this total, 1,316 gene clusters were present in all 70 genomes analyzed based on a conservative BSR value of ≥ 0.8 (S3 Table), representing the conserved EPEC core genome.
Interrogation of the LS-BSR patterns and comparative analysis indicated that no two strains exhibited the same pattern and thus the isolates in this collection were not duplicates and did not result from expansion of single clones. No gene clusters were found exclusively in strains from individuals with any of the three clinical outcomes examined. However, we identified 9 clusters that were significantly more prevalent in LI strains than in matched NSI strains, but not matched AI strains and 286 clusters that were significantly more prevalent in LI strains than in matched AI strains, but not matched NSI strains (Table 2). In addition, 7 clusters were more prevalent in comparisons of LI strains with both NSI and AI strains. These 302 clusters may be considered genes putatively associated with LI. We also identified 16 clusters significantly more prevalent in NSI strains, but not AI strains, than in matched LI strains and 62 clusters significantly more prevalent in AI strains, but not NSI strains, than matched LI strains, as well as 13 more prevalent in both NSI and AI strains than LI strains (Table 2). Thus, 91 gene clusters have putative associations with nonlethal infection. The greater number of clusters identified in comparisons between LI and AI strains than between LI and NSI strains indicates, as expected, that EPEC strains from patients who did not survive differ more from those of control subjects with no symptoms than they do from strains of patients with nonlethal MSD.
The lists of all gene clusters significantly associated with clinical outcome in comparisons of LI with NSI, LI with AI, and LI with both NSI and AI are shown in S4–S6 Tables. These tables also display the results of pre-specified subgroup analyses comparing only tEPEC strains, only strains from well-matched pairs, and only tEPEC from well-matched pairs, which are summarized in Table 2. For example, when only tEPEC strains are considered, 14 of the 20 gene clusters that were either more prevalent in LI strains than both NSI and AI strains or vice versa remained significant for both comparisons. Additional gene clusters remained significant in comparisons between tEPEC strains from LI and NSI or LI and AI, but not both. Subgroup analyses of strains only from well-matched pairs with propensity scores < 3 reduced the number of significant clusters to five. Similarly, 74 clusters were more prevalent in LI than AI and 18 more prevalent in AI than LI when only well-matched strain pairs were considered, 71 clusters were more prevalent in LI than AI and 26 more prevalent in AI than LI when only tEPEC strains were compared, and 15 were more prevalent in LI than AI and 5 more prevalent in AI than LI when only well-matched pairs with tEPEC were considered. For LI comparisons with NSI, these numbers were 7 and 1, 5 and 6, and 4 and 3, respectively (Table 2). Some clusters were significant only in subgroup analyses (S4–S6 Tables).
The largest categories of loci associated with outcome identified encode bacteriophage-related and hypothetical proteins. Several genes that have plausible links to virulence were also overrepresented according to clinical outcome. A gene cluster encoding closely related alleles of rfbB was significantly associated with LI, both in comparisons to NSI and LI strains. RfbB, a dTDP-glucose 4,6-dehydratase, is part of the O-specific LPS synthesis pathway (www.ecocyc.org). A gene cluster encoding the Wza polysaccharide capsule transporter protein (www.ecocyc.org) was significantly more prevalent when all LI and matched AI strains were considered, but not when only tEPEC or well-matched strains were compared.
Interestingly, some gene clusters encoding putative virulence genes were significantly associated with non-lethal infection. Clusters encoding the T3S effectors EspJ and OspB were significantly more common in AI than matched LI strains, although significance was not maintained in the pre-specified subgroup analyses limited to well-matched or tEPEC-only strains (S5 Table).
The T3S effector protein NleG, an E3 ubiquitin ligase (Wu et al., 2010), shows extensive sequence variation, with 11 nleG gene clusters identified in the current collection and many strains having multiple alleles. Against this backdrop emerged a remarkable degree of bias with regard to the presence of certain nleG gene clusters among matched isolates from LI and AI subjects (S5 Table). Whereas, in 12 such pairs, cluster 6826 was present in the LI strain, but absent in the AI strain, in only one such pair was this cluster present in the AI strain and absent in the LI strain. Conversely, gene cluster 4759 and gene cluster 6719 were present in the AI strain and absent in the LI strain in six and seven pairs, respectively, whereas in no pairs was either cluster present in the LI strain and absent in the AI strain. Thus, it appears that different nleG alleles are strongly associated with the extremes of virulence observed in this study.
Because of the extremely large number of clusters identified compared to the relatively small number of matched subjects available, correction for multiple comparisons virtually eliminates the potential for statistically significant associations. Therefore, we sought to determine whether an association we identified could be replicated in a larger data set. Ideally, additional matched LI-NSI and LI-AI pairs would be examined; however, we know of no source of additional LI strains. However, 63 additional NSI strains and 56 additional AI strains from GEMS were available for testing. Clusters that were significantly associated, either positively or negatively, with lethal outcome in the LI versus AI analysis (S5 Table), but not the LI versus NSI analysis (S4 Table) or vice versa, might reasonably be expected to differ in prevalence between NSI and AI strains. The various nleG clusters mentioned above met these criteria as they showed strong discordance in comparisons of matched LI and AI pairs, but no significant association with outcome in matched LI and NSI pairs. Additionally, as the T3S system is linked to virulence, and nleG was the only T3S effector both positively and negatively associated with outcome, we decided to focus on this locus. Indeed, nleG cluster 6826 was identified by PCR in 43 of 63 (68%) of NSI strains and 28 of 56 (50%) of AI strains (P = 0.043). Conversely, nleG cluster 6719 was present in 10 of 63 (16%) NSI strains and 18 of 56 (32%) AI strains (P = 0.037). However, no significant association with outcome was found for nleG cluster 4759.
Since the initial identification of tEPEC as a cause of infant diarrhea, it has been appreciated that this infection can be deadly [1,22]. This early observation has been reaffirmed in the modern era [23], most recently by the GEMS, in which isolation of tEPEC among infants 1–11 months of age who have MSD more than doubled the risk of death during the follow-up period [15]. Indeed, 6.4% of such infants did not survive for 60 days. While the cause of death in these infants is uncertain, it is likely that host susceptibility influenced by factors such as passively transferred maternal immunoglobulins, genetic predisposition, degree of malnutrition and concurrent infection conspired with specific bacterial virulence factors to produce a deadly outcome. The relative contribution of genetic variation in bacterial virulence to clinical outcome is unknown.
We used a novel genomic epidemiological approach to test the hypothesis that variation in the presence and absence of genes is associated with lethal outcome in EPEC infection. The average E. coli genome has approximately 5000 genes, only about 2000 of which are shared among all members of the species [9,24,25]. The balance of genes not only determines the pathovar of the isolate, but dictates a remarkable degree of variation among strains within each pathovar. We reasoned that some of the variability in tEPEC gene content could exert a strong influence on clinical outcome.
To test our hypothesis, we generated draft genome sequences of 70 EPEC strains including 24 tEPEC strains associated with LI, 23 matched strains from children with NSI, and 23 matched strains from children with AI. To minimize the contributions of host factors, we used a propensity-matching method to select the closest available matches for each LI isolate.
Among the 70 genomes analyzed, we identified 14,412 distinct gene clusters, defined as potential protein-coding DNA sequences differing by 10% or more from all other sequences. As expected, many of these clusters were unevenly distributed among LI, NSI and AI strains. Among the 392 gene clusters that showed statistically significant associations with lethality, the majority encodes hypothetical or bacteriophage proteins, the potential contribution to virulence of which is difficult to assess. However, the contribution of several identified bacterial factors to severe clinical outcomes is entirely feasible.
Since the discovery of EPEC, a limited number of O-antigen serogroups has been associated with disease (Kauffmann and Dupont, 1950;Taylor and Charter, 1952). A focus on the O-antigen as a virulence determinant by early investigators was superseded by the identification of specific virulence genes such as those that encode the T3S system and the BFP that now define EPEC [26]. However, the contribution of particular O-antigen types to virulence remains undefined. An rfbB gene cluster encoding dTDP-glucose 4,6-dehydratase, an enzyme required for synthesis of O-specific lipopolysaccharide, was significantly associated with LI in comparisons to both NSI and AI strains (S6 Table). Gene clusters most closely related to wblO and wblQ potentially encoding glucose-1-phosphate thymidylyltransferase and UDP-4-amino-4-deoxy-L-arabinose-oxoglutarate aminotransferase, respectively, had similar associations and may also be involved in O-antigen synthesis. Thus, the findings in this study resurrect the possibility that particular O-antigens or other surface carbohydrates may modulate tEPEC virulence.
The associations of particular nleG gene clusters with clinical outcome were particularly noteworthy. We not only identified different nleG clusters associated with lethal versus asymptomatic infections and vice versa, but were also able to confirm the expected difference in prevalence between additional strains from nonlethal symptomatic versus asymptomatic children for two of these clusters. NleG proteins are E3 ubiquitin ligases, and the remarkable variety of nleG genes, including 14 alleles in the Sakai strain of enterohemorrhagic E. coli, has been previously noted [27]. The cellular targets of NleG proteins have not yet been identified, but it is tempting to speculate that the diversity of NleG proteins is related to substrate specificity, which in turn influences clinical outcome. Further studies to test this hypothesis are indicated.
To our knowledge, this is the first study to use genome sequencing to identify associations between E. coli variability and clinical outcome, although less comprehensive genetic analyses have been performed in the past. In a study limited to aEPEC strains from children with diarrhea and controls in Norway, the prevalence of 182 virulence genes of various E. coli pathovars was compared using an oligonucleotide array [28]. Genes from a pathogenicity island including those encoding T3S system effectors NleB, NleE, lymphostatin and EspL were significantly more common in strains from children with diarrhea. Interestingly, those same genes were identified in a completely different context [29]. Seventy-two Shiga toxin producing E. coli strains were specifically tested for these four genes by PCR. Their prevalence was greater in strains associated with two markers of increased severity: those from outbreaks compared to sporadic cases and those associated with hemolytic-uremic syndrome. Given the large differences in patient populations, strains and methods between these studies and the current study, it should perhaps not be surprising that we did not find associations of these genes with lethal infection in the current study.
Our study has a number of weaknesses, although we know of none that could have been avoided. Although we took great care to optimize matching, differences between LI cases and NSI cases and AI controls persist, particularly with regard to those pairs for which no tEPEC strain was available. The greater heterogeneity of aEPEC in comparison to tEPEC strains [9] adds further genomic complexity, potentially obscuring results. Our pre-specified analysis limited to tEPEC strain pairs is useful in this regard, but at the cost of reducing power to detect differences. Although our collection of 24 tEPEC strains associated with LI is unprecedented in its size, the number of strain pairs studied is relatively small for an epidemiological study, a major factor limiting our ability to detect differences in the prevalence of genes according to clinical outcome. This issue is especially problematic given the large number of statistical comparisons (14,412) made. Accordingly, traditional approaches such as Bonferroni corrections and false discovery rate calculations to limit the possibility of type I statistical errors were impractical, as these methods would have yielded no residual significant associations at a considerable cost of type II errors. These considerations render broad-scale investigations of finer detail differences in our data, such as single-nucleotide polymorphisms or particular alleles of known virulence factors, entirely unrealistic in the absence of specific a priori hypotheses. Thus, the associations we identified must be viewed cautiously as requiring further validation.
Our study also has notable strengths, including a highly relevant and objective clinical outcome (mortality), a rigorous matching procedure, the comprehensive detail afforded by highly redundant draft genome sequencing, and the pre-specified subgroup analyses designed to mitigate the effects of imperfect strain matching. Our results are supported by the biological plausibility of some of the loci we identified. Our results are further supported by the expected observation of larger differences in comparisons between LI and AI strains than between LI and NSI strains. Furthermore, differences in the prevalence of two versions of nleG in a larger set of strains validated the association found in the initial analysis, adding additional confidence to our conclusions.
This unique analysis of genetic variation among tEPEC strains according to clinical outcome provides valuable information linking specific genes to risk of and protection from lethality. The list of such genes provides fertile grounds for further investigations with the ultimate goal of preventing severe EPEC disease.
|
10.1371/journal.pcbi.1005657 | β-adrenergic signaling broadly contributes to LTP induction | Long-lasting forms of long-term potentiation (LTP) represent one of the major cellular mechanisms underlying learning and memory. One of the fundamental questions in the field of LTP is why different molecules are critical for long-lasting forms of LTP induced by diverse experimental protocols. Further complexity stems from spatial aspects of signaling networks, such that some molecules function in the dendrite and some are critical in the spine. We investigated whether the diverse experimental evidence can be unified by creating a spatial, mechanistic model of multiple signaling pathways in hippocampal CA1 neurons. Our results show that the combination of activity of several key kinases can predict the occurrence of long-lasting forms of LTP for multiple experimental protocols. Specifically Ca2+/calmodulin activated kinase II, protein kinase A and exchange protein activated by cAMP (Epac) together predict the occurrence of LTP in response to strong stimulation (multiple trains of 100 Hz) or weak stimulation augmented by isoproterenol. Furthermore, our analysis suggests that activation of the β-adrenergic receptor either via canonical (Gs-coupled) or non-canonical (Gi-coupled) pathways underpins most forms of long-lasting LTP. Simulations make the experimentally testable prediction that a complete antagonist of the β-adrenergic receptor will likely block long-lasting LTP in response to strong stimulation. Collectively these results suggest that converging molecular mechanisms allow CA1 neurons to flexibly utilize signaling mechanisms best tuned to temporal pattern of synaptic input to achieve long-lasting LTP and memory storage.
| Long-term potentiation of the strength of synaptic connections is a mechanism of learning and memory storage. One of the most confusing aspects of hippocampal synaptic potentiation is that numerous experiments have revealed the requirement for a plethora of signaling molecules. Furthermore the degree to which molecules activated by the stress response modify hippocampal synaptic potentiation and memory is still unclear. We used a computational model to demonstrate that this molecular diversity can be explained by considering a combination of several key molecules. We also show that activation of β-adrenergic receptors by the stress response appears to be involved in most forms of synaptic potentiation, though in some cases unconventional mechanisms are utilized. This suggests that novel treatments for stress-related disorders may have more success if they target unconventional mechanisms activated by β-adrenergic receptors.
| Synaptic plasticity is one of the cellular mechanisms underlying learning and memory. In the hippocampus, long-term potentiation (LTP) has been implicated not only in acquisition, consolidation and retrieval of spatial memories, but also contextual fear extinction [1–4]. Several neuromodulatory systems contribute to both synaptic plasticity and fear memory [5], including pathological memory retention such as post-traumatic stress disorder (PTSD). One of the most potent regulatory systems is the noradrenergic system, which is activated by arousal, emotion and stress. Experimental evidence shows that norepinephrine is elevated in the hippocampus in mouse models of PTSD [6, 7]; however, its contribution to long term plasticity is unclear and this lack of knowledge hinders the development of treatments for fear memory disorders.
Numerous experiments investigating long-lasting LTP have revealed the requirement for a plethora of signaling molecules (reviewed in [5, 8]). Experimental protocols that induce long-lasting LTP activate diverse signaling pathways, which may interact competitively or cooperatively. For example, long-lasting LTP evoked by multiple trains of high-frequency electric stimulation requires protein kinase A (PKA) only if the inter-train interval is greater than 60 sec [9, 10]. These networks of signaling pathways may converge on common targets, such as extra cellular regulated kinase (ERK), which is required for most forms of long-lasting LTP [11–14]. Alternatively, some components of those signaling pathways are location specific and function in restricted spatial compartments such as spines or dendritic submembrane. Those observations pose the key question of whether this diversity of mechanisms can be explained by collectively considering the combined molecular network.
Another type of unexplained diversity of mechanisms underlying induction of long-lasting LTP is introduced by neuromodulation. To date, β-adrenergic receptor (βAR) activation has been considered essential for only a subset of experimental protocols, usually for weak electric stimulation. Conversely, commonly used βAR antagonists, such as propranolol, do not affect long-lasting LTP elicited by strong electric stimulation.
The idea that βAR activation is not essential for long-lasting forms of LTP was undermined by recent experiments suggesting that conventional βAR antagonists do not block all downstream signaling pathways. Though βARs typically are coupled with stimulatory G protein (Gs), phosphorylated βARs decouple from Gs and couple with inhibitory G protein (Gi). Both Gs-activated and βAR coupled to Gi-activated signaling pathways converge on a common target, ERK [15–17], which is required for long-lasting LTP. The ability of propranolol to recruit ERK [18], suggests that long-lasting LTP evoked by strong stimulation with or without propranolol might require βAR signaling to ERK. This hypothesis is supported by recent experiments showing that a complete βAR antagonist blocks long-lasting LTP induced by strong electric stimulation [19]. Therefore, βAR activation might play a pivotal role for many forms of long-lasting LTP.
To investigate whether the diverse experimental evidence can be unified by considering activation of multiple signaling cascades and address the role of βAR activation in occurrence of long-lasting LTP, we develop a spatial, mechanistic model of signaling pathways underlying induction of long-lasting forms of LTP. We evaluate spatio-temporal dynamics of key kinases that activate molecular pathways reported to play an essential role in long-lasting forms of LTP. We show that the combined elevation of several molecules in the spine and in the dendrite can predict the induction of long-lasting LTP, and our results suggest that activation of the βARs may be essential for all forms of LTP. These findings may help unravel the contribution of the noradrenergic system to learning and memory and help with the development of treatments for fear and anxiety disorders.
To investigate how temporal pattern of synaptic activation determines which signaling pathways are activated, we employed a multi-compartmental, stochastic reaction-diffusion model of calcium and cAMP activated signaling pathways (Fig 1). The model was adapted from an existing model of a dendrite plus spine of a CA1 hippocampal pyramidal neuron [20]. The signaling pathways included calcium-calmodulin activated molecules, such as calcineurin (PP2B); and phosphodiesterase 1B (PDE1B), cAMP activated molecules: Epac and PKA, and interactions between calcium and cAMP pathways via Inhibitor1. The previously published model [20] was modified by adding neurogranin (Ng) [21], a calmodulin buffer, implicated in LTP and learning [22, 23]. Most importantly we added several pathways downstream of β2AR [24] to the model.
βARs in CA1 pyramidal neurons are activated by norepinephrine and mainly couple to stimulatory G proteins [25, 26]. The activated α subunit of Gs (Gαs GTP) synergistically enhances cAMP production by calcium-calmodulin bound adenylyl cyclase 1 (AC1). Elevations in cAMP, produced by either prolonged stimulation of β2AR or increases in intracellular calcium, activate PKA, which can phosphorylate β2AR. There are four sites of heterologous phosphorylation [27] on the β2AR [28], whose phosphorylation leads to alternative G protein coupling. In the model, a single phosphorylation event decouples the β2AR from Gs, and the fully phosphorylated β2AR then binds inhibitory G protein (Gi). The β2ARs are phosphorylated in a cooperative and distributive manner [29], which yields an ultrasensitive response [30–32]. Note that both Gi [15–17] and β-arrestin [33] have been implicated in ERK recruitment to the pβ2AR; thus, in our model the Gi binding to pβ2AR could alternatively represent β arrestin binding. Kinetic constants of the model are presented in S1 Table.
The morphology of the model comprised one spine attached to a 2 μM dendrite or 8 spines attached to a 20 μM long dendrite with 0.6 μM diameter (Fig 2). In all cases, the dendrite and spines were subdivided into voxels to accurately simulate spatial aspects of signaling molecules. Molecules diffused between spine and dendrite with a coupling coefficient proportional to the surface area of the spine neck. The layer of voxels immediately adjacent to the membrane was considered the submembrane domain. AC (type 1 and 8), PKA holoenzyme, G proteins and the β2ARs were localized and anchored both in this submembrane domain and the spine head. The diffusible molecules included cAMP, ATP, calcium, all forms of calmodulin (CaM), CaMKII, β2AR agonists and antagonists, Inhibitor-1 and Epac. Their diffusion constants are listed in S2 Table. Initial conditions were either taken from the prior model [20], experimental publications (e.g. quantity of neurogranin [21]), or adjusted to reproduce experimentally measured concentrations of dependent molecules, e.g. the balance of AC and PDE was adjusted to produce a 30 nM basal cAMP concentration [34, 35].
Different forms of LTP are evoked by different stimulation patterns [36]; thus, we performed simulations using seven, well characterized, stimulation protocols (Table 1). Four of them experimentally elicit L-LTP, one results in an early form of LTP (E-LTP) and the remaining two stimulation protocols do not produce LTP, though one (LFS) elicits brief depression. Electric stimulation of Schaeffer collaterals results in activation of post-synaptic NMDA receptors and action potentials, thus each stimulation pulse was simulated in the model as calcium injection both into the spine to represent NMDA receptors, and into the dendrite to represent activation of voltage dependent calcium channels. Electric stimulation in the hippocampus is accompanied by norepinephrine release [37], which was modeled as ligand influx. Bath applied isoproterenol (ISO) was simulated by injecting sufficient ISO to produce a 1 μM concentration. We started stimulation after 300 sec of simulation to ensure the model had reached equilibrium. Steady state was confirmed by running simulations for 900 sec in the absence of stimulation and visually assessing that activity of each molecular specie was stationary.
Several data sources were used to adjust calcium pulse amplitudes for all stimulation protocols. To stimulate calcium influx during 100 Hz trains of electric stimulation (HFS), we used release probabilities from [38] which provides changes in the amplitudes of calcium pulses in the spine during high frequency trains. We assumed that amplitudes of consecutive calcium pulses in the dendrites are uniform, because they result from full height action potentials. To calculate absolute amplitudes of calcium pulses, we constrained calcium concentration in the spine and in the dendrite to match experimental data [39]: 10 μM in the spine and 2 μM in the dendrite. This pattern of calcium pulses was used in all stimulation protocols using trains of HFS: 1 train of 100 Hz (HFS), four trains of 100 Hz given 3 sec apart (4xHFS-3s), four trains of 100 Hz given 80 sec apart (4xHFS-80s) and bath applied ISO followed by 1 train of 100 Hz (ISO+HFS). For the 5 Hz (LFS) stimulation protocol, spine calcium pulses were of the same amplitude, and equal to the amplitude of the first pulse of the HFS train [39].
In order to estimate the temporal pattern and amplitude of neuromodulation elicited by electric stimulation, we used a model (Eq 1) describing vesicle release [40]. This model assumes that synaptic resources can be found in three states: inactive (I), recovered (R) and effective (E; released). u represents release probability, which decays with a time constant τf and increases with each action potential (AP) by a fraction of USE. After the arrival of the AP a fraction of recovered resources (uR) becomes effective (E) i.e. gets released. Effective resources, E, become inactive with a time constant τi. Inactive resources, I, recover with a time constant τr. The Dirac delta function is denoted as δ(t − tAP) and has value 1 at t = tAP and 0 otherwise. ASE is the absolute synaptic efficacy (response amplitude produced by complete release of all the neurotransmitter). We tuned the vesicle release model on experimental data to voltammetric measurements of norepinephrine release in the rat Ventral Bed Nucleus Stria Terminalis following electric stimulation of noradrenergic projection pathways [41] (S6 Table). Using this model we estimated norepinephrine release for stimulation patterns in Table 1. The spatial distribution of norepinephrine during and following release was in agreement with a spatial gradient of neuromodulators [42], namely a high concentration in the spine release site (1 μM) and lower at the dendrite.
PKA phosphorylates NMDA receptors, which increases the amplitude of calcium influx through these receptors [43, 44]. This enhancement of NMDA mediated calcium influx has been observed with bath application of ISO. Thus, for the case of ISO+LFS, calcium influx was increased by 50% [45].
We modeled propranolol (1 μM; [46]) ICI-118,551 (100 nM; [19]) and carvedilol (10 μM [47]) by allowing it to bind the β2AR [48] (S1 Table), and then both propranolol- and carvedilol-bound β2AR were able to bind with Gi and form a target representing ERK activation. Binding affinity was constrained so that carvedilol produces one third the Gi bound β2AR compared to that of isoproterenol as has been measured experimentally [18].
We used a stochastic simulation technique, as many molecular populations are small. In such case activations fluctuate greatly about the mean within such small compartments [49, 50]. Similarly, diffusion of second messenger molecules out of the spines and along the thin dendrites is subject to random variation. The model was implemented using an efficient mesoscopic stochastic reaction-diffusion simulator NeuroRD [51], version 2.1.10, because the large numbers of molecules in the morphology described (Fig 2) made tracking individual molecules in microscopic stochastic simulators computationally expensive. This simulator uses reflective boundary conditions (molecules attempting to diffuse out of the morphology were reflected back into the morphology). Model simulations used a time step of 2.9 μs. A single simulation of 900 sec (of the dendrite with 1 spine) takes 4.5 days on a Intel Xeon CPU E5-2620 2.00GHz processor.
Based on results from our prior studies, simulations were repeated four or eight times using a different random seed. Eight simulations were used for stimulation protocols whose signature exhibited a large standard deviation relative to the mean. To determine whether the combination of stimulation and βAR ligand would induce L-LTP or not, we analyzed the duration of combined molecular activations (signatures) in the spine and in the dendrite above their respective thresholds. The statistical analysis used SAS (version 9.4, SAS Institute, NC). Student’s T test (SAS procedure TTEST) was applied to each condition to evaluate whether the duration above threshold was significantly greater than the duration threshold of 10 sec. For the multi-spine simulations, we used the SAS procedure GLM to perform a two-way analysis of variance using condition (adjacent or separate) and stimulation (spine was stimulated or not) as factors. All model simulation files are available from modelDB (https://senselab.med.yale.edu/ModelDB/showModel.cshtml?model=190304).
We validated the model by comparing activity of AMPA receptor (AMPAR) phosphorylation and PKA-mediated Gs-Gi switching with independent, published experimental results. To validate the PKA-mediated Gs-Gi switching we simulated bath application of 1 μM of isoproterenol in a model over-expressing Epac (8 time the amount used in other simulations). The model’s Epac activity was compared with the response of genetically encoded Epac-sh150 (monitoring cAMP activity in hippocampal CA1 neurons) to 1 uM ISO [52]. Fig 3 shows the model’s Epac activity and fluorescence traces of distal dendrites in response to bath application of 1 μM of isoproterenol (ISO) and confirms that the model accurately captures the decay in cAMP activity while isoproterenol is still present due to phosphorylation of βARs (and phosphodiesterases). For comparison we chose fluorescence traces of small, tertiary dendrites, which had similar diameter to the dendrite diameter used in the model.
Next we validated the model of AMPAR phosphorylation by comparing phosphorylation AMPAR at Serine 845 and Serine 831 to experimentally measured values. In the model bath application of 1 μM of ISO yields 200% increase in phosphorylation of Serine 845 and no discernible phosphorylation of Serine 831, which is in agreement with values reported in hippocampal CA1 neurons after bath applying 1 μM of ISO [53, 54]. These comparisons confirm the parameters describing inactivation mechanisms (both Gs-Gi switching and PDE4 phosphorylation) of cAMP and PKA activity for AMPAR.
Our goal was to explain the diverse literature on molecular dependence of long-lasting forms of LTP induction. We evaluated whether the spatio-temporal dynamics of molecular signaling pathways can explain and predict which stimulation patterns produce long-lasting LTP. We constructed a model of signaling pathways (Fig 1) that regulate long-lasting forms of LTP in hippocampal CA1 pyramidal neurons in NeuroRD [51] using the morphology of a dendrite with one spine (Fig 2). We simulated seven experimental protocols (Table 1), four of which elicit long-lasting forms of LTP, one of which results in E-LTP, and two of which cause no lasting change in synaptic efficacy. Our goal was to create a simple set of equations to explain all the outcomes, and also molecular dependence of seven protocols. In designing the equation, we concentrated on the activity of molecular species that are implicated in spine-specific and dendrite-specific changes and accompany long term plasticity.
We quantified the spatio-temporal dynamics of molecular species that are known to play a role in the induction of long-lasting forms of LTP, including PKA [9, 12, 55, 60], calcium-calmodulin-dependent protein kinase II (CaMKII) [58, 61–63] and exchange protein directly activated by cAMP (Epac) [14]. These molecules were activated either by calcium pathways or by the βAR coupling either to Gs or Gi. We empirically determined two equations that we called ‘signatures’ to predict the occurrence of long-lasting LTP. The first one summed normalized activity of key molecular species in the spine, the second one summed normalized activity of key molecular species in the dendrite. We assumed that if the experimental protocol enhanced activity of key molecular species in the spine, then spine specific changes would be induced and, similarly, if the experimental protocol enhanced activity of key molecular species in the dendrite, then dendrite specific changes would be induced. To evoke long-lasting forms of LTP both spine specific and dendrite specific changes needed to be induced.
The spine molecular signature trace (referred to as the spine signature) evaluates the initiation of plasticity processes in the spine by calculating time dependent increases in CaMKII, Epac, and PKA activity in the spine:
S spine ( t ) = Δ pCaMKII ( t ) max Δ pCaMKII + Δ Epac ( t ) max Δ Epac + Δ PKA ( t ) max Δ PKA (2)
where ΔEpac(t) is the fold increase in cAMP bound Epac, ΔpCaMKII(t) is fold increase in phosphorylated CaMKII,
ΔPKA(t) is the fold increase in phosphorylation of PKA targets. Max ΔX is a normalization value equal to the maximum activation of molecular specie X among the seven control protocols, where the maximum activation was calculated as the mean (over trials) of the peaks (for each trial). If the spine signature exceeds its threshold for more than 10 sec, spine-specific changes are induced.
The dendritic signature represents spatially non-specific plasticity processes, and takes into account molecular species: PKA, Epac and CaMKII:
S dendrite ( t ) = Δ Epac ( t ) max Δ Epac + Δ pCaMKII ( t ) max Δ pCaMKII + Δ ( pInhibitor 1 ( t ) + pPDE 4 ( t ) ) max Δ ( pInhibitor 1 + pPDE 4 ) + Δ G i ( t ) max Δ G i , (3)
For the dendritic signature, the PKA activity is subdivided into two terms: inhibitory G protein (Gi(t)) which represents phosphorylated β2AR, and other phosphorylated PKA targets: Inhibitor-1 and PDE4. We have subdivided the PKA activity into these two parts to evaluate the role of Gs-Gi switching (and β-arrestin) in synaptic plasticity, and also to evaluate the role of novel β2AR antagonists. Δ(pInhibitor1(t) + pPDE4) represents PKA phosphorylation of other phosphoproteins included in the model for LTP induction. If the dendritic signature exceeds its amplitude threshold for more than the 10 sec duration threshold, dendrite-specific changes are induced.
We chose a relatively short duration threshold as it has been shown that the temporal window of CaMKII activation required for synaptic plasticity and learning is narrow [64], less than 1 minute. To induce long-lasting forms of LTP, both the spine- and dendrite-specific changes must be induced.
The spatial approach allowed us to monitor changes in the phosphorylation of the AMPA receptor subunit GluA1 (AMPAR) in the PSD (Fig 2). We monitored AMPAR phosphorylation (pAMPAR) because it is correlated with E-LTP [70]. To evaluate induction of E-LTP for the seven control protocols, the only additional parameter added was a threshold on AMPA receptor phosphorylation. HFS, ISO+HFS, 4xHFS-3s, 4xHFS-80s and ISO+LFS each cause three-fold increases in phosphorylation of AMPA receptors resembling E-LTP (Fig 10A), whereas ISO causes a smaller increase in phosphorylation of AMPA receptors (Fig 10B), which is in agreement with [71]. Thus, though explaining E-LTP was not a goal of the model, an emergent property was that the model correctly predicts the development of E-LTP. Because only a single additional parameter was added to evaluate the outcomes of seven stimulation protocols, these results are considered an additional validation of the model.
A question of major importance for information processing is which events triggered by synaptic plasticity are spatially specific. Recent experiments using glutamate uncaging at single spines suggest that uncaging induced structural plasticity is spine specific [72]. On the other hand, some molecules, such as Ras, can diffuse into nearby spines, reducing the threshold for LTP at those spines [73, 74]. In addition to spatial specificity, other experiments suggest that stimulation of multiple spines may either cooperate with each other [75] or compete for resources [74]. Thus, the next set of simulations investigated whether electrically induced synaptic plasticity exhibits spatial specificity, i.e., what is the extent of diffusion of key molecules to adjacent spines. We used a 20 μM dendrite with 8 dendritic spines, applied 4xHFS-80s and evaluated stimulation of two adjacent spines (1.5 μm apart) and two non-adjacent (8 μm apart), i.e. separated, spines. Because the model is intrinsically a spatial model, extension of the morphology to a larger dendrite with additional spines requires no changes to reaction rates, molecule concentrations and surface densities, or the equation and thresholds for the signatures.
Both stimulation of separated and adjacent spines produce spine and dendritic signatures that exceed the threshold, and thus are able to induce L-LTP. Fig 11C shows that the dendritic signature exceeds the threshold throughout the dendritic branch. In contrast, Fig 11B reveals some degree of spatial specificity in the spine signature. Statistical analysis shows that for both adjacent and separated spine stimulation, molecular signatures of stimulated spines is greater than molecular signature of unstimulated spines (GLM, stimulus spacing and stimulation as factors, F(2,61) = 163, F >.0001; factor stimulation:P < 0.0001, factor spacing: P = 0.623. For both adjacent and separated spine stimulation, the duration of the spine signature above threshold of stimulated spines is significantly greater than the duration threshold 10 sec, (t-test, T(7) < 0.0001 for both adjacent and separated spine stimulation). In contrast, spine signatures of unstimulated spines are not above threshold for greater than 10 sec (t-test, T(7) = 0.9 for upper threshold, 0.06 for lower threshold for separated spine stimulation; T(7) = 0.79 for upper threshold, 0.016 for lower threshold for adjacent spine stimulation). For both adjacent and separated stimulation protocols, the CaMKII and Epac of the non-stimulated spines is lower than that of the stimulated spine, which is consistent with the gradients observed experimentally [76].
The ability to predict long-lasting forms of LTP does not depend on the precise details of the molecular signatures; instead the LTP predictions are similar for a range of thresholds, and for slight variations in the signature equations. The kinase-to-phosphatase balance, evaluated by molecular signatures, is thought to control direction of synaptic plasticity [36]. There are at least two ways of assessing this balance: either measuring the quantity of phosphorylated targets of kinases and phosphatases Eq (2), or assessing a ratio of kinase activity to phosphatase activity. Importantly, LTP predictions of our model are similar when the spine molecular signature evaluates the ratio of kinases (CaMKII and PKA) to phosphatases (PP1 and PP2B) (S7 Table).
The figures show a threshold range to demonstrate that the model makes the same predictions for any threshold value between the upper and lower thresholds, and does not require a precisely set threshold. To further assess robustness of our results, we evaluated individual simulations (realizations of protocols), that were executed with different random seeds. Note that the stochastic simulation includes a variation in injected quantity, which propagates (in some pathways with amplification) to yield as much as 30% variation in quantity of molecule activation. Tables 2 and 3 show that, despite variability in the time-course, the signatures for each realization of the long-lasting LTP eliciting protocols cross their thresholds for more than 10 sec uninterrupted. Further analysis (Tables 2 and 3) shows that these results are statistically significant. In addition increasing the time the spine signature remains over the threshold to 15 sec, does not significantly change the number of individual simulations that exceed the spine threshold (S8 Table).
To further evaluate robustness of the results, we repeated simulations with variations of two sets of parameters. The first set of parameter variations lowered both AC and PDE4 concentration by 30%. The second set of parameter variations increased AC concentration by 30% and PDE4 concentration by 20%. In both cases, AC and PDE4 quantities were varied together to maintain a 30 nM basal cAMP concentration. Fig 12 shows the mean duration that the spine or dendritic signatures remained above their respective thresholds. Though the signatures varied significantly with parameter variation and trial (as shown by the standard error of the mean), in all cases both signatures were exceeded only for those stimulation protocols that experimentally yield LTP. It is also worth noting that simulations of models with higher AC levels were more noisy because of competition for calmodulin.
To predict long-lasting forms of LTP we developed a stochastic reaction-diffusion model of a dendrite with spines. We looked at activity of the key molecular species during the first 10 min following plasticity induction, because long-lasting LTP is blocked by protein kinase inhibitors applied during or immediately after induction of LTP [57, 77]. A relatively short duration above the threshold is in agreement with [64], showing that temporal window of CaMKII activation required for synaptic plasticity and learning is narrow. We devised a set of molecular signatures: one in the spine and one in the dendrite, that predict induction of long-lasting forms of LTP. We demonstrated that two molecular signatures can explain the results of a large number of experimental protocols. Additional simulations suggested the complex role of the βAR activation in long-lasting forms of LTP. The spatial aspect of these simulations was critical, as a single molecular signature that calculated a spatial average of molecular activity was unable to predict the induction of all forms of long-lasting forms of LTP. Fig 12 clearly shows that the relationship between the dendritic signature and the spine signature depends on the stimulation protocol.
Separate molecular signatures in the spine and in the dendrite represent distinct phenomena. Two signatures can be viewed as corresponding to synaptic tagging and capture [63, 65], a theory explaining how signaling molecules in different spatial compartments play different roles in L-LTP. Synaptic tagging involves labeling of specific dendritic spines that are to undergo long term plasticity, and capture implies that a spatially non-specific signal induces synthesis of plasticity related proteins (PRPs), and in some cases, initiates transcription [78]. PRPs are synthesized locally or trafficked up the dendrite and captured by tagged spines to stabilize synaptic strength. Crossing the threshold by the spine molecular signature can be viewed as setting the tag and crossing the threshold by the dendritic molecular signature corresponds to sending the signal initiating the synthesis of PRPs.
In constructing the spine molecular signature, we evaluated molecules that are implicated in synaptic tagging, AMPA receptor insertion, actin remodeling and structural plasticity [72, 79–82] (Fig 13). Blocking CaMKII activity [61–63] has been shown to block tagging, and CaMKII also is implicated in the actin remodeling underlying structural plasticity [83–85] by triggering SynGAP dispersion from synaptic spines [86]. PKA is required for synaptic tagging [56, 66, 87, 88] and is implicated in structural plasticity. PKA modulates the activity of LIM kinase [89, 90], which phosphorylates (and inhibits) cofilin allowing for actin polymerization. Cofilin-mediated actin dynamics regulates spine morphology and AMPAR trafficking during synaptic plasticity [91, 92]. Epac anchors in the PSD [93] and triggers changes to spine cytoskeleton via Rap1 activation [94]. Interestingly, synapses stimulated by HFS while blocking PKA activity fail to be tagged [88], whereas ISO+HFS stimulation while blocking PKA still yields L-LTP [14]. Our simulations suggest that this seemingly contradictory result arises from the difference between the amount of Epac provided by HFS alone versus ISO+HFS. The plausibility of the spine signature is evident from its time course, which is comparable to the dynamics of molecular activation measured using live cell imaging [80].
The molecular signature in the dendrite takes into account molecules that play a role in synthesis of PRPs (Fig 13). Both PKA and Epac activate ERK via Rap1 regulation [95–98]. Also, PKA phosphorylation of β2AR can produce ERK activation by switching the β2AR coupling from Gs to Gi [15–17], though this has not been directly demonstrated in neurons. ERK has been shown to be critical in L-LTP [12, 13, 55, 99–101] and the synthesis of PRPs [61]. Both PKA and ERK can phosphorylate CREB, a molecule directly implicated in transcription. CaMKII is required for regulation of protein synthesis via phosphorylation of cytoplasmic polyadenylation element binding protein [102, 103] in hippocampal plasticity, but see [61, 62]. Though both spine and dendritic signatures incorporated the same molecules, they have different downstream targets in the spine and in the dendrite. Thus the two molecular signatures set the stage for future models that incorporate control of actin dynamics in the spine and ERK activation in the dendrite.
Several other models have evaluated molecular dependence and temporal sensitivity of L-LTP induction. The most comprehensive model of signaling pathways leading to transcription of mRNA [104] demonstrated that different temporal stimulation patterns could recruit different mRNAs. In agreement with their results, our simulations showed that different stimulation patterns produced different patterns of elevation of various kinases. It would be quite interesting to couple our dendritic model to downstream modules of the model presented in [104] to evaluate control of transcription by L-LTP stimulation patterns. Several other models investigated synaptic tagging and capture [105–107] at hippocampal CA3-CA1 synapses. All of these models were able to predict various aspects of the synaptic tagging and capture hypothesis. Nonetheless, these models used simplified and abstract equations for activation of key kinases and phosphatases; thus it is not clear how well they could extrapolate to alternative stimulation patterns. Another model [108] also used streamlined equations for activation of key kinases and phosphatases, but included a model of histone deacetylation, which regulates transcription [109]. That model suggested that promoting histone acetylation while simultaneously slowing cAMP degradation could help in restoring L-LTP, which is impaired in mouse models of Rubinstein-Taybi syndrome, a condition resulting in lower levels of CREB binding protein, which reduces transcription.
Our simulations of a dendrite with multiple spines are consistent with the spatial specificity of homo- and heterosynaptic plasticity suggested by imaging of spine morphological plasticity. Stimulation of two spines on the same branch produces a dendritic signature that crosses the threshold along the entire branch, regardless of the spatial configuration of those stimulated spines. This result is consistent with [75], showing that one train of 5 Hz stimulation applied to two spines on the same branch saturates ERK activation in that branch. During these simulations, spine signatures of the unstimulated spines are elevated, although lower than those of the stimulated spines. This observation is consistent with the gradients observed experimentally [76]. Furthermore, the increase in signature of non-stimulated spines is consistent with the observation of a reduced LTP threshold heterosynaptically [73]. It is, however, also possible that not all spines will exhibit potentiation due to competition for resources, as in [74]. Our model does not take into account this competition, but such a model would allow only the spines with the highest signatures to capture PRPs, and thus non-stimulated spines with lower signatures would not exhibit LTP. The agreement between these simulations and experiments suggests the model could be used to predict the spatial pattern of LTP in response to in vivo like stimulation patterns.
We evaluated AMPAR phosphorylation by CaMKII and PKA as an indicator of E-LTP, and found agreement between our simulations and experimental results [70, 110, 111]. The brief duration of the AMPAR phosphorylation in our model is likely due to absence of AMPAR re-cycling mechanisms [112]. Previous work has shown AMPAR recycling contributes to bistability [113], and insertion of a phosphorylated AMPAR may protect it from dephosphorylation. Alternatively, AMPAR phosphorylation may only be a trigger for insertion, and the time course of E-LTP may reflect the removal of AMPARs in the synapse.
Induction of long-lasting LTP initiates a cascade of complex molecular interactions; therefore signaling pathway modeling is a useful approach to facilitate understanding of this complexity. In addition to confirming the plasticity outcome and molecular dependence for numerous LTP induction protocols, our model makes several experimentally testable predictions. Our model suggests that βAR signaling through non-conventional pathways is necessary in the dendrite, therefore ICI-118,551, a complete βAR antagonist, will likely block long-lasting LTP induced with 4xHFS-80s, a model prediction that needs to be tested experimentally. Moreover, the model suggests that both conventional (Gs-activated) and non-conventional (Gi-activated) pathways are required for ISO+LFS and ISO+HFS to produce long-lasting LTP, therefore we predict that bath application of carvedilol, which blocks norepinephrine binding but allows Gi recruitment, will not induce long-lasting LTP. Simulations of bath application of carvedilol followed by one, two and three trains of HFS shows that high enough calcium might substitute for Gs activation in L-LTP induction, but that both Gs and Gi might be necessary for L-LTP induction using LFS. Though our model focuses on βAR signaling, CA1 neurons express dopamine receptors, which have been implicated in some forms of long-lasting LTP [114]. If such receptors are shown to undergo switching of Gs to Gi coupling, then these receptors also may contribute to a plethora of long-lasting forms of LTP. In summary, our model suggests that the non-linearity of signaling pathway interactions may explain why experimentally blocking any of the molecules included in our signature can disrupt long-lasting LTP.
|
10.1371/journal.pntd.0001798 | Involvement of CD4+ Foxp3+ Regulatory T Cells in Persistence of Leishmania donovani in the Liver of Alymphoplastic aly/aly Mice | Visceral leishmaniasis (VL) is a chronic and fatal disease in humans and dogs caused by the intracellular protozoan parasites, Leishmania donovani and L. infantum (L. chagasi). Relapse of disease is frequent in immunocompromised patients, in which the number of VL cases has been increasing recently. The present study is aimed to improve the understanding of mechanisms of L. donovani persistence in immunocompromised conditions using alymphoplastic aly/aly mice. Hepatic parasite burden, granuloma formation and induction of regulatory T cells were determined for up to 7 months after the intravenous inoculation with L. donovani promastigotes. While control aly/+ mice showed a peak of hepatic parasite growth at 4 weeks post infection (WPI) and resolved the infection by 8 WPI, aly/aly mice showed a similar peak in hepatic parasite burden but maintained persistent in the chronic phase of infection, which was associated with delayed and impaired granuloma maturation. Although hepatic CD4+Foxp3+ but not CD8+Foxp3+ T cells were first detected at 4 WPI in both strains of mice, the number of CD4+Foxp3+ T cells was significantly increased in aly/aly mice from 8 WPI. Immunohistochemical analysis demonstrated the presence of Foxp3+ T cells in L. donovani–induced hepatic granulomas and perivascular neo-lymphoid aggregates. Quantitative real-time PCR analysis of mature granulomas collected by laser microdissection revealed the correlation of Foxp3 and IL-10 mRNA level. Furthermore, treatment of infected aly/aly mice with anti-CD25 or anti-FR4 mAb resulted in significant reductions in both hepatic Foxp3+ cells and parasite burden. Thus, we provide the first evidence that CD4+Foxp3+ Tregs mediate L. donovani persistence in the liver during VL in immunodeficient murine model, a result that will help to establish new strategies of immunotherapy against this intracellular protozoan pathogen.
| The protozoan parasite Leishmania donovani is the causative agent of visceral leishmaniasis (VL) with a variety of outcomes ranging from asymptomatic to fatal infection. In the last decade, an increasing number of VL cases in immunocompromised conditions have been reported. Loss of the control of parasite persistence causes relapse of the disease in these patients. To clarify why parasite persistence and disease are caused in an immunocompromised condition, we examined L. donovani infection in alymphoplastic aly/aly mice that completely lack lymph nodes and have disturbed spleen architecture. Although parasites grew in the liver of aly/+ mice for the first 4 weeks post infection (WPI) and parasites were eliminated by 8 WPI, we found that parasites persisted in the liver of aly/aly mice with the ineffective of granuloma formation to kill the parasites. These aly/aly mice showed significant increases in CD4+Foxp3+ regulatory T cells in the liver. Consequently, we treated infected mice with anti-CD25 or anti-FR4 mAb to inhibit the function of Tregs, and found significant reductions in both hepatic Foxp3+ cells and parasite burden. These results clearly demonstrated for the first time that the expansion of CD4+Foxp3+ Tregs is involved in hepatic L. donovani persistence in immunodeficient murine model.
| Visceral leishmaniasis (VL) is a chronic and fatal disease caused by the intracellular protozoan parasites Leishmania donovani and L. infantum (chagasi), which infect a range of mammalian hosts, including humans, dogs and rodents [1]. Liver, spleen, bone marrow (BM) and lymph nodes are the major sites for parasite growth and disease pathology. Transplantation of infected kidney, liver, heart, lung, pancreas or BM has been shown to cause VL in transplant recipients, indicating lifelong parasite persistence in the viscera [2]. Moreover, malnutrition is a risk factor for the development of VL [3]. Recent experiments in protein energy-, zinc- and iron-deficient mice suggest that this effect is mediated primarily through functional failure of the lymph node barrier and increased early visceralization of the parasites [4]–[6]. Loss of the control of parasite persistence in VL causes the reactivation of parasites and relapse of the disease is frequent in the immunocompromised patients, in which the number of visceral leishmaniasis cases has been increasing recently [7]. However, the mechanisms underlying the parasite persistence in the immunocompromised condition have not been clearly clarified. To develop effective prophylactic or therapeutic strategies against VL, understanding of the precise immune mechanisms including T-cell functions in the chronic stage of infection is required [8].
The role of secondary lymphoid organs for immune responses to Leishmania infection has not been investigated. The aly/aly mouse is an autosomal recessive natural mutant C57BL/6 strain that carries a point mutation within the gene encoding NF-κB inducing kinase (NIK) [9], which prevents the induction of the non-canonical NF-κB pathway [10]. The aly/aly mice lack all lymph nodes and Peyer's patches with the abnormal architecture of spleen and thymus and exhibit severely impaired humoral response [9]. This mutant mouse strain has been used to examine the role of secondary lymphoid organs for immune responses to intracellular pathogens, including Mycobacterium leprae, Listeria monocytogenes, vesicular stomatitis virus, vaccinia virus, lymphocytic choriomeningitis virus and human T-cell leukemia virus [11]–[14], and different susceptibilities to these pathogens have been reported.
Organ-specific immunity has been described in various experimental VL studies in mouse models [15], [16]. The liver is the site of an acute but resolving infection. In contrast, the spleen becomes a site of parasite persistence with associated immunopathological changes [17]. In BALB/c and C57BL/6 mice, the inflammatory granuloma reaction around infected Kupffer cells is developed and the infection is resolved by 4–8 weeks after infection [18]. However, low levels of hepatic parasite persistence for 6–12 additional months occur and administration of anti-CD4 antibodies result in the relapse of hepatic quiescent L donovani infection [19], suggesting that CD4+ T cells are required for the maintenance of acquired immunity and prevention of relapse. However, no additional data explaining the underlying mechanisms of CD4+ T cell-mediated control of persistent parasites have been presented. Cellular and molecular interactions mediated by Kupffer cells, monocytes, CD4+ and CD8+ T cells and a number of cytokines and chemokines are required for effective hepatic granuloma formation [16]–[18], [20], [21]. Defects in these cellular and molecular factors cause ineffective parasite clearance from the liver, but most murine studies have focused on the first few weeks of infection and not the persistent stage of infection [18].
The present study is aimed to improve the understanding of mechanisms of L. donovani persistence in an immunocompromised condition. Our data presented herein offered a novel insight into the involvement of CD4+Foxp3+ regulatory T cells (Tregs) in L. donovani persistence in the liver of immunodeficient aly/aly mice. Moreover, treatment of infected aly/aly mice with anti-CD25 or anti-FR4 mAb revealed the significant reductions in both hepatic Tregs and parasite burden. These results suggest that manipulation of Tregs may provide a promising immumotherapeutic strategy for VL.
Female ALY® NscJcl aly/aly and aly/+ mice of 6–8 weeks of age were purchased from CLEA Japan, Inc. (Tokyo, Japan). Mice were maintained, inoculated and sacrificed within a safety facility of Hokkaido University.
A virulent line of L. donovani (MHOM/SU/62/2S-25M-C2) [22] was maintained by passage of the frozen stabilized parasites in NNN medium containing 5% defibrinated hemolyzed rabbit blood. Then, parasites were consecutively sub-passaged in liquid M199 medium supplemented with 15% heat-inactivated fetal calf serum (HIFCS), 25 mM HEPES and 50 µg/ml gentamycin. The stationary growth phase of subcultures with less than five passages was used for mouse inoculation. Mice were infected by injecting stationary phase promastigotes (5×107) intravenously via the lateral tail vein and were sacrificed at 1, 2, 4, 8, 12, 16 and 28 weeks post infection (WPI). One group of non-infected animals was used as naïve control.
This study was carried out under the guidance of the Institute for Laboratory Animal Research (ILAR). All animals were housed in a facility in strict accordance with the recommendations in the Guidelines for the Care and Use of Laboratory Animals of Graduate School of Veterinary Medicine, Hokkaido University, which was based on Fundamental Guidelines for Proper Conduct of Animal Experiment and Related Activities in Academic Research Institutions under the jurisdiction of the Ministry of Education, Culture, Sports, Science and Technology, Japan and approved by the American Association for Accreditation of Laboratory Animal Care (AAALAC) international. The protocol was approved by the Committee on the Ethics of Animal Experiments of Hokkaido University (Permit Number: 10-0009).
Giemsa-stained impression smears of the liver were prepared and parasite burden was determined as Leishman-Donovan Units (LDU), in which LDU is the number of amastigotes per 1,000 host nuclei, multiplied by the liver weight in gram [23].
Genomic DNA (gDNA) was isolated from different tissues, including liver, spleen, BM, blood, heart, lung, kidney, brain and skin, using the QIAamp® DNA Mini Kit (Qiagen, MA, USA). Real-time quantitative (qPCR) assays were performed on the StepOne™ and the StepOnePlus™ Real-Time PCR Systems (Applied Biosystems, CA, USA), following the manufacturer's instructions. A typical 20-µl reaction mixture contained approximately 100 ng gDNA, 1× SYBR® Premix Ex Taq™ II (Takara, Tokyo, Japan), 0.4 µM each primer (Table S1) and 1× Rox™ Reference Dye. All samples were run in triplicate and underwent an initial 30 sec incubation step at 95°C, followed by 40 cycles of 5 sec at 95°C and 30 sec at 65°C for the Leishmania surface protease gp63 gene or 60°C for the mouse brain-derived neurotrophic factor (mBDNF) gene [24], [25]. The average threshold cycle of amplification (Ct) values was determined, and standard deviation (SD) of all the reaction was analyzed by the software provided with the instrument. The relative amounts of the gp63 gene were then calculated using standard curve method normalized to the amounts of the mBDNF gene.
The livers were fixed in 10% neutral phosphate-buffered formalin. Paraffin-embedded organs were cut into 4 µm-thick sections, followed by staining with hematoxylin and eosin for light microscopy. For the detection of parasites, liver sections were subjected to indirect immunohistochemical staining using L. infantum-infected dog serum (1∶1000 dilution) [26] and horseradish peroxidase (HRP)-conjugated goat anti-dog IgG heavy and light chain antibody (1∶300; Bethyl Laboratories, TX, USA). Peroxidase was visualized using 3,3′-diaminobenzidine (DAB)-H2O2 (Wako, Tokyo, Japan) and the sections were counterstained with Mayer's hematoxylin before dehydration and mounting.
Hepatic immune responses were categorized into (1) “No granuloma”: no inflammation with no mononuclear cell (MNC) around the parasitized Kupffer cells; (2) “Immature granuloma”: less than 10 MNCs around the parasitized Kupffer cells; (3) “Mature granuloma”: epithelioid cells and more than 10 MNCs around the parasitized Kupffer cells; and (4) “Involuting granuloma”: devoid of amastigotes and tissue inflammatory nearly resolved [18], [23]. The number of infected foci with each tissue response including “No granuloma”, “Immature granuloma”, “Mature granuloma” and “Involuting granuloma” was counted for 25 consecutive microscopic fields per mouse liver at ×400 magnification.
Hepatic mononuclear cells were isolated using a 33% (vol/vol) Percoll solution, as described elsewhere [27]. Briefly, livers were minced, pressed through a stainless steel mesh and suspended in RPMI1640 medium (Sigma, MO, USA) supplemented with 3% HIFCS (wash buffer). After washing, the cells were resuspended in 33% Percoll solution containing heparin (100 U/ml) and centrifuged at 800× g for 30 min to remove liver parenchymal cells. The pellet was treated with an RBC lysis solution (155 mM NH4Cl, 10 mM KHCO3, 0.1 mM EDTA), washed and re-suspended in 2.4G2 mAb solution to block the Fc receptor before staining with antibody. Antibodies used for FACS included PE-labeled rat anti-mouse CD4 (L3T4) (BD Pharmingen, CA, USA), FITC-labeled rat anti-mouse CD8a (Lyt-2) (BD Pharmingen), APC-labeled rat anti-mouse Foxp3 (FJK-16s) (eBioscience, CA, USA) and the proper isotype staining control, according to the manufacturer's instructions. Flow cytometry analysis of the labeled cells was performed on a FACS Calibur (BD Pharmingen), running the Cell Quest program provided with the instrument. Lymphocytes were identified by forward scatter (FSC) and side scatter (SSC) characteristics, gated and further analyzed with Cell Quest software (BD Pharmingen) or FlowJo software V. 5.7.2 (Tree Star Inc., OR, USA).
Immunohistochemical analysis of the 4 µm-thick paraffin-embedded sections of the liver was performed to determine the localization of Foxp3+ Tregs. After deparaffinization and rehydration, heat-induced epitope retrieval (HIER) was conducted by autoclaving at 100°C for 17 min using Target Retrieval Solution (pH 9.0) (Dako, Uppsala, Sweden). Endogenous peroxidase was blocked by incubating sections in 0.3% H2O2 in absolute methanol for 30 min at 4°C, followed by flushing with water and incubation with 10% goat serum for 1 h at room temperature (RT) to block crystallized receptor fragments. The sections were incubated overnight with rat anti-mouse/rat Foxp3 mAb, clone FJK-16s (eBioscience), in 1∶100 diluted with 0.1% Triton X in PBS (pH 7.4). For negative control sections, PBS was used instead of the primary antibody. After washing three times in PBS (5 min each), sections were incubated in 1∶100 biotin-conjugated goat anti-rat IgG (H+L) antibody (Invitrogen, MD, USA) for 30 min at RT. Sections were then washed, which was followed by incubation with streptavidin-peroxidase conjugate (Histofine SAB-PO® Kit) for 30 min at RT. The streptavidin-biotin complex was visualized with DAB-H2O2 solution, pH 7.0, for 4 min. Sections were washed in distilled water, and finally counterstained with Mayer's hematoxylin. The mean counts of Foxp3-expressing cells were assessed microscopically at 400× magnification by counting a total of 25 consecutive fields. The number of immunoreactive cells was estimated in each hepatic granuloma assembly. Values are expressed as the means of immunoreactive cells present in 25 fields.
Double immunofluorescence staining was also conducted to locate Tregs in the liver. Formalin-fixed and paraffin-embedded liver sections were subjected to the deparaffinization, rehydration and HIER as described above. After blocking of crystallized receptor fragments with 10% goat serum, sections were incubated overnight with rat anti-mouse/rat Foxp3 mAb (clone FJK-16s; 1∶100; eBioscience) at 4°C. Then, the sections were incubated with FITC-goat anti rat IgG (1∶200; Zymed, CA, USA) for 30 min at RT and successively incubated in 10% donkey or rabbit serum to block the crystallized receptor fragments. For double staining of Foxp3-expressing cells and T cells, the sections were incubated with rabbit anti-mouse CD3 mAb (1∶200; Nichirei) overnight at 4°C and then with TRITC-donkey anti rabbit IgG (1∶200; Abcam, MA, USA) for 30 min at RT. On the other hand, double staining of Foxp3-expressing cells and L. donovani amastigotes was conducted using L. infantum-infected dog serum (1∶1,000) and TRITC-goat anti dog IgG (1∶200; Rockland, PA, USA). Finally, the sections were mounted using a Fluoromount™ (DBS, CA, USA) and examined under an IX70 confocal microscope (Olympus, Tokyo, Japan).
Laser microdissection (LMD) was performed in RNase-free conditions as described previously [28]. Cryosections of 7 µm thickness were prepared from the frozen livers of naïve and infected mice and embedded in Tissue-Tek OTC compound (Sakura, Tokyo. Japan). The sections were mounted on glass slides pre-coated with LMD films (Meiwafosis, Osaka, Japan) and fixed with absolute methanol for 3 min at 4°C. After staining with 0.5% toluidine blue for 10 sec, approximately 20 “Mature granulomas” were microdissected from each frozen liver sample by using Ls-Pro300 (Meiwafosis).
Total RNA was purified from the frozen whole liver tissue and microdissected “Mature granulomas”, using the RNAqueous®-Micro Kit (Ambion, Texas, USA). Expression levels of Foxp3, TGF-β and IL-10 mRNA were determined by quantitative RT-PCR (qRT-PCR) using the PrimeScript™ RT Reagent Kit (Takara) and the relative number of these molecules to 1000 housekeeping glyceraldehyde-3-phosphate dehydrogenase (GAPDH) was calculated using a standard curve method. The PCR reaction was performed as described above using primers shown in Table S1 [29], [30].
At 26 WPI, three L. donovani-infected aly/aly mice were intraperitoneally injected three times every other day with 0.5 mg of rat anti-mouse CD25 mAb (clone PC61; Biolegend, CA, USA), 0.05 mg of rat anti-mouse FR4 mAb (clone TH6; Biolegend) or 0.5 mg of rat IgG (Jackson ImmunoResearch, PA, USA) as a control. The mice were euthanized at 10 days post-antibody injection for examination of host responses as described above.
Statistical differences between aly/aly mice and aly/+ mice at the indicated time points were tested using Student's t-test (Microsoft Excel software) and two-way ANOVA as well as post hoc Bonferroni test (Prism software version 5, GraphPad, CA, USA). All data are presented as the mean values ± SE unless otherwise stated. p<0.05 was considered as statistically significant.
Long-term persistence after clinical cure of the primary infection is a characteristic feature of many intracellular pathogens, including protozoan parasites of the genus Leishmania, but the underlying mechanisms are not fully understood [31]. We measured parasite burdens in the livers of aly/+ and aly/aly mice for up to 28 WPI by two different methods. The number of amastigotes in hepatic impression smears was expressed as LDU (Figure 1A), and relative amounts of Leishmania gp63 gene to mBDNF gene were determined by qPCR (Figure 1B). In aly/+ mice, parasite burden peaked at 4 WPI and reduced to near-baseline levels by 8 WPI. In aly/aly mice, parasite burden also peaked at 4 WPI but the maximum parasite burden was lower than that of aly/+ mice. Although the parasite load decreased by 8 WPI as observed in aly/+ mice, the parasite persisted in the liver of aly/aly mice during the observation period of 28 WPI (Figure 1). Persistent L. donovani infection was also demonstrated in the spleen and BM of both mice strains but the parasite burden was much higher in aly/aly mice during the chronic phase of infection. Nevertheless, parasite was not detected in the skin and internal organs, such as lung, kidney, heart and brain during infection by qPCR (data not shown).
Efficient granuloma development around infected Kupffer cells is a key event in the control of hepatic L. donovani infection [8], [16], [18]. The infected foci in the liver were examined and made a quantitative analysis of granuloma formation around the parasitized Kupffer cells. The progression of granuloma formation from “No granuloma” to “Immature granuloma”, “Mature granuloma” and finally “Involuting granuloma” was observed in aly/aly mice as well as aly/+ mice (Figure S1), indicating that aly/aly mice have ability to generate hepatic cell-mediated immunity to some extent as shown in the previous study [32].
The number of infected foci was well correlated with the hepatic parasite loads (Figure 1); the number of foci in aly/+ mice reached a peak at 4 WPI and was drastically reduced at 8 WPI (Figure 2A) and the involuting granuloma was well formed (Figure 2B). However, the number of involuting granuloma in the liver of aly/aly mice were much less than those in aly/+ mice at 4 and 8 WPI (Figure 2C) while the 30–40% of the infection foci with no granulomas was found in the liver of aly/aly mice at 4–16 WPI. These may reflect that effective but insufficient clearance of the parasites in granuloma of aly/aly livers renders persistent release of the parasites, which results in increased proportion of infected Kupffer cells in the later stages.
Foxp3+ Tregs influence immunity to viral, bacterial or parasitic infections [33]. To begin to characterize the mechanism by which parasites persist in the liver, we examined whether Tregs expand in the livers of aly/aly mice and where they localize during L. donovani infection.
Flow cytometry analysis of hepatic lymphocytes revealed no expansion of CD8+Foxp3+ T cells in the liver during L. donovani infection in either strain of mice (Figure 3A). In contrast, CD4+Foxp3+ T cells were first detected at 4 WPI in both strains of mice. In aly/aly mice, the proportions of CD4+Foxp3+ T cells to CD4+ T cells (Figure. 3B) as well as the absolute number of CD4+Foxp3+ T cells (Figure 3C) were higher than those of aly/+mice especially at 8–16 WPI although the total number of hepatic CD4+ T cells was not significantly different between aly/+ and aly/aly mice (Figure S2).
There have been no reports describing the localization of Foxp3-expressing cells in the liver during VL. To address this, we stained Foxp3 in liver sections of naïve and L. donovani-infected aly/+ and aly/aly mice. Foxp3-expressing cells were localized in the “Immature granuloma” and “Mature granulomas” as well as the perivascular areas of infected aly/aly mice. Furthermore, the density of Tregs increased, especially in the perivascular areas, during the course of infection (Figure 4A and B). Development of such abnormal lymphocyte infiltration or neo-lymphoid aggregates at perivascular areas is a feature found in aly/aly and other alymphoplastic mice [32]. In addition, the frequency of “Mature granulomas” containing more than 5 Tregs increased during infection in aly/aly mice (5% at 4 WPI, 18% at 12 WPI and 39% at 28 WPI), suggesting the accumulation of Tregs at sites of inflammatory foci. On the other hand, Foxp3-positive Tregs were limited to the parenchyma, granulomas and perivascular areas at 4 WPI and hardly detectable in the liver of infected aly/+ mice at 12 WPI (Figure 4A and B).
Double immunofluorescence analysis of hepatic granuloma revealed that Foxp3-expressing cells (green in Figure 5A) and CD3+ cells (red in Figure 5A) were present in the granuloma, and Foxp3+ cells expressed CD3 molecules (Figure 5A-merged image). Some CD3+Foxp3+ cells (yellow arrows in Figure 5A-merged image) were adjacent to the CD3+Foxp3− cells (pink arrows in Figure 5A-merged image). In addition, L. donovani amastigotes (red in Figure 5B) were surrounded by Foxp3+ cells (green in Figure 5B) in the hepatic granuloma. These results suggested that the interaction among parasitized cells (Kupffer cells), CD3+Foxp3+ cell (Tregs) and CD3+Foxp3− cells (non Tregs, probably CD4+ and/or CD8+ effector T cells).
Evidence has accumulated regarding the essential roles of Tregs in the control of a variety of physiological and pathological immune responses, but it is still obscure how Tregs control other lymphocytes at the molecular level [34]. Quantitative RT-PCR was performed for Foxp3, IL-10 and TGF-β mRNA levels in the whole liver and micro-dissected “Mature granulomas” liver tissue samples of L. donovani-infected aly/aly mice. The Foxp3 mRNA expression was increased after infection in the whole liver (Figure 6A) and mature granuloma samples (Figure 6B). Although the TGF-β mRNA transcripts showed similar levels at 4 and 12 WPI in both tissue samples, the levels of IL-10 mRNA markedly increased in mature granuloma but not whole liver samples at 12 WPI (Figure 6A and B), suggesting that IL-10 may be involved in function of Tregs.
Manipulation of Tregs by treatment with antibodies has been used to examine the roles of Tregs in many infectious diseases [33]. Effects of anti-CD25 and anti-FR4 mAb on hepatic immune responses in L. donovani-infected aly/aly mice at 26 WPI were examined. Ten days after injection with anti-CD25 or anti-FR4, reduction in Foxp3 mRNA expression was observed (Figure 7A). This reduced Foxp3 mRNA expression was associated with decreases in parasite burden (Figure 7B) and infected foci (Figure 7C). Instead, the frequency of “Mature granulomas” was increased after treatment with especially anti-FR4 mAb (Figure 7D), suggesting that depletion of Tregs can activate hepatic cellular immune responses and accelerate parasite killing. Furthermore, immunohistochemical analysis confirmed a reduction in Foxp3-immunoreactive cells in the liver parenchyma, granulomas (Figure 7E) and perivascular neo-lymphoid areas (data not shown).
In the present study, aly/aly mice were used as an immunodeficient VL murine model and immunohistopathologically investigated during L. donovani infection for up to 28 WPI. CD4+Foxp3+ T cells were increased in the granulomas and perivascular areas of the liver in the chronic phase and the impairment of granuloma maturation was observed. The depletion of Tregs by the administration of either anti-CD25 or anti-FR4 mAb resulted in significant reductions in hepatic Tregs, infected foci and parasite burden. To our knowledge, this is the first definitive evidence that CD4+Foxp3+ Tregs are involved in hepatic L. donovani persistence in a murine model of VL.
The aly/aly mice have been used to examine the role of secondary lymphoid organs on immune responses in various infection models. Disruptive architecture of the thymus and spleen could affect the development and expansion of T cells. Several studies using bone marrow chimeras between aly/aly and wild type mice showed that antiviral CTL responses were clearly improved in the wild type environment [11]. However, expansion of CD25+CD4+ Treg is impaired in the spleen of aly/aly mice [35]. This suggests that expansion of functional CD4+Foxp3+ Treg in the liver of aly/aly mice during L. donovani infection is likely related to the parasite persistence but not to the structural defects of secondary lymphoid organs although this possibility will be confirmed by BM chimera experiments in future.
The NIK gene mutation may contribute to other immune defects due to the partial blocked NF-κB activation [10], [36]. NF-κBp52 knockout mice showed less parasite burden in the liver, perhaps due to less number of B cells (unpublished data; Ato M., Kaye PM). NF-κBp50 (NF-κB1) would be important for TNF/TLR signaling which is involved in canonical TLR/TNFR signaling for activation of dendritic cells and macrophages. NIK is associated in CD40/LT-αβR but not in TNF. CD40 signaling is one of DC activation factors, but the function of DC of aly/aly are controversial. Yamada et al [36] has reported that DC from aly/aly mice exhibit grossly normal development and function. However, Tamura et al [35] had reported that DCs from aly/aly mice showed impaired antigen presentation ability. Lower hepatic parasite loads was unexpectedly observed in aly/aly mice than aly/+ mice in the first 4 WPI. This may be not due to lower number of the sessile Kupffer cells (unpublished data), but associated with the strong innate immunity as reported during Listeria monocytogenes infection in aly/aly mice [12]. Partial hepatic granuloma progression and neo-lymphoid aggregates in aly/aly mice imply that mice lacking secondary lymphoid tissues can still generate T cell-mediated immune responses to some extent [32].
Although anti-CD25 mAbs have been used for depletion of Tregs in various experimental cases, administration of anti-FR4 mAb also reduced Treg numbers and provoked effective tumor immunity [37]–[39]. In the present study, 10 days after the third injection of infected aly/aly mice with anti-CD25 and anti-FR4 mAb, the hepatic parasite burdens were reduced by 88% and 89% of that of control mice, respectively (Figure 7B). Likewise, treatment with either anti-CD25 or anti-FR4 mAb also reduced parasite burdens in the spleen and BM (Figure S3). The reason why anti-FR4 mAb was more effective than anti-CD25 mAb in reducing parasite burden is unknown, but the present study is the first to report effectiveness of anti-FR4 mAb to control systemic infection of L. donovani in mice. Thus, anti-FR4 antibodies may be an alternative measure to manipulate Tregs in chronic VL. However, since anti-CD25 mAb can also affect effector T cells and effective immunity [39], and anti-FR4 mAb can also deplete a small population of CD4+ Foxp3− T cells in the lymph node [38], probably including IL-10-producing conventional CD4+ T cells, further studies of the role of Tregs in VL are required.
Studies of Tregs in cutaneous leishmaniasis demonstrated the involvement of CD4+CD25+ Tregs in cutaneous leishmaniasis caused by L. major [40], [41] and L. amazonensis in mice [42] and by L. braziliensis in humans [43]. Regarding VL, the role of Tregs is uncertain and the primary source of IL-10 is controversial. In the spleen of VL patients in India, CD4+CD25−Foxp3− cells were identified as the major producers of IL-10 [44]. In L. infantum-infected BALB/c mice, CD4+CD25+ Foxp3+ cells expanded in a pooled fraction of draining lymph nodes and spleen cells at 7 and 28 days of infection [45]. In L. donovani-infected BALB/c mice, the number of splenic CD4+ CD127dimCD25+GITR+ T cells expressing higher Foxp3 and IL-10 increased at 21 days of infection [46]. IL-10 production by splenic CD4+CD25−Foxp3− IL10+ T cells, representing type 1 regulatory T (Tr1) cells, was a strong correlate of disease progression in L. donovani-infected C57BL/6 mice [47]. Further analyses using quantitative RT-PCR of IL-10 and Foxp3 transcripts in selected populations of CD25+ and CD25− enriched hepatic CD4+ T cells, and/or by intracellular cytokine staining, will elucidate the issue. Nevertheless, in the present study, Treg and IL-10 augment immunosuppressive effects in hepatic granuloma of L. donovani-infected aly/aly mice. Maintenance of relatively higher expression levels of TGF-β in the chronic phase of the infection in aly/aly mice may be related to the generation and maintenance of CD4+Foxp3+ Tregs [48] rather than the inhibition of granuloma maturation [49].
In conclusion, we focused on immune responses to the chronic phase of murine VL caused by L. donovani infection in an immunodeficient host. In the last decade when the number of visceral leishmaniasis in immunocompromised patients has been increasing, our data presented herein offered a novel insight into the possibly involvement of CD4+Foxp3+ Tregs in persistent L. donovani infection in the liver of immunodeficient hosts. The manipulation of Tregs may provide a promising immumotherapeutic strategy for VL.
|
10.1371/journal.pgen.1002491 | A Meta-Analysis and Genome-Wide Association Study of Platelet Count and Mean Platelet Volume in African Americans | Several genetic variants associated with platelet count and mean platelet volume (MPV) were recently reported in people of European ancestry. In this meta-analysis of 7 genome-wide association studies (GWAS) enrolling African Americans, our aim was to identify novel genetic variants associated with platelet count and MPV. For all cohorts, GWAS analysis was performed using additive models after adjusting for age, sex, and population stratification. For both platelet phenotypes, meta-analyses were conducted using inverse-variance weighted fixed-effect models. Platelet aggregation assays in whole blood were performed in the participants of the GeneSTAR cohort. Genetic variants in ten independent regions were associated with platelet count (N = 16,388) with p<5×10−8 of which 5 have not been associated with platelet count in previous GWAS. The novel genetic variants associated with platelet count were in the following regions (the most significant SNP, closest gene, and p-value): 6p22 (rs12526480, LRRC16A, p = 9.1×10−9), 7q11 (rs13236689, CD36, p = 2.8×10−9), 10q21 (rs7896518, JMJD1C, p = 2.3×10−12), 11q13 (rs477895, BAD, p = 4.9×10−8), and 20q13 (rs151361, SLMO2, p = 9.4×10−9). Three of these loci (10q21, 11q13, and 20q13) were replicated in European Americans (N = 14,909) and one (11q13) in Hispanic Americans (N = 3,462). For MPV (N = 4,531), genetic variants in 3 regions were significant at p<5×10−8, two of which were also associated with platelet count. Previously reported regions that were also significant in this study were 6p21, 6q23, 7q22, 12q24, and 19p13 for platelet count and 7q22, 17q11, and 19p13 for MPV. The most significant SNP in 1 region was also associated with ADP-induced maximal platelet aggregation in whole blood (12q24). Thus through a meta-analysis of GWAS enrolling African Americans, we have identified 5 novel regions associated with platelet count of which 3 were replicated in other ethnic groups. In addition, we also found one region associated with platelet aggregation that may play a potential role in atherothrombosis.
| The majority of the variation in platelet count and mean platelet volume between individuals is heritable. We performed genome-wide association studies in more than 16,000 African American participants from seven population-based cohorts to identify genetic variants that correlate with variation in platelet count and mean platelet volume. We observed statistically significant evidence (p-value<5×10−8) that 10 genomic regions were associated with platelet count and 3 were associated with mean platelet volume. Of the regions that were significantly associated, we found 5 novel regions that were not reported previously in other populations. Three of these 5 regions were also associated with platelet count in European Americans and Hispanic Americans. All these regions contain genes that are either known to have or potentially may have a role in determining platelet count and/or mean platelet volume. We further found that one of these regions was also associated with agonist-induced platelet aggregation. Further studies will determine the exact role played by these genomic regions in platelet biology. The knowledge generated by this and other studies will not only help us better understand platelet biology but can also lead us to the discovery of new anti-platelet drugs.
| While platelets play a fundamental role in hemostasis, they are also important in the development of atherosclerosis and arterial thrombosis [1]. An elevated platelet count has been associated with adverse clinical outcomes after thrombolysis or coronary intervention in patients presenting with acute myocardial infarction and moderate reductions in platelet count by thrombopoietin inhibition were associated with reduced thrombogenesis in a primate model [2]–[4]. The heritability of variation in platelet count is substantial with estimates ranging from 54% to more than 80% [5]–[8]. In the GeneSTAR study, a cohort included in the current meta-analysis, the heritability of platelet count is 67% [9].
Like platelet count, an elevated mean platelet volume (MPV) is also associated with adverse cardiovascular events and its reported heritability is as high as 73% [8], [10]–[12]. The heritability of MPV in the GeneSTAR cohort was 71% [9]. Recent genome-wide association studies (GWAS) and meta-analyses have identified genetic variants associated with these two platelet traits in Caucasians and a Japanese population [13]–[15]. A recent meta-analysis in the CARe Project, involving genotyping of about 50,000 single nucleotide polymorphisms (SNPs) in 2,100 candidate genes, also reported two genetic variants associated with platelet count in African Americans [16]. The genetic variants reported to date explain only a small fraction of the heritability in platelet count and MPV, providing an opportunity for new studies to discover additional genetic variants of importance [15]. Moreover, African Americans have higher platelet counts than Caucasians and additional genetic variants may contribute to this difference [17]. Because of the different allele frequencies and linkage disequilibrium patterns in populations of European and African ancestry, we anticipated that we might discover new genetic loci associated with platelet count and MPV in an African American population compared to Caucasians [18].
We performed a meta-analysis of 7 GWAS studies that included African-American subjects in the Continental Origins and Genetic Epidemiology Network (COGENT) in order to identify novel genetic variants associated with platelet count and MPV.
We performed a GWAS analysis of platelet count in an African American discovery sample of 16,388 individuals from 7 population-based cohorts (Table 1). The MPV meta-analysis included all subjects from three cohorts and a subset of subjects from two other cohorts (n = 4,531). Following stringent genotyping and imputation quality control procedures (as outlined in the Methods section), over 2.2 million SNPs were available for analysis in each cohort (Table 1). The results of association studies and the genomic-control corrected QQ plot for the combined African-African GWAS analysis for platelet count and MPV are shown in Figure 1 and Figure 2 and study specific QQ plots and genomic inflation factors are reported in Figures S3 and S4 and Table S1. The Jackson Heart Study (JHS) cohort contains a few hundred related individuals. This resulted in a high genomic inflation factor for platelet count and a few other traits, as previously described in Lettre et al [19]. Within the CARe Consortium, Lettre et al have done several analyses involving simulated phenotypes as well as empirical data (lipids, BMI) and have shown that for JHS, genomic control-correction is an appropriate way to control for the small sub-group of related individuals. A list of all genome-wide significant SNPs with regional plots for platelet count and MPV can be found in Tables S2 and S3 and Figure S1. Cohort-specific QQ-plots and association results for index SNPs associated with platelet count or MPV are summarized in Figure S2 and Table S4. Of the 10 loci on 7 chromosomes that reached GWAS threshold (p<5×10−8) in the platelet count meta-analysis, five have not been reported in previous platelet count GWAS studies in any population and 8 loci have not been reported previously in African Americans (Figure 1). The MPV meta-analysis identified three loci, each one on different chromosomes; two of these loci were also associated with platelet count at GWAS threshold in the current study (Figure 2). One MPV-associated locus has been reported in African Americans before, and two of these three loci have been associated with MPV in Caucasians in prior studies [15], [16]. A sex-specific meta-analysis did not reveal any heterogeneity for the allelic effect between the two sexes and did not uncover any additional loci. Thus, the sex-specific results are not reported here.
The first of the novel loci from platelet count meta-analysis is located on chromosome 6p22. The best SNP (rs12526480; p = 9.1×10−9) in this region is located in the intron of the leucine-rich repeat containing 16A gene (LRRC16A). The minor allele (G) of rs12526480 was associated with decreased platelet count. Ten additional SNPs in the region had p<10−6 (Table 2 and Table S2). The LRRC16A gene encodes a protein called ‘capping protein ARP2/3 and myosin-I linker’ (CARMIL), which plays an important role in cell-shape change and motility. Genetic variants in LRRC16A have been previously reported to be associated with serum uric acid levels [20], nephrolithiasis [21] and markers of iron status [22] but there have been no reports of any association with either platelet count or other platelet phenotypes. In the three European American cohorts, rs12526480 was statistically significant in one cohort (p = 0.01) and near nominal significance in the combined meta-analysis (p = 0.06) with an effect size and direction similar to that observed in African Americans. In Hispanic Americans, rs12526480 was not significantly associated with platelet count (Table 3). Given the proximity of the LRRC16A gene to the hemochromatosis (HFE) gene and the well-known reciprocal relationship between platelet count and iron stores, we additionally assessed the association between rs12526480 and red cell phenotypes in the COGENT African Americans. There was no evidence of association between LRRC16A genotype and hemoglobin, hematocrit, red cell count or mean corpuscular volume in the 16,388 African Americans, nor was there any evidence of association between rs12526480 genotype and serum ferritin in 672 African Americans from CARDIA or 2,126 from JHS. Nor did adjustment for red cell phenotype or iron status alter the relationship between platelet count and rs12526480 genotype. Finally, we had uric acid levels available in 943 African Americans from CARDIA; again there was no association with LRRC16A genotype (Table S5).
The second locus is on chromosome 7q11 where two SNPs in intronic regions of the CD36 gene (rs13236689; p = 2.8×10−9 and rs17154155; p = 1.1×10−8) reached GWAS significance threshold, while 8 additional SNPs had p<10−6. rs13236689 and rs17154155 are in close linkage disequilibrium (r2 = 0.90 in the HapMap Yoruban population). After conditioning on rs13236689 in the association analysis, rs17154155 did not remain statistically significant (p = 0.39). Of the three European American cohorts, rs13236689 was statistically significant in the WHI cohort (p = 0.05) but not in the meta-analysis of all three studies (p = 0.07, Table 3). The CD36 gene encodes a thrombospondin receptor (platelet glycoprotein IV) which is present on the surface of platelets and several other cells [23]. rs17154155 has been reported to be associated with platelet function as well as with platelet expression of CD36 [24], [25].
In the third locus on chromosome 10q21, 71 SNPs reached GWAS threshold and 57 additional SNPs had p<10−6. Two non-synonymous common variants of unknown functional significance, rs 10761725 (resulting in serine to threonine substitution) and rs1935 (resulting in glutamate to aspartate substitution), in this region also crossed the GWAS threshold. All 128 SNPs in this region appear to be in strong linkage disequilibrium based on Yoruban HapMap data. The most significant SNP in this region, rs7896518 (p = 2.3×10−12), is located in an intron of the jumonji domain containing 1C (JMJD1C) gene. SNPs in this region have been reported to be associated with MPV (rs2393967) and with native platelet aggregation in platelet-rich plasma (rs10761741 in Caucasians and rs2893923 in African Americans) but not with platelet count [15], [26]. For rs7896518, data were available from 2 European American cohorts and meta-analysis found a significant association reaching GWAS threshold (p = 2.61×10−9) with similar direction of effect size (Table 3).
The fourth novel locus was located on chromosome 11q13. The most significant SNP (rs477895; p = 4.9×10−8) was in an intron of the BCL2-associated agonist of cell death (BAD) gene, while 23 other SNPs had p<10−6. For rs477895, all replication cohorts had effect sizes in a direction similar to African Americans and one European American and the Hispanic cohorts reached statistical significance (p = 4.48×10−3 and p = 0.04 respectively). Meta-analysis of the three European American cohorts also found significant association of rs477895 with platelet count (P = 1.71×10−3, Table 3). The protein encoded by the BAD gene inhibits the activity of the BCL-xL and BCL-2 proteins and thus has a pro-apoptotic effect [27]. Phospholipase C β3 protein encoded by another gene at this locus, PLCB3, is also known to be present in platelets and its deficiency results in impaired platelet function in mice [28]. This locus also contains SLC22A11 and SLC22A12, two genes that encode solute carrier proteins and previous GWAS have found association of genetic variants in these genes with serum uric acid levels [20]. Of the two genes, the transcript of SLC22A11 is present in significant amount in platelets as is the transcript for BAD [29]. Interestingly, a SNP about 20 kbp upstream of SLC22A11, rs4930420, almost reached GWAS threshold (p = 9.16×10−8, r2 with rs477895 = 0.21) and four additional SNPs in complete LD with rs4930420 (r2 = 1) had p-values<10−6. By examining the actual linkage disequilibrium patterns in this region in COGENT, and by performing conditional regression analysis in more than 8,400 African Americans from the WHI cohort simultaneously adjusting for BAD rs477895 and SLC22A11 rs4930420, we demonstrate that there are likely at least 2 independent platelet count association signals in this region and that the BAD and PLCB3 polymorphisms appear to represent the same association signal (Table S6).
The fifth novel locus was on chromosome 20q13 where one SNP in the SLMO2 gene exceeded GWAS significance threshold (rs151361; p = 9.4×10−9) while 2 other SNPs had p<10−6. One of these two SNPs was located in the first intron of TUBB1 gene (rs6070696; p = 2.5×10−7) and was 16.3 kbp downstream of the lead SNP (YRI HapMap r2 = 0.6). The TUBB1 gene encodes a beta1 tubulin, which plays an important role in megakaryopoiesis [30]. All replication cohorts had effect sizes in the direction similar to African Americans for rs151361 but only one European American study reached statistical significance (p = 0.01). The meta-analysis of the three European American replication cohorts also found a statistically significant association between rs151361 and platelet count (p = 1.1×10−3, Table 3).
In addition to identifying novel loci, we also replicated 5 previously reported loci at GWAS significance threshold and 3 other loci that were highly significant in our study but not at GWAS significance level (Table S7). The strongest signal in our platelet count meta-analysis was from chromosome 6p21 (SNP with the lowest p-value = rs210134; p = 2.3×10−15) located in the BAK1 gene, a locus that has been reported previously in Caucasians, Japanese, and African American populations [13]–[16]. We also found strong associations between platelet count and loci on chromosomes 6q23 (rs9494145; p = 2.8×10−9), 7q22 (rs342293; p = 1.6×10−8), and 12q24 (rs6490294; p = 4.8×10−9), all of which have been previously reported for Caucasians but not for African Americans [15]. Finally, we confirmed the association of a genetic variant rs8109288 (p = 5.0×10−10) in the tropomyosin 4 (TPM4) gene at chromosome 19p13 that has been previously reported for African Americans in a candidate gene study [16]. In our replication cohorts, rs8109288 was associated with platelet count in meta-analysis of European American cohorts and in Hispanic Americans (p = 2.6×10−8 and 0.02 respectively). We were also able to confirm the association of all previously reported SNPs (or a nearby SNP in the same LD block) with platelet count at a p<0.05 (Table S5).
Of the three loci we identified at GWAS significance level for MPV, 2 have been previously reported to be associated with MPV in Caucasians, and one has been reported previously in African Americans. The association which has been previously reported in African Americans was of the A-allele of rs8109288 in TPM4 with increased MPV (p = 3.3×10−9); the same SNP was also associated with platelet count in this study. TPM4, a protein with a major role in stabilizing the cellular cytoskeleton, is present in platelets [31]. In the 7q22 region, we found that the SNP with the lowest p-value for MPV (rs342296; p = 1.4×10−11) was different from the SNP most associated with platelet count (rs342293; p-value = 5.84×10−11) although the two SNPs were only 684 bp apart and are in the same LD block (r2 = 0.92 based on HapMap II YRI) [15]. We also replicated a locus associated with MPV on 17q11 (rs11653144; p = 4.2×10−8) at GWAS significance threshold [15]. Of the 10 additional previously reported loci for MPV, we found statistically significant associations with 7 of them although these associations did not reach GWAS significance threshold (Table S8). For the loci that we were unable to replicate, we found other nearby SNPs with p<0.05. The direction of effect for all SNPs was not similar to the previously reported study of individuals of European ancestry suggesting that the alleles at the causal loci may be different between the two populations.
Three regions (7q11, 7q22, 10q21) containing four SNPs (rs13236689, rs342296, rs342293, rs7896518) have already been shown to be associated with platelet aggregation [24]–[26], [32]. Therefore, the SNPs with the lowest p-values in each of the remaining 8 regions (Table 4) identified for either platelet count or MPV were examined for their association with platelet aggregation in 832 African-American individuals from the GeneSTAR study. Of the 8 SNPs, 3 were associated with a significant change in agonist-induced platelet aggregation but only one exceeded the Bonferroni-corrected significance threshold of 0.005 (Table 4). The minor allele (C) of rs6490294 in the ACAD10 gene (12q24) was associated with increased ADP-induced platelet aggregation (p = 0.002). Variants in this region have been previously reported to be associated with coronary artery disease [15]. The minor allele (A) of the 2nd SNP, rs8109288, in the TPM4 gene, was associated with decreased arachidonic-induced platelet aggregation (p = 0.03) and a trend towards decreased aggregation with ADP (p = 0.09). The minor allele (G) of the 3rd SNP, rs151361, in the SLMO2 gene, was associated with increased ADP-induced platelet aggregation (p = 0.008). The last 2 SNPs were nominally significant but did not exceed the Bonferroni-corrected significance threshold.
We report the first meta-analysis of GWA studies of platelet count and MPV in a large number of African American participants from 7 population-based cohorts. We have identified 5 novel loci associated with platelet count of which three were replicated in the European American cohorts and one in the Hispanic cohort. None of these new African-American platelet loci have been reported previously in any racial group. In addition, we have confirmed that several loci previously reported in Europeans or Japanese are also associated with these platelet phenotypes in African Americans. We have further shown that 3 of the 8 loci (with one exceeding Bonferroni-corrected threshold), for which there have been no previously known association with platelet aggregation, are also associated with differences in platelet function using a subset of our African American sample.
Interestingly, the 5 novel platelet count loci are intragenic and 4 of these genes are known to have some role in platelet formation or biology. Platelets are small anucleate blood cells that are released from the cytoplasm of much larger bone marrow precursor cells known as megakaryocytes. One of the novel findings is the association of LRRC16A gene with platelet count. The protein encoded by the LRRC16A gene, capping protein ARP2/3 and myosin-I linker (CARMIL), plays an important role in actin-based cellular processes. Actin filaments are essential for end-amplification of pro-platelet processes during megakaryocyte maturation [33]. CARMIL exposes the barbed ends of actin filaments by binding to and then dislodging the capping protein from the actin filament [34]. Capping proteins are up-regulated during megakaryocyte maturation and LRRC16A is differentially expressed in megakaryocytes compared to other blood cells [35], [36]. The capping protein binding region of the CARMIL protein resides in the later part of the protein (940–1121 amino acid residues), which is a highly conserved region from protozoa to vertebrates. The majority of the residues in this region are critical for the anti-capping protein activity of CARMIL [37]. The rs12526480 genetic variant identified in our study is located in the latter part of the gene and may be in LD with a functional mutation in this conserved region. Any mutation that decreases the ability of CARMIL to dislodge capping protein from the barbed ends of the actin filament may result in abnormal megakaryocyte maturation and decreased platelet formation which is consistent with the direction of effect we observed in our study.
Another novel finding not reported in earlier GWA studies is the association of platelet count with CD36, a gene that encodes a receptor present on the surface of platelets, megakaryocytes, and several other cells. CD36 has a wide variety of ligands including thrombospondin [23]. Both CD36 and thrombospondin genes are up-regulated during megakaryocyte maturation and binding of thrombospondin-I to CD36 inhibits megakaryopoiesis, thus potentially providing a feedback mechanism for control of megakaryopoiesis [34], [36], [38]. The exact mechanism through which activation of CD36 inhibits megakaryopoiesis is unclear but may involve activation of extrinsic apoptotic mechanisms [39].
The most significant SNP associated with platelet count (rs210134 in BAK1) in our study is in complete LD with the most significant BAK1 SNP reported to be associated with platelet count in individuals of European ancestry (rs210135, r2 = 1 with rs210134 in HapMap II YRI, p = 2.18×10−14 in the current study). While the magnitude of effect is similar, the direction of effect is opposite suggesting that the allele at the causal locus is different in the two ethnic groups. A candidate gene study in African Americans has reported another SNP (rs449242, r2 = 0.81 with rs210134 in HapMap II YRI) in BAK1 and the direction of effect is similar to our study (Table S5) [16]. In addition to confirming the association of genetic variants in the pro-apoptotic BAK1 gene with low platelet count, we have identified and replicated a variant in another pro-apoptotic gene, BAD, that is associated with low platelet count. The protein encoded by BAD acts as a sensor for apoptotic signals upstream of BAK and activates BAK through indirect mechanisms [27]. The identification of these two genes in the intrinsic apoptotic pathway highlights the importance of the apoptotic process in modulating platelet lifespan in the circulation, which is one of the mechanisms that regulate platelet count [40]. Interestingly, this region also contains genetic variants associated with serum uric acid levels [20], however, the mechanism through which uric acid levels may be associated with platelet count remains unclear.
Genetic variants in the JMJD1C gene have been previously reported to be associated with MPV in Caucasians but not with platelet count. Conversely, we found several SNPs in this region that reached GWAS significance threshold for association with platelet count but none with MPV and we replicated the lead SNP in European Americans at GWAS threshold. In a GWAS study of platelet aggregation in Caucasians, the minor allele (T) of rs10761741 was associated with an increase in epinephrine-induced platelet aggregation in Caucasians [26]. JMJD1C gene is a histone demethylase and appears to be involved in steriodogenesis [41]. In addition to its association with platelet aggregation and MPV, previous GWAS have found genetic variants in this gene to be associated with serum levels of alkaline phosphatase and lipoprotein particle size and content [42]–[44].
In addition to confirming the finding of association of A-allele of rs8109288 in TPM4 gene with lower platelet count [16] and replicating this finding in European Americans, we also confirmed the association of the A-allele of this SNP with increased MPV and found a nominally significant association with decreased platelet aggregation. TPM4 gene expression is higher in megakaryocytes than other blood cells or other hematopoietic cells [35], [45]. Tropomyosin proteins play a central role in actin-based cytoskeletal changes and there appears to be biological plausibility for an effect of genetic variants on megakaryocyte maturation and platelet aggregation [46].
The final novel locus in the SLMO2 gene was also replicated in European Americans but SLMO2 gene has no known role in megakaryocyte biology. However, the variant is located within 13 kb of the TUBB1 gene, which is essential in the formation of normal mature platelets. The TUBB1 gene encodes beta1-tubulin that is exclusively expressed in platelets and megakaryocytes and forms a component of microtubules [30]. Loss of function mutations in TUBB1 gene have been reported in the literature and result in thrombocytopenia, large platelets, and increased risk of intracranial hemorrhage in men [47], [48]. The G-allele of the rs1513691 variant is associated with increased platelet count, decreased MPV, and increased aggregation, which may point towards a gain in function mutation in this region.
All previously reported loci that were also significantly associated with platelet count or MPV at GWAS threshold in our study have known biological roles in platelet biology. Two of these regions, 6q23 and 12q24, have pleiotropic effects with the 6q23 region associated with several hematological traits [13], [15], [49] and the 12q24 region associated with celiac disease and coronary artery disease [15]. More importantly, we also found that the 12q24 locus was associated with platelet aggregation after Bonferroni adjustment for multiple comparisons and thus may provide a mechanistic explanation of its role in development of coronary artery disease. The GG genotype of the most significant SNP in the 7q22 region, rs342293, is known to be associated with higher PIK3CG mRNA levels in platelets [32]. SNPs at this locus are also associated with platelet aggregation, pulse pressure, and carotid artery plaque [26], [50], [51]. TAOK1 is an important regulator of the mitotic progression and may also play a role in the apoptosis of cells [52], [53].
Our study included over 16,000 participants with platelet count and over 4500 participants with MPV measured and we were able to identify loci that explain between 0.16–0.33% of the variance in platelet count and loci that explain 1–1.5% of the variance of MPV (Table S9). Overall, the loci we identified explain up to 7% of the variance in platelet count and up to 6% of the variance in MPV, assuming that the each of these loci is independent. However, for both platelet count and MPV, the estimated heritability is >50%. Therefore, for each of these traits, the majority of heritability remains unexplained. One of the limitations of GWA studies is the limited power to detect effects caused by genetic variants with frequency <5%. We hypothesize that a significant proportion of the heritability of platelet count and MPV may be explained by variants with frequency <5%. Alternatively, there may be a large number of additional common variants that affect these traits, but have more modest effects.
In conclusion, we have conducted a meta-analysis of GWAS studies of platelet count and MPV in a large African American population and identified novel genetic variants in regions with genes that are likely to have a role in platelet formation. Furthermore, we have replicated 3 of the 5 novel loci in European Americans and one in Hispanic Americans. The novel regions identified may provide a focus for further research in improving our understanding of the biology of megakaryocyte maturation and platelet survival. In addition, we examined the effect of the genetic variants associated with platelet count and MPV on platelet function, and found 3 of these genetic variants to be associated with agonist-induced platelet aggregation of which one crossed Bonferroni-corrected significance threshold. Whether these newly identified genetic variants contribute to the risk of coronary artery disease or myocardial infarction, or to disorders associated with hyper- or hypo-aggregation of platelets, merits further investigation.
The 7 studies included in this meta-analysis belonged to COGENT and enrolled 16,388 African American participants. The supplementary text contains a detailed description of each participating COGENT study cohort (Text S1). All participants self-reported their racial category. Additional clinical information was collected by self-report and clinical examination. All participants provided written informed consent as approved by local Human Subjects Committees. Study participants who were pregnant or had a diagnosis of cancer or AIDS at the time of blood count were excluded. We also excluded subjects who were outliers in the analysis of genetic ancestry (as determined by cluster analysis performed using principal component analysis or multi-dimensional scaling) or who had an overall SNP missing rate >10%.
Fasting blood samples for complete blood count (CBC) analysis were obtained by venipuncture and collected into tubes containing ethylenediaminetetraacetic acid. Platelet counts and MPV were performed at local laboratories using automated hematology cell counters and standardized quality assurance procedures. Methods used to measure the blood traits analyzed in this study have been described previously for ARIC, CARDIA, JHS, Health ABC, WHI, and GeneSTAR [54]–[58]. Platelet count was reported as 109 cells per liter, and was recorded in all 16,388 study participants. Information on MPV was available in a subset of 4,612 participants from five COGENT study cohorts (ARIC, GeneSTAR, Health ABC, HANDLS, and JHS) and was reported in femto liters (10−15 L). All the phenotypes were approximately normally distributed and we did not perform any data transformations.
Genotyping was performed within each COGENT cohort using methods described in Text S1. Affymetrix chips were used in the ARIC, CARDIA, JHS, and WHI studies and Illumina chips were used in GeneSTAR, HANDLS, and Health ABC. DNA samples with a genome-wide genotyping success rate <95%, duplicate discordance or sex mismatch between genetic estimates of gender and self-report, SNPs with genotyping failure rate >10%, monomorphic SNPs, SNPs with minor allele frequency (MAF) <1%, and SNPs that mapped to several genomic locations were removed from the analyses. Because African-American populations are recently admixed, we did not filter on Hardy-Weinberg equilibrium p-value. Instead, significantly associated SNPs were later examined for strong deviations from Hardy–Weinberg equilibrium and/or raw genotype data was examined for abnormal clustering. Participants and SNPs passing basic quality control were imputed to >2.2 million SNPs based on HapMap II haplotype data using a 1∶1 mixture of Europeans (CEU) and Africans (YRI) as the reference panel. Details of the genotype imputation procedure are described further under Supplemental Methods. Prior to meta-analyses, SNPs were excluded if imputation quality metrics (equivalent to the squared correlation between proximal imputed and genotyped SNPs) were less than 0.50.
Differences in platelet count may affect platelet function and aggregation [59]. In addition, younger platelets have higher MPV than older platelets and are more reactive [60]. We hypothesized that the genetic variants that determine platelet count and MPV may also affect platelet aggregation. To examine this hypothesis, we used agonist-mediated platelet aggregation assays, which can provide information about the different aspects of platelet aggregation. For these assays, platelet aggregation agonists, such as collagen or ADP, are added to whole blood or platelet-rich plasma and platelet aggregation is measured after a specified amount of time (300 seconds). We performed platelet aggregation assays in the participants of the GeneSTAR cohort. Blood samples were obtained as described above, and platelet aggregation in whole blood was measured as reported previously [57]. Briefly, in vitro whole blood impedance in a Chrono-Log dual-channel lumiaggregometer (Havertown, Pa) was performed after samples were stimulated with arachidonic acid (0.5 mmol/L, intra-assay CV = 24%), collagen (5 µg/mL; intra-assay CV = 9%), or ADP (10 µmol/L; intra-assay CV = 46%). Maximal aggregation within 5 minutes of agonist stimulation was recorded in ohms.
For all cohorts, genome-wide association (GWAS) analysis was performed using linear regression adjusted for covariates, implemented in either PLINK v1.07, R v2.10, or MACH2QTL v1.08 [61], [62]. Allelic dosage at each SNP was used as the independent variable, adjusted for primary covariates of age, age-squared, sex, and clinic site (if applicable). The first 10 principal components were also incorporated as covariates in the regression models to adjust for population stratification (Text S1). For GeneSTAR, family structure was accounted for in the association tests using linear mixed effect (LME) models implemented in R [63]. Although the JHS has a small number of related individuals, extensive analyses have shown that results were concordant using linear regression or LME, after genomic control [19]. Therefore, results are presented for JHS using linear regression. For imputed genotypes, we used dosage information (i.e. a value between 0.0–2.0 calculated using the probability of each of the three possible genotypes) in the regression model implemented in PLINK or MACH2QTL (for cohorts with unrelated individuals) or the Maximum Likelihood Estimation (MLE) routines (for GeneSTAR).
For both platelet phenotypes, meta-analyses were conducted using inverse-variance weighted fixed-effect models to combine beta coefficients and standard errors from study level regression results for each SNP to derive a combined p-value and effect estimate [64]. Study level results were corrected for genomic inflation factors (λGC) by incorporating study specific λGC estimates into the scaling of the standard errors (SE) of the regression coefficients by multiplying the SE by the square-root of the genomic inflation factor. The inflation factors for all completed analyses are presented in Table S1. To maintain an overall type 1 error rate of 5%, a threshold of α = 5×10−8 was used to declare genome-wide statistical significance. Between-study heterogeneity of results was assessed by using Cochrane's Q statistic and the I2 inconsistency metric. Meta-analyses were implemented in the software METAL [64] and were performed independently by two analysts to confirm results. To examine whether there were any differences between males and females, sex-specific GWAS were conducted in each cohort. The results for each SNP were pooled and heterogeneity of allelic effects between females and males was examined using the meta-analysis methods as implemented in GWAMA software [65].
To assess whether the loci previously reported to be associated with the platelet phenotypes in Europeans, Japanese, and African Americans were replicated in the COGENT African-Americans, we examined the meta-analysis results for each index SNP in the regions previously reported, including consistency of direction of effect. If the reported index SNP was not significant at p<0.05 we examined adjacent SNPs and reported the closest SNP with p<0.05 along with its distance from the index SNP.
To examine the association of genotype on platelet aggregation in the GeneSTAR cohort, linear mixed models were used with additive models adjusting for age and sex, and taking into account familial correlation between the individuals.
|
10.1371/journal.ppat.1002787 | The Cholesterol-Dependent Cytolysin Signature Motif: A Critical Element in the Allosteric Pathway that Couples Membrane Binding to Pore Assembly | The cholesterol-dependent cytolysins (CDCs) constitute a family of pore-forming toxins that contribute to the pathogenesis of a large number of Gram-positive bacterial pathogens.The most highly conserved region in the primary structure of the CDCs is the signature undecapeptide sequence (ECTGLAWEWWR). The CDC pore forming mechanism is highly sensitive to changes in its structure, yet its contribution to the molecular mechanism of the CDCs has remained enigmatic. Using a combination of fluorescence spectroscopic methods we provide evidence that shows the undecapeptide motif of the archetype CDC, perfringolysin O (PFO), is a key structural element in the allosteric coupling of the cholesterol-mediated membrane binding in domain 4 (D4) to distal structural changes in domain 3 (D3) that are required for the formation of the oligomeric pore complex. Loss of the undecapeptide function prevents all measurable D3 structural transitions, the intermolecular interaction of membrane bound monomers and the assembly of the oligomeric pore complex. We further show that this pathway does not exist in intermedilysin (ILY), a CDC that exhibits a divergent undecapeptide and that has evolved to use human CD59 rather than cholesterol as its receptor. These studies show for the first time that the undecapeptide of the cholesterol-binding CDCs forms a critical element of the allosteric pathway that controls the assembly of the pore complex.
| The CDCs are a large family of pathogenesis-associated pore-forming toxins that are expressed by many Gram-positive pathogens. The conserved undecapeptide motif of the CDCs has been regarded as the signature peptide sequence for these toxins, yet its function has remained obscure. The studies herein show that the undecapeptide forms a critical structural element in the allosteric pathway that couples membrane binding to cholesterol to the initiation of distal structural changes, which are required for the assembly of the pore forming complex. These studies provide the first insight into the function of this highly conserved sequence and show that through evolution this pathway is missing in the CD59-binding CDCs.
| The cholesterol-dependent-cytolysin (CDC) family of toxins consists of over 25 members that are produced by many different species of Gram-positive bacterial pathogens [1] and contribute in various ways to the pathogenesis of these organisms [2], [3], [4], [5]. Members of this family exhibit high levels of homology in their primary structures (40–70%) and in the crystal structures of their soluble monomers [6], [7], [8], [9]. The region within the CDC primary structure that exhibits the highest degree of sequence identity is an 11-residue peptide known as the undecapeptide or tryptophan-rich motif, which is located near the C-terminus of the molecule in domain 4 (D4) (Fig. 1). The undecapeptide (ECTGLAWEWWR) is the signature motif for the CDCs [1] and so proteins exhibiting this peptide sequence have a high probability of belonging to the CDC family. The pore forming mechanism of the CDCs that use cholesterol as their receptoris highly sensitive to changes in the primary structure of the undecapeptide [6], [10], [11], [12], [13], [14], [15]. These studies suggest that the undecapeptide plays an important role in the CDC pore-forming mechanism, yet since Iwamoto et al. [16] began studying the effects of chemically altering the undecapeptide in 1987 its contribution to the pore forming mechanism of the CDCs has remained elusive.
The undecapeptide is located at the tip of D4 of the CDC structure, as shown in the structure of the CDC produced by Clostridium perfringens, perfringolysin O (PFO) (Fig. 1). D4 also contains the cholesterol recognition/binding motif (CRM) and two other short loops (L2 and L3) near the undecapeptide (reviewed in [17]). Upon recognition of membrane cholesterol by the CRM,loops L2 and L3 insert into the membrane. These interactions anchor the monomers in a perpendicular orientationto the membrane surface where the tip of D4 is anchored to the membrane surface and the top of D3 resides about 113 Å above the membrane surface [18], [19], [20]. Although the sidechains of several residues of loops L2 and L3 and the undecapeptide insert into and anchor the monomers to the membrane they do not penetrate deeply into the bilayer core [19], [21].
It had been generally accepted in the field that the undecapeptide motif wasthe CRM of the CDCs, although this function had never been demonstrated unambiguously. An early study by Iwamoto et al. [16] showed that chemical modification of the undecapeptide cysteine caused independent defects in both binding and pore formation. Since that time it has been shown that mutation of many of the undecapeptide residues often affects both binding and pore formation [6], [10], [11], [12], [13], [14], [15]. We recently showed, however, that the CRMresides in the nearby D4 loop L1 (Fig. 1) and is comprised of a threonine-leucine pair that is strictly conserved in all known CDCs [22]. Upon cholesterol binding by the CRM the nearby loops L2 and L3 and the conserved undecapeptide insert into the bilayer surface and anchor the monomer in a perpendicular orientation to the membrane surface [19], [21], [23], [24]. Membrane binding in conjunction with monomer-monomer interactions [20] initiates and drives a dramatic series of secondary and tertiary structural changes in D3, which is about 60 Å distant from the tip of D4 (Fig. 1). These structural changes are necessary for the assembly of the membrane bound monomers into the large oligomeric pore complex [23], [24], [25], [26], [27]. Soluble monomers of PFO do not exhibit these D3 structural changes, even at the high concentrations required for crystallization of the protein [28]: membrane binding is required to initiate the structural changes in D3 [20], [25].
As indicated above, the pore-forming mechanism of PFO-like CDCs is highly sensitive to mutations in the undecapeptide [6], [10], [11], [12], [13], [14], [15]. Furthermore, the conformational changes in the PFO undecapeptide, reflected by the membrane insertion of its tryptophan residues, are conformationally coupled to the structural changes in TMH1 required for the formation of the β-barrel pore [20]. This observation suggests that the undecapeptide of PFO is involved in the allosteric coupling of membrane binding to the initiation of the D3 structural changes that are necessary for monomer-monomer interaction and the formation of the oligomeric β-barrel pore complex.
A small family of CDCs, typified by Streptococcus intermedius intermedilysin (ILY) use human CD59 as their receptor, rather than cholesterol [29], [30], [31]. The D3 structural changes in ILY can be initiated by binding to human CD59 in membranes that are largely, though not completely depleted of cholesterol [32]. ILY still requires a CRM-mediated membrane interaction with cholesterol to maintain its anchor to the membrane surface (it disengages from CD59 during prepore to pore conversion [22], [33]), but it remains unclear if cholesterol binding also participates in initiation of the D3 structural changes necessary for assembly of the oligomer pore complex. Interestingly, in contrast to the CDCs that use cholesterol as their receptor, the pore forming mechanism of ILY is comparatively insensitive to mutations within its undecapeptide [6], which suggests that it may not play as significant of a role in the pore forming mechanism of these toxins.
In the present study we performed a detailed molecular analysis of a point mutation in the undecapeptide of PFO that reduces its pore-forming activity 100-fold, whereasthe analogous mutation has no significant effect on the mechanism of ILY [6]. In PFO this mutant blocks all measurable structural transitions in D3 and prevents the stable interaction of membrane-bound monomers. We further show that the effect of this mutation on the activity of PFO is similar to that observed for cholesterol bound native ILY in the absence of CD59. These results show that the undecapeptide of PFO is a critical structure within the allosteric pathway of PFO that couples cholesterol binding to the initiation of structural changes within D3, which lead to the formation of the β-barrel pore. We further show that this pathway appears to be missing in the CD59-binding ILY, so that assembly of its pore complex is initiated by its interaction with CD59 rather than cholesterol.
Arg-468 is the last residue of the PFO undecapeptide (ECTGLAWEWWR), as well as in the ILY undecapeptide (GATGLAWEPWR). Substitution of the PFO undecapeptide at this residue with alanine decreases its hemolytic activity 100-fold (Table 1), whereas substitution of the analogous residue in ILY has little effect on the activity [6]. A series of mutants were generated for Arg-468 of PFO to examine the effects of size, length and charge of the residue atposition 468 on the hemolytic activity of PFO (Table 1). Neither conservative nor non-conservative substitutions were tolerated: all substitutions decreased hemolytic activity ≥100-fold. Based on the crystal structure of PFO the only intramolecular contacts established by Arg-468 are hydrogen bonds between its sidechain NH1 and the CRM carbonyls (Fig. 1B). Interestingly, this contact is lost in the ILY monomer (Fig. 1C), which presumably results from differences in its undecapeptide structure [6]. We selected the PFOR468A mutant for further studies into the defect(s) induced by substitution of the Arg-468 residue on the PFO pore-forming mechanism.
We have shown that a conserved Thr-Leu pair in Loop 1, and not the undecapeptide, is responsible for CDC binding to membrane cholesterol [22], yet mutations within the conserved undecapeptide were often observed to affect binding [6], [13], [34]. To confirm that the loss of hemolytic activity by the PFOR468A mutant was not due solely to a defect in binding we examined the ability of the mutant to bind to human RBCs by flow cytometry. In order to prevent cell lysis at high concentrations of toxin, derivatives of native PFO and PFOR468A were generatedin which an engineered disulfide was introduced between residuesThr-319 in β4 and Val-334 in β5 that prevent the rotation of β5 away from β4 in domain 3 (PFOβ4β5 and PFOR468A•β4β5). The engineered disulfide thereforepreventsthe formation of a functional pore by blocking the intermolecular interaction of β1 of one monomer with β4 of another monomer [25]. A third cysteine was substituted at residue Asp-30 in both mutants, which is at the amino terminus of PFO, so that specific fluorescent probes could be introduced into these mutants. This mutation does not affectthe structure of PFO or its function [35].
At the highest concentration of PFOR468Awe observed about a 50% decrease in binding to hRBCs compared to PFO, although at lower concentrations this difference was greater (Fig. 2A). The decrease in binding, however, doesn't account for the 100-fold decrease in cytolytic activity. PFO follows an ordered series of coupled conformational changes that are initiated by binding [19], [20], [23], [25], therefore the major defect induced by the PFOR468A mutation affects an event after binding, which then prevents formation of the pore complex.
Upon membrane binding PFO monomers oligomerize into large SDS-resistant prepore complexes containing approximately 36 monomers [23]. An oligomerization assay was performed with the PFOR468A mutant. Due to the lower binding affinity observed in Fig. 2A we increased the concentration of human red blood cells (hRBCs) to ensure complete binding of PFOR468A. The concentration of hRBCs in the binding studies in Fig. 2A was maintained at 4×106/ml whereas in the oligomerization assay shown in Fig. 2B the concentration hRBCs ranged from 2.5×107 to 2.5×108/ml, which wasapproximately 6–60 fold higher than the concentration used in the flow cytometry assay. After the toxins were allowed to bind, the samples were solubilized with SDS without heating and separated by SDS-agarose gel electrophoresis (SDS-AGE), which separates the monomer and oligomer forms [36]. Native PFO readily formed SDS-resistant oligomers at all concentrations of RBCs (Fig. 2B) whereas PFOR468Adid not form detectable levels of SDS-resistant oligomers (Fig. 2B). Therefore, the major defect in the PFOR468A pore-forming mechanism follows binding and prevents the formation of an SDS-stable oligomer.
Several structural transitions in domain 3 are initiated by membrane binding, which are required for oligomerization and pore formation [25], [26], [27]. One of these structural transitions is the rotation of β-strand 5 (β5) away from the adjacent β-strand 4 (β4) of the core β-sheet in domain 3 (Fig. 1A), whichcontributes to the formation of the SDS-resistant prepore oligomer [25]. Rotation of β5 away from β4 allows the formation of edge-on hydrogen bonds between the peptide backbones of β4 and β1 of two membrane-bound monomers [25].
The disruption of the β4/β5 interaction can be followed spectroscopically using the environmentally sensitive fluorescent probe, NBD (7-nitrobenz-2-oza-1,3 diazole) [27], which is positioned on the sulfhydryl of a cysteine substituted for Val-322 in β4 (Fig. 1A). Val-322 is buried under the residues of β5 and so a probe positioned here is in a hydrophobic pocket. The fluorescence emission of NBD is quenched by water, therefore as β5 rotates away from β4 the NBD positioned in β4 moves from a nonpolar to polar environment, which results in a decrease in its fluorescence emission intensity as it is exposed to the aqueous milieu [25]. The PFOR468A•V322C-NBD mutant exhibited virtually no change in the NBD emission compared to functional PFOV322C-NBD (Fig. 3). These results show that the rotation of β5 away from β4 does not occur in membrane bound PFOR468A.
The studies above show that PFOR468Adoes not form SDS-resistant oligomers, which is likely due to the loss of the intermolecular β1–β4 interaction of monomers. This observation, however, did not rule out the possibility that PFOR468Amonomers could still form a SDS-sensitive oligomer. To determine whether PFOR468Aformed SDS-sensitive oligomers the PFOR468A monomer association was examined using fluorescence resonance energy transfer (FRET). A cysteine was substituted for the amino terminal Asp-30 and labeled with either donor fluorophore (D) (Alexa Fluor 488) or acceptor fluorophore (A) (Alexa Fluor 568). A mixture containing a 4∶1 molar ratio of A-labeled PFOR468A or unlabeled PFOR468A(U) to D-labeled toxin was incubated with membranes and fluorescence emission intensity of D was measured.
When membrane-bound PFO monomers associate to form the prepore oligomer the distance (R0) between D and Afluorescent dyes on the monomers decreases, which results in the FRET-dependent quenching of the D emission (R0 is typically<10 nm) [37]. As expected, we observed an A-dependent quenching of the D emission for functional PFO as it oligomerized [35], [37], whereas no change in the donor fluorescence was observed for PFOR468A(Fig. 4).This result shows that the PFOR468A monomers do not interact, or only form transient interactions that cannot be detected by FRET. FRET requires the donor and acceptor pair be at a fixed distance during the lifetime of the donor emission, which for Alexa-488 is approximately 4 ns [38]. Therefore the PFOR468A monomers are, at most, only interacting briefly within a timeframe that is shorter than the fluorescence lifetime of the Alexa dye.
The D2–D3 interface is disrupted in order to extend the D3 α-helical bundle into transmembrane β-hairpin 1 (TMH1) [27], which together with TMH2 ultimately contribute to the formation of the membrane spanning β-barrel pore [26], [27]. First, the α-helical bundle that forms TMH1 must break its interaction with D2 to unravel and form the extended β-hairpin structure, which eventually inserts into the bilayer as part of the β-barrel pore [27]. Disruption of the TMH1 contact with D2 can be measured by placing a NBD probe on a cysteine substituted for Asn-197inTMH1 (Fig. 1). Asn-197 resides at this interface and undergoes a nonpolar to polar transition as the α-helical bundle breaks contact with D2 and unravels to form the extended β-hairpin [27]. The subsequent insertion of the β-barrel pore can be followed by placing a NBD probe on cysteine-substituted Ala-215 in TMH1, which undergoes a polar to nonpolar transition as its sidechain inserts into the bilayer core [27].
As expected, the fluorescence emission of the NBD probe on cysteine substituted Asn-197 in native PFO decreases to less than 25% of its initial value as the α-helical bundle disengages from its interface with D2 (Fig. 5A, left panel). Also, as expected, the fluorescence emission of the NBD probe located at position 215 in TMH1 of PFO increases as it makes the transition from its polar environment in the soluble monomer to its membrane embedded position in the β-barrel pore (Fig. 6A, left panel). In contrast, little change was detected in the fluorescence emission of the NBD probe at both locations in membrane bound PFOR468A•N197C-NBD, showing that TMH1 did not disengage from its interface with D2 (Fig. 5A, right panel) and insert into the membrane (Fig. 6A, right panel).
As shown above, the membrane-bound monomers of PFOR468A do not interact and the D3 structural transitions that lead to the insertion of the β-barrel pore do not occur in PFOR468A: in essence the monomers remain inert after binding. Therefore, we next determined whether functional PFO could form chimeric oligomers with PFOR468Aand drive these structural transitions.
The same experiments were performed as in Figs. 5A and 6A except that a 4∶1 ratio of unlabeled PFO or PFOR468A was mixed with the labeled species prior to their addition to the liposomes. As expected, the relative emission intensity of the NBD probe was similar when each fluorescence species was mixed with a 4-molar excess of the unlabeled homologous protein. For PFON197C-NBD compare the left panels of Fig. 5A and Fig. 5B and for PFOA215C-NBD compare the left panels of Fig. 6A and Fig. 6B. Similarly, no change was observed in the NBD emission for PFOR468A•N197C-NBD (compare Fig. 5A, right panel to Fig. 5B center panel) and for PFOR468A•A215C-NBD (compare Fig. 6A, right panel to Fig. 6B center panel) when they were mixed with a 4-molar excess of unlabeled PFOR468A.
However, when a 4-molar excess of unlabeled PFO was mixed with the NBD-labeled species of PFOR468A it drove the disruption of the D2–D3 interface and insertion of the β-barrel pore. We observed the expected decrease in the fluorescence emission of the NBD probe located at the D2–D3 interface (compare the right and left panel of Fig. 5), as β5 swings away form β4. Also, the relative emission intensity increased as the probe located at position 215 inserted into the bilayer (compare the right and left panels in Fig. 6). Furthermore, the change in the emission intensity of the NBDin PFOR468A•A215C-NBD when mixed with a 4-molar excess of PFO was quantitatively similar to that observed for PFOA215C-NBD alone or mixed with the unlabeled PFO. Therefore, nearly all of the PFOR468A•A215C-NBD TMHs were converted to a membrane-inserted state.
These results show that functional PFO can form sufficient intermolecular contacts with PFOR468A to efficiently drive the disruption of its domain 2–3 interface and the membrane insertion of its β-barrel. Hence, PFOR468Ais competent to undergo the necessary D3 structural changes and insert its TMHs into the membrane, but is unable to initiate these changes because it is missing the allosteric signal that is initiated by membrane binding. These data also indicate that the rate of binding of the PFOR468A monomers to the membrane surface is not significantly different from that of the native PFO monomers, otherwise the PFO monomers would preferentially interact with each other before interacting with PFOR468A, which would have resulted in a less efficient conversion of the PFOR468A monomers to an inserted state.
Our previous studies suggested that cholesterol binding by ILY was not necessary to trigger the D3 structural changes that are necessary for the formation of the oligomeric complex [32]: it appeared that CD59 binding, not cholesterol binding, initiated the D3 structural changes. Subsequent studies showed that the ILY CRM must initiate a cholesterol-dependent interaction to trigger the membrane insertion of loops L1–L3, which is necessary to anchor ILY to the membrane when it disengages from CD59 during prepore to pore conversion [33]. Therefore, if ILY could bind directly to cholesterol, in the absence of CD59, we predict that this interaction alone would not trigger the formation of the pore complex, as control of this process has been transferred to the CD59-binding site [30].
Although ILY does not bind significantly to cholesterol-rich cell membranes that lack human CD59 [29], we unexpectedly discovered that it binds well to cholesterol-rich liposomes, even better than PFO (Fig. 7A). Furthermore, this binding is dependent on the CRM, as a CRM knockout (ILYDM) lacksdetectable binding to liposomes (Fig. 7A). Therefore, does this CRM-mediated binding trigger the D3 structural changes like PFO and formation of a β-barrel pore? To address this question we first generated cholesterol-rich liposomes with entrapped 5(6)-carboxyfluorescein (CF) and then treated them with PFO or ILY. The fluorescence emission of the concentrated liposome-trapped dye is quenched, but if the dye is released from the liposome its fluorescence emission increases upon dilution as it is released from the liposome [39], [40], [41]. PFO exhibited a dose-dependent release of the dye as evidenced by the increased emission of the dye as the concentration of PFO was increased. Although about twice as much ILY as PFO is bound to the liposomes, the ILY released less than 6% of dye released by PFO at the highest concentrations (Fig. 7B). We confirmed that the β-barrel of ILY was not inserting by measuring the insertion of TMH1. A NBD probe was position in TMH1 at cysteine-substituted His-242, which is a membrane facing residue in the β-barrel [42].Consistent with the lack of pore formation, ILY did not insert its β-barrel into the liposomal membranes (Fig. 7C).
Although pores were not forming, it was possible that the D3 β4–β5 interactionwas disrupted upon ILY binding to cholesterol-rich liposomes. The disruption of the β4–β5 interaction is detected by a decrease in the emission intensity of an NBD probe positioned in β4 as β5 rotates away from β4 its exposes the probe to the aqueous milieu thereby quenching its emission. This transition did not occur in the liposome bound ILY (Fig. 8), although it does occur in ILY bound to human CD59 containing cell membranes [32].Collectively these results suggest that while ILY can bind to cholesterol-rich POPC liposomes, like PFOR468A its binding does not trigger the D3 structural transitions necessary to initiate the formation of the oligomeric pore complex.
The PFO pore forming mechanism is highly sensitive to changes in the undecapeptide structure [6], [13], [16], but until now the molecular basis for its role in the CDC mechanism has been elusive. The studies herein show that mutation of the undecapeptide arginine residue uncouples membrane binding from the D3 structural transitions, which are necessary for the assembly of the pore complex.This mutation blocks all detectable structural changes in D3and prevents the stable interaction of the membrane-bound monomers. In essence, the structure of membrane bound monomers of this mutant appears relatively unchanged from that of the soluble monomer. Hence, for the first time these studies demonstrate a function for the conserved undecapeptide, which forms a critical structural elementin the allosteric pathway that couples membrane binding to the D3 structural changes that lead to pore formation. The studies also show that binding initiates changes through this allosteric pathway that allow monomer-monomer interaction, but it is the monomer-monomer interactions that subsequently drive the major D3 structural transitions that are required for formation of the oligomeric pore complex.Furthermore, these studies show that this pathway has been lost in a CD59-binding CDC, which was a necessary evolutionary step towards transferring control of this process from the cholesterol-binding site to the CD59-binding site.
These studies show that mutation of Arg-468 disrupts the allosteric signal that couples binding to the D3 structural changes that lead to the formation of the oligomeric pore complex and that functional derivatives of PFO can drive the major structural changes in D3, which are necessary for membrane insertion of the TMHs of PFOR468A. Hotze et al. showed that monomer-monomer contact could drive the major structural transitions in D3 in a mutant of PFO that was trapped in a prepore complex. Here we show that PFO can drive these changes in PFOR468A, which is trapped at a much earlier stage where the monomers cannot interact with each other. Our data suggest that membrane binding is allosterically coupled to structural changes in PFO, which facilitate monomer-monomer interaction, but alone this allosteric pathway does not drive the major D3 structural transitions (i.e., disengagement of D2–3 interface and the β4–β5 interaction): these changes are driven by the subsequent interaction of monomers.
Soluble monomers of PFO do not interact and form oligomers, even at the high concentrations required for crystallization [7], [28]. Therefore, PFO membrane binding must initiate structural changes in the monomers that facilitate their interaction. The monomer-monomer interactions then drive the major conformational changes within domain 3, whichare required for the formation of the β-barrel pore [35]. In this way PFO ensures an efficient assembly of the oligomeric pore complex on the target membrane. Membrane-bound PFOR468A monomers did not appear to form interactions that were of sufficient duration to be detected by FRET. For FRET to occur the donor and acceptor fluorophores must be at a fixed distance for a time that is equal to or greater than the half-life of the donor fluorescence emission, which is approximately 4 nsec for the Alexa-488 dye [38]. Therefore, the mutation of Arg-468 appears to prevent the changes in the monomer structure that allows monomers to initially interact and form stable contacts that then drive the D3 structural changes. This mutation results in a membrane-bound monomer that appears to retain the structure of the soluble monomer, which cannot form any detectable intermolecular interactions.
In the crystal structure of PFO the only contacts made by Arg-468 are hydrogen bonds between the NH1 of its guanidinium group and the backbone carbonyls of the CRM Thr-Leu pair (Fig. 1). Therefore, its substitution with alanine only prevents the formation of these two hydrogen bonds. This contact is interesting because it hydrogen bonds with the CRM, and may help stabilize the CRM structure in PFO. Hence, this contact may explain why mutation or chemical modification of the undecapeptide affects binding, as well as assembly of the oligomeric pore complex of PFO [6], [13], [16]. We cannot know with certainty, however, that this contact is essential to the allosteric pathway, only that the substitution of Arg-468 disrupts the allosteric pathway that couples membrane binding to the formation of the pore complex. The crystal structures of the cholesterol-binding CDCs PFO [7], suilysin (SLY) [43] and anthrolysin O (ALO) [9] have revealed that the undecapeptide 3D structure is highly variable: no two undecapeptides 3D structures have been shown to be the same [7], [9], [43], even though their undecapeptide primary structures are identical. It is possible that these differences are due to an inherent flexibility of the undecapeptide and/or crystal contacts that affect the structural arrangement of the undecapeptides in the crystals. Hence, the conformational coupling of binding to the D3 structural changes may proceed through different undecapeptide mediated contacts in the CDCs. Alternatively, if the structure of the undecapeptide is flexible, as is suggested by the crystal structures, then membrane binding may lock it into a specific conformation that transmits the allosteric signal to D3, which cannot be achieved in PFOA468A.
Functional PFO can drive D3 conformational changes in PFOR468Aand the membrane insertion of itsTMHs in chimeric oligomers comprised of both proteins. Therefore, PFOR468A is structurally competent to form a pore, but lacks the conformational signal that initiates the necessary changes in its structure that facilitate the formation of stable intermolecular contacts. The fact that native PFO can drive these structural changes in PFOR468A indicates that it can establish a sufficient number of contacts with the PFOR468A monomers to drive these conformational changes in the latter. No stable intermolecular interactions of the PFOR468A monomers alone were detected by FRET showing that they do not interact, or that the interactions are transient and only exist on a timescale that cannot be detected by FRET. The ability of functional PFO derivatives to interact with PFOR468A indicates that at least one of the surfaces of PFOR468Ais accessible to the functional PFO derivatives, which allows the functional PFO derivatives to dock with PFOR468A.This interaction allows the functional PFO derivatives to establish contact with andsubsequently drive the structural changes in PFOR468AD3 that are necessary for the formation of the oligomeric pore complex.
It is clear thatPFOR468Abindingwas also affected by the Arg-468 to alanine mutation. If Arg-468 does make contact with the CRM carbonyls, as suggested by the PFO crystal structure, then it is possible that this substitution partially destabilized the CRM structure thereby affected binding. However, avidity may be a more important factor that contributes tothe difference in binding of wildtype PFO and PFOR468A. Wildtype PFO and all mutants thereof generated to date still form membrane oligomers (most are represented herein). Oligomerizationis an important component of the binding interaction due to the avidity of the oligomeric complex versus the binding affinity of a single monomer. Oligomerization of PFO begins shortlyafter binding [37], [44], thus binding of wildtype PFO and its derivatives reflects the avidity of the oligomeric complex rather than the affinity of single monomer.PFOR468A is the first mutant that has been described, whichis trapped in a monomer state on the membrane.Hence, therelatively poorbinding exhibited by PFOR468A may actually reflect the true binding interaction of native PFO monomers in the absence of oligomerization. It is also important to note that in the experiments in which functional PFO was used to drive the structural transitions in PFOR468A that we obtained near quantitative conversion of these transitions with a 4∶1 molar ratio of functional PFO to PFOR468A. If PFOR468A monomers bound the membrane with a significantly lower affinity than native PFO monomers then the probability of the native PFO monomers interacting with the PFOR468A monomers would be decreased and therefore it would be unlikely that we would have observed the near quantitative conversion of the PFOR468A monomers to a membrane-embedded state.
The CD59-binding CDCs, ILY [29], vaginolysin (VLY) [31] and lectinolysin (LLY) [45], [46] exhibit undecapeptides with significant changes to their primary structures, most notably a proline substitution for the second conserved tryptophan (consensus, ECTGLAWEWWR;ILY, GATGLAWEPWR; VLY, EKTGLVWEPWR; LLY, EKTGLVWEPWR). Unlike PFO, ILY does not maintain the hydrogen bond contacts between Arg-495 and the CRM (Fig. 1). This may be one of contacts in the cholesterol-dependent allosteric pathway that was disrupted during the evolution of the CD59-binding site, which was necessary to transfer of control of the D3 structural changes from the cholesterol-binding site to the CD59-binding site. Consistent with this scenario is the observation that substitution of the analogous arginine residue in ILY has little effect on the ILY pore-forming mechanism [6]. We have shown herein that when ILY binds to cholesterol in the absence of CD59 it remains largely inert on the membrane, similar to what we observed for PFOR468A. These data suggest that through evolution ILY has lost the allosteric pathway that couples cholesterol binding to the D3 structural changes in order to transfer control of the assembly of the oligomeric complex to the CD59-binding site [33], [42], [46].
Recently others have proposed that the membrane attack complex/perforin (MACPF) family of proteins may exhibit a CDC-like pore forming mechanism [47], [48], [49], [50], [51]. This proposal is based on the presence of a conserved protein fold that is similar to D3 of the CDCs [7], which we have shown forms the β-barrel pore structure of the CDCs [26], [27]. The MACPF proteins play important roles in immune defense as well as in the pathogenesis of eukaryotic pathogens such as Toxoplasma gondii [52]and Plasmodium falciparum [53], [54], [55]. These proteins exhibit little sequence homology with the CDCs and do not exhibit an undecapeptide motif. It is possible, however, that they will alsoexhibit an analogous allosteric mechanism to regulate the assembly of their pore complex.
These studies provide the first evidence that shows the conserved undecapeptide plays an integral role in the allosteric coupling of cholesterol-mediated membrane binding to distal structural changes, which are necessary for the monomer-monomer interactions that drive the assembly of the β-barrel pore.
The genes for native ILY and PFO were cloned into pTrcHisA (Invitrogen) as described previously [27], [42]. All mutations were made in native ILY (naturally cysteine-less) or the cysteine-less PFO derivative (PFOC459A) backgrounds. The various CDCs and their derivatives are summarized in Table 2. All chemicals and enzymes were obtained from Sigma, VWR and Research Organics. All fluorescent probes were obtained from Molecular Probes (Invitrogen). Polyclonal anti-PFO antibody was affinity purified from hyperimmune rabbit serum. Secondary antibody goat anti-rabbit-HRP was obtained from BioRad. Sterols were obtained from Steraloids and lipids were obtained from Avanti Polar Lipids.
PCR QuikChange mutagenesis (Stratagene) was used to make the various amino acid substitutions in native ILY or PFOC459A and DNA sequences of the PFO and ILY mutants were determined by the Oklahoma Medical Research Foundation Core DNA Sequencing Facility. The expression and purification of recombinant toxins and derivatives inEscherichia coliBL21 DE3 were carried out as previously described [27], [56]. Purified protein was stored in HBS [100 nM NaCl, 50 mM HEPES; (pH 7.5)], 50 µMtris(2-carboxyethyl)phosphine (TCEP) and 10% (vol/vol) glycerol at −80°C.
The labeling of PFO, PFOR468A and ILY cysteine-containing derivatives with IANBD [iodoacetamido-N,N′-dimethyl-N-)7-nitrobenz-2-oxa-1,3-diazolyl)ethylene-diamine; Molecular Probes] was carried out as previously described [27], [42]. Toxin derivatives were labeled using a 20-fold molar excess of the probe overnight at room temperature (22°C).The labeling reactions for the PFOV322C derivatives also contained 3 M guanidine hydrochloride to increase the efficiency of labeling.
Following the modification with the probes the mixtures were passed over a Sephadex G-50 column equilibrated in HBS. The labeled samples were made 10% (vol/vol)in glycerol and stored at −80°C. Proteins were typically labeled at an efficiency of 80–100%.
The cytolytic activity of the toxins and their derivatives on human red blood cells (hRBCs) was measured as previously described [27] except that the procedure was adapted to a microtiter plate format. Briefly, fresh human RBCs (hRBCs) were washed and suspended to 5% in phosphate buffered saline (PBS). The PFO and its derivatives were serially diluted in 2-fold steps in a microtiter plate at a final volume of 50 µl per well to which 50 µl of a 5% suspension of hRBCs was added and incubated for 1 hour at 37°C. After incubation, unlysed RBCs were removed by centrifugation of the plate at 3400×g for 10 min. The EC50for hemolysis (effective concentration of toxin for 50% hemolysis)was determined by quantifying hemoglobin release by measuring the absorbance of the supernatantat 540 nm using a DU640B spectrophotometer (Beckman).
Liposomes containing 1-palmitoyl-2-oleoyl-sn-glycero-3-phophocholine (POPC) and cholesterol at a ratio of 45∶55 mol% were prepared as previously described [27]. Carboxyfluorescein-containing liposomes were made by adding CF [5(6)-carboxyfluorescein]to the cholesterol/lipid mixture in HBS at a concentration of 50 mM before extruding [40]. After extrusion, the encapsulated liposomes were then passed over a Sephadex G-50 column in HBS pH 7.5 to separate unencapsulated CF from liposomes.
Surface plasmon resonance (SPR) was performed with a BIAcore 3000 system (Oklahoma Medical Research Foundation) using an L1 sensor chip (Biacore) as previously described [22]. Binding analysis was performed as previously described [22] with the following modification: nine consecutive 10 µl injections of the toxins and their derivatives (100 ng per injection) in HBS were passed over the liposome-coated chip at a flow rate of 10 µl/min.
The pore forming activity of PFO and ILY and ILY DM was measured by incubating serial dilutions of toxin with 100 µl of a 1∶1000 dilution of carboxyfluorescein (CF)-containing liposomes in HBS for 1 h at 37°C. Samples were read on a Victor3V Wallac 1420 Multilabel counter (Perkin Elmer) using wavelength settings optimized for high count fluorescein detection.
Two-fold serial dilutions of PFOβ4β5and PFOR468A•β4β5labeled with Alexa-488 were incubated with washed human RBCs (1×106 cells) in PBS for 30 min at 4°C.Samples were then brought to a final volume of 500 µl with cold PBS and analyzed by a FACSCalibur flow cytometer (University of Oklahoma Health Sciences Center), gating on live cells. The emission wavelength was 530 nm and the excitation was 488 nm with a bandpass of 30 nm. The disulfide locked β4β5 versions of PFO and PFOR468A have cysteines substituted for residues Thr-319 and Val-334 [25], which forms a disulfide that prevents β5 from rotating away from β4. This disulfide prevents the lysis, but not binding to the RBCs during flow cytometry [25].
The geometric mean fluorescence of RBCs alone was subtracted from the experimental data for both PFO derivatives and the net fluorescence was graphed using GraphPad Prism.
SDS-agarose gel electrophoresis was performed as previously described [36]. Briefly, samples were incubated with different concentrations of washed hRBCs, for 30 min at 37°C.Toxins were maintained at 440 nM and the hRBCs concentrations ranged from 2.5×107/ml to 2.5×108/ml in a final volume of 40 µl. Samples were solubilized with SDS sample buffer and the complexes were analyzed on a 1.5% SDS-agarose gel (100 V, 120 min) and then transferred to nitrocellulose membranes. Protein bands were identified using rabbit anti-PFO antibody followed by horseradish peroxidase tagged goat anti-rabbit secondary IgG. The bands were visualized using a chemiluminescent substrate (ECL Western Blotting Detection Reagents, Amersham/GE Healthcare) and autoradiography.
Human erythrocyte (hRBC) ghost membranes were prepared as previously described with some modifications [27], [42]. After hypotonic lysis of the hRBCs for 15 min at 4°C in lysis buffer [5 mM sodium phosphate (monobasic), pH 7.5, 1 mM EDTA], cytoplasmic constituents were separated from the membranes by dialysis with 2 L of the lysis buffer by recirculation through a Vivaflow 200 0.2 µm PES cassette (Sartorius Stedim Biotech). Membrane protein content was quantified using the Bradford method (Bio-Rad Protein Assay, Bio-Rad Laboratories, Inc.) as previously described [27].
All fluorescence intensity measurements were performed using a Fluorolog-3 Spectrofluorometer with the fluorescence software (Horiba JobinYvon). NBD measurements were made using the following settings: an excitation wavelength of 480 nm and an emission wavelength of 540 nm with a bandpass of 5 nm. Emission intensity was scanned between 500 and 600 nm at a resolution of 1 nm with an integration time of 0.1 sec. In a typical experiment, labeled and unlabeled samples containing 10 µg total toxin each were incubated with hRBC ghost membranes (equivalent to 300 µg of membrane protein) or 20 µl liposomes in HBS for 15 min at 37°C before taking spectral measurements. For all experiments the fluorescence intensity of the unlabeled samples was subtracted from that of the fluorescent probe-labeled samples in order to control for the intrinsic fluorescence of the sample in the absence of the probe.
FRET analysis was performed as previously described [35] with the following changes. The PFO and PFOR468A derivatives were labeled with either Alexa Fluor 488 (donor, D) or Alexa Fluor 568 (acceptor, A).Parallel samples were prepared containing 10 µg of D-labeled toxin mixed with a 4-fold molar excess of either A-labeled toxin or unlabeled (U) toxin in a total volume of 2 ml. To correct for light scattering and direct excitation of the acceptor, a sample was prepared in parallel in which unlabeled PFO or PFOR468A (U) replaced the donor-labeled PFO to create the UA sample, therefore net DA = DA-UA. The samples were mixed in the presence of hRBC ghost membranes (equivalent to 300 µg of membrane protein) for 15 minutes at 37°C and the donor emission intensity measured from 500 nm to 600 nm. The donor emission intensity of samples in which unlabeled PFO derivatives replaced donor-labeled PFO derivatives was measured and subtracted from the donor-labeled samples to control for any intrinsic fluorescence of the toxin or direct excitation of the acceptor.
|
10.1371/journal.pntd.0002104 | S. haematobium as a Common Cause of Genital Morbidity in Girls: A Cross-sectional Study of Children in South Africa | Schistosoma (S.) haematobium infection is a common cause of genital morbidity in adult women. Ova in the genital mucosal lining may cause lesions, bleeding, pain, discharge, and the damaged surfaces may pose a risk for HIV. In a heterogeneous schistosomiasis endemic area in South Africa, we sought to investigate if young girls had genital symptoms and if this was associated with urinary S. haematobium.
In a cross-sectional study of 18 randomly chosen primary schools, we included 1057 schoolgirls between the age of 10 and 12 years. We interviewed assenting girls, whose parents had consented to their participation and examined three urines from each of them for schistosome ova.
One third of the girls reported to have a history of genital symptoms. Prior schistosomal infection was reported by 22% (226/1020), this was associated with current genital symptoms (p<0.001). In regression analysis the genital symptoms were significantly associated both with urinary schistosomiasis (p<0.001) and water contact (p<0.001).
Even before sexually active age, a relatively large proportion of the participating girls had similar genital symptoms to those reported for adult genital schistosomiasis previously. Anti-schistosomal treatment should be considered at a young age in order to prevent chronic genital damage and secondary infections such as HIV, sexually transmitted diseases and other super-infections.
| Urogenital schistosomiasis (Bilharzia) is a common cause of gynecological disease in adult women. Reports to date indicate that genital lesions in adults become chronic and that the damages make women susceptible to HIV. This is the first study on urogenital schistosomiasis in pre-pubertal girls. We interviewed girls aged 10 to 12 years of age for urinary and gynecological symptoms. The research assistants did not know the schistosomiasis infection status in the school or the individuals. We collected three urines that were examined for schistosome eggs. We found that a significantly increased number of girls with urinary schistosomiasis have stinking, bloody discharge, ulcers, tumors and a burning sensation in their genitals. This indicates that gynecological damages due to schistosomiasis start before sexual activity, and before menstruation. By preventing urogenital schistosomiasis in girls we may have an innovative opportunity to reduce teenage HIV transmission and gynecological disease. This study presents a new aspect of a neglected disease affecting more than 100 million females, long overdue for mass intervention.
| Urogenital schistosomiasis causes gynecological morbidity in adult women [1], [2]. Schistosoma (S.) haematobium is primarily known for its effect on the urinary tract, but in endemic areas schistosomiasis may be the most common cause of genital morbidity and mucosal lesions [3]. An estimated 390 million females are at risk of schistosomiasis infection [4], [5]. It is second only to malaria in terms of public health impact of the parasitic diseases, with more than 100 million females infected, 85% of them live in rural parts of Africa.
Previous studies on urogenital schistosomiasis have been conducted in adult women of childbearing age. S. haematobium ova when deposited in the female reproductive tract seem to be equally distributed in the different genital parts, but are most commonly identified in the cervix and the vagina [6], [7], [8], [9]. Both viable and dead ova may cause tissue reactions, morbidity and symptoms long after contact with infested waters [10], [11]. The disease may manifest itself in both the genital and urinary tract and may be found exclusively in the genitals [6], [12]. In young girls there have only been a few case reports, hypothesizing that the pre-pubertal predilection site is in the vulva [7], [8], [9], [13], [14]. This may partly be because gynecological inspections are not prioritized in rural areas, controversial in virgins, but also because the causal relationship between schistosomiasis and genital lesions in young females has not been explored on a large scale [14], [15].
It has been hypothesized that female genital schistosomiasis poses an increased risk to secondary infections such as human papillomavirus and other STDs. Most importantly, these women have been found to have significantly more HIV [16], [17], [18], [19]. In the wake of the HIV epidemic and a realistic prospect of successful anti-schistosomal mass-treatment programs this study sought to explore if girls before sexual debut had signs of genital disease [20]. In Ugu District, South Africa the study aimed to explore the association between gynecologic symptoms and urinary S. haematobium in young girls.
We carried out a cross-sectional study in 18 primary schools, which were randomized for inclusion from 309 primary schools in the area. We invited all girls aged 10–12 years in the included schools. Girls who were absent on the days of the invitations were excluded, as were girls with serious illnesses, or if their guardians or they refused.
The schools were visited between September 2009 and November 2010 in the predominantly rural Ugu District, KwaZulu Natal, South Africa, an S. haematobium endemic area, which covers 5866 km2 (Figure 1). It has an estimated population of 700 000 almost exclusively isiZulu speaking people, 84% reside in the rural areas, 51% are below the age of 20 years and 55% are female [21].
The study was approved by The Biomedical Research Ethics Administration, UKZN 2009, Ref BF029/07; by the Department of Health, Pietermaritzburg, 2009, Ref HRKM010 - 08; by the Norwegian ethics committee, Ref 469 - 07066a1.2007.535, 2007, and the Departments of Health (2008) and Education (2009) in Ugu District. The Helsinki Declaration was followed. All members of the group, including students and research assistants had passed exams in Good Clinical Practice and signed Declarations of Confidentiality. The interviewers did not know the study subjects beforehand. Prior to the study there were information meetings for the parents, principals, school governing bodies and/or teachers of each school. Informed consent was given by each girl, and the parents/guardians signed consent forms. Identifying information was stored separate from the interview information (in separate towns). All were informed about their right to withdraw and to abstain from answering questions without negative consequences. In order to protect girls from stigmatization the disease was discussed in general terms as urinary schistosomiasis. Treatment for schistosomiasis was offered to all, and all were informed about possible side effects. A private psychologist was hired by the project to take care of referred cases as felt necessary; for psychological, practical and legal issues. When other medical help was required, the girls were referred to a government clinical facility, or offered private care if government services were unavailable. For ethical and community liaison reasons the project staff was not involved in any physical or psychological examinations after referral.
The interviews (30 minutes duration) were conducted face to face in isiZulu (the local language) by trained female fieldworkers. Questions were asked about recent (the last week) or previous urogenital symptoms of itch, burn, ulcers or tumors (swelling, lumps) in the genitals, malodorous discharge, color of discharge and feeling of a burning sensation in the genitals, as well as red urine, dysuria, urge and stress incontinence [1], [2]. They were also asked questions about confounders for bloody discharge (menstruation and red urine), malodorous discharge (sexual intercourse and sexual abuse) and for burning sensations in the genitals (sexual intercourse and dysuria). The girls in a pilot study (same age) had no concept of the local anatomy and we therefore decided to not include questions on exact localization of e.g. tumors. In case the child did not seem to understand, terms were explained and if she seemed too shy/uncomfortable to answer the interviewers were instructed to move to the next question. The discharge color was defined using a custom-made color chart. The study population was not familiar with the details of the questionnaire on beforehand. A water body was defined as a river, dam, lake, stream or pond. Each child was asked if she carried out any of seven specific water-related activities known in the study area (playing/swimming, washing/bathing, laundry, washing blankets, collecting water, fishing and crossing) [22]. Furthermore, there were validated demographic, social and psychological components in the questionnaire that will not be analyzed here.
The researchers aimed to visit each school at least three times. We obtained urine samples from each girl on three consecutive days between 10 am and 2 pm. After gentle tilting we deposited 10 ml of urine into a container with 1 ml methylate-formalin solution and the same week we investigated the specimens by microscopy [23]. After centrifuging, we transferred all of the precipitate onto microscopy slides; the last amount was washed with water before transferred. If the mean number of eggs of the three specimens was higher than 50 per 10 ml urine the infection was classified as high-intensity [24]. One stool sample per child was collected for Kato Katz and analyzed for Ascaris lumbricoides, Ancylostoma duodenale, Taenia solium, Trichuris trichiura and S. mansoni. If there was at least one ovum in a specimen it was defined as positive.
Based on data from studies in adults we estimated the prevalence of genital schistosomiasis to be 30% and urinary schistosomiasis 40% [3]. We hypothesized that the expected prevalence of genital ulcers in the schistosomiasis exposed to be 9% and in the unexposed 4%. To detect a difference with a significance level of 5% and a power of 80%, the sample size would have to be 511 unexposed and 341 schistosomiasis exposed young women, in all 852 subjects. The information was recorded on paper; the personal information sheet was separated from the other information as soon as the record number had been secured. Data was entered into EpiData (interview) or Excel (urine) and subsequently exported into IBM SPSS version 19 (Chicago, Illinois, USA). Chi-square and odds ratios (OR) with 95% confidence interval (CI) were used to compare impacts of water contact or current urinary schistosomiasis infection on genital symptoms. In order to study the impact of other variables (for example menstruation or red urine), logistic regression analysis was applied with a 5% significance level; variables were included if the P-value in the crude association was less than 0.2 and if the Spearman rank correlation coefficient was below 0.7. When there were less than 10 cases, the variable was not included in regression analysis. The statistical analysis was computed using SPSS.
Schools that were randomized for inclusion were visited in no particular order. We invited all pupils aged 10 to 12 years. All schools were visited several times in order to collect guardian consent forms and to find as many students as possible. The parents of 1241/1948 (64%) pupils in 18 schools provided consent. On the days we were in the schools we were able to include 1057 assenting girls. In the first 13 schools, where there was adequate time before exams, the consent forms were returned and signed by 92% (1109/1201) of the parents. The pupils were recruited from grades 1 to 7, median grade 5.
Answers were recorded as missing if the child chose not to answer or did not understand the topics. Seven percent of the girls (71/1019) had started menstruating. Only 5% (51/981) said they had been tested for HIV. Three out of 980 (0.3%) knew they had HIV, as many as 495 said they did not know and 77 girls did not reply to this question. Less than one percent (7/1017) reported to have had intra-vaginal sex. However, two percent (22/953) reported to have been sexually abused. They were referred to psychosocial follow-up. All in all 24 of 1019 young girls had experienced voluntary or involuntary vaginal sex.
Out of the 1057 girls who were interviewed, 970 submitted at least one urine for examination, and out of these 791 submitted three urine samples. S. haematobium eggs were found in 32% (312/970) of the girls. High-intensity urinary infection was found in 28% (88/312) of these. In those who had ova in the urine, the mean intensity of infection was 52 eggs/10 ml urine (range 1–624/10 ml). Among the 658 girls with negative urine specimens, 79% (522) submitted three negative urine samples. There was neither any difference in urinary schistosomiasis infection intensity nor presence of symptoms between those who had submitted three urine samples versus one sample.
Thirty five percent reported to have had genital symptoms (356/1018), and as many as 17% (172/1008) reported genital symptoms the last week. Eighteen girls reported having a genital tumor or an ulcer this last week. Table 1 shows the association between urinary schistosomiasis and symptoms in girls. Controlled for confounders in multivariate analyses the table shows that urinary schistosomiasis remained associated with bloody discharge, a burning sensation in the genitals, genital ulcers, tumors and incontinence. Having had vaginal sex was not significantly associated with any of the symptoms; however the variable ‘vaginal sex’ was forced into the multivariate analyses. It did not influence the association between the symptoms and urinary schistosomiasis or water contact. Likewise, having soil-transmitted helminths did not influence the associations (data not shown) and only one person had S. mansoni. The discharge color was white in 51% (84/164) of the cases; cream color in 35% (58/164) and yellow in 9% (14/164). The discharge had streaks/traces of red in 13% (11/87) of the cases, light red in 66% (57/87) of the cases and an even lighter shade of red (light pink) in 9% (8/87). Patients with symptoms were referred but not investigated by the project.
In order to explore the urinary negative girls in more detail they were first divided into two groups (Figure 2), those in high-endemic school versus those in low endemic. The pupils from the low-endemic schools were further split into two, those who admitted water body contact and those who did not. Figure 2 shows that symptoms are significantly more common in those that have a high intensity of infection. The figure also highlights that low-endemic schools (the ‘most negative’ in the district) have a low prevalence of genital symptoms. Amongst the girls with high-intensity schistosomiasis almost 50% had genital symptoms, compared to less than 5% in the negative girls who lived in non-endemic areas and had no water contact. These girls denied having genital tumors, ulcers, bloody or smelly discharge. Bloody discharge was found in the high-endemic schools only, notably also in those individuals of these schools who were negative for schistosomiasis in three urines (Figure 1, category III). Urge incontinence and genital itch were both relatively constant in the high-endemic schools, but significantly higher than in the low-endemic schools.
We found a significant association between water contact and all the listed symptoms (data not shown). However, among the 364 who denied water contact, 21% (76) had urinary schistosomiasis. Sixty three percent of the girls reported water contact (667/1057), among these 606 submitted urines and 39% (236/606) had S. haematobium ova in urine. Among the girls with three negative urines, 54% (281/522) reported water contact. These reported significantly more genital symptoms the last week than their peers without water contact (p = 0.001).
Twenty two percent of the young girls (226/1020) reported having had urinary schistosomiasis previously. This was significantly associated with current genital symptoms such as bloody discharge (Chi square, p<0.001), malodorous discharge (p<0.001), genital itch (p<0.017) and genital ulcer (p = 0.001). Furthermore 29% (251/853) knew of a family member who had had urinary schistosomiasis or red urine and these factors too were associated with all the queried symptoms (p≤0.002), except genital tumor (p = 0.08).
Twelve percent (129/1057) reported that they had been treated for schistosomiasis previously. Girls who said that they had not been treated had significantly more urinary schistosomiasis than those who had been treated (p<0.001), but not more symptoms (sample size small and p-values range from p = 0.43 to p = 1.0).
In an S. haematobium endemic area girls aged 10 to 12 years with schistosomiasis had significantly more often unpleasant symptoms such as genital ulcers, bloody discharge, malodorous white to yellow cultured discharge, genital itch or tumors than those without this infection. Even before sexual debut and independent of menstruation more than 40% of girls with S. haematobium ova in urine reported having had gynecological symptoms previously, one third reported having it the last week or ‘always’. Girls living in endemic areas without urinary schistosomiasis also had significantly more genital symptoms than their peers in low-endemic schools. This study shows that urinary schistosomiasis, water contact, history of red urine and family history of schistosomiasis (‘Isichenene’ in Zulu) are also associated with the full range of symptoms. As shown in adults previously, the history of water contact was an excellent predictor for genital symptoms also in girls [3], [7], [8], [9], [25]. The girls who had been treated for schistosomiasis previously had the same symptoms as those who denied having received treatment.
In adults the grainy sandy patches have been found to be diagnostic of S. haematobium infection and are significantly associated with discharge [1], [2], [3]. However, the findings in this young population could not be corroborated by a clinical examination. These results are therefore circumstantial, since the gynecological symptoms are not specific for genital schistosomiasis. Without the physical examination and intravaginal tests we cannot confirm schistosomiasis as the etiological factor. Furthermore, vaginal discharge, ulcers and genital itch may have other causes that were not controlled for in this study, such as the sexually transmitted diseases, atopic, irrigative dermatitis or other dermatoses like psoriasis or lichen sclerosis, lice, scabies, or non-specific etiology [26], [27]. Furthermore, one cannot preclude that the current symptoms, although caused by infection with S. haematobium in the lower genital tract may make the genital mucosa more susceptible to super-infections by other agents such as bacterial infections, which in turn may cause the reported symptoms. Likewise, the association between water contact and symptoms may be influenced by social and other practices. Poor perinea hygiene may be more common in a group that has limited access to water; and cultural cleansing rituals may also be hypothetical reasons for the association between water contact and symptoms [28].
In this study three urines were collected and the presence of schistosome ova defined the urinary schistosomiasis positive group. However, this study confirms that even urinary negative cases in endemic areas have gynecological morbidity. Hence the prevalence is most likely higher in this population. A more sensitive diagnostic method, such as antigen detection or PCR, would likely have made the reported finding more apparent, though this was not possible in our study [29].
Some girls denied having had water contact, but were found to still have S. haematobium ova in urine. Girls may be shy, worried about repercussions or not be able to differentiate between urinary and genital symptoms. Some girls were ignorant of some phenomena in the questionnaire such as menstruation or discharge. The interviewers – all female – were trained to explain the differences, however information sessions using dolls followed by more thorough questioning of genital symptoms could perhaps have produced more reliable answers and less under-reporting. This was not done in the present study. Further, the most reliable method to determine water contact is by direct observation, although many schistosomiasis studies have used self-reported water body data [30], [31].
It is well documented that many adult women may have genital schistosomiasis even without having detectable schistosome ova in the urine [3], [12], [32]. Urine investigations may therefore be of limited use in the diagnosis of genital schistosomiasis. Studies have shown that female genital schistosomiasis may cause pathologic blood vessel morphology and fragile blood vessels that may lead to mucosal bleeding [3], [33], [34]. Bloody discharge may be a result of this. Inter-menstrual bleeding, post-coital bleeding, malodorous and abnormally colored discharge and genital itch have been found to be associated with S. haematobium ova in the genitals of adults, even after correcting for sexually transmitted diseases [1], [19], [25], [34].
The girls in this study report the same symptoms as adult women in previous studies [1], [2]. One may fear that young children's genital mucosa are already imbued with calcified S. haematobium ova [16], [25], [35]. Childhood water contact may start very early and these girls may have had S. haematobium infection for several years [36]. One study found that such lesions were refractory to treatment in adults, whereas treatment received before the age of 20 years seemed to offer some protection against genital mucosal pathology [25]. Even so, the morbidity prevalence levels were unacceptably high even in those who had received treatment once in childhood and treatment may have to be given in infanthood in order to prevent genital damage [36]. Furthermore, siblings and people sharing the same water bodies should be given simultaneous treatment in order to reduce re-infection rate and intensity; treatment should be given in low-transmission seasons, and the effect should be secured by several rounds [37].
At the present time there are no suitable tools for the diagnosis of genital schistosomiasis in girls. Abnormal malodorous or bloody genital discharge are mucosal symptoms [2]. In our young study population gynecological investigations were not possible for cultural and technical reasons [3]. Further studies are needed to triangulate the analyses of (1) symptoms and (2) water contact/family history with (3) the objective mucosal findings in genital schistosomiasis.
For the rural clinician history taking and urine analyses are simpler sources of information than gynecological examinations, and especially so in virgins. The findings in this study suggest that young girls in S. haematobium endemic areas have gynecological symptoms as a result of schistosoma infection. Mucosal damages may be present as these young girls enter into their first sexual relationships, making them particularly susceptible to HIV or human papillomavirus infection [16], [17], [36]. Further studies are needed to explore the effects of treatment on the prolific symptomatic manifestations and on decreasing the susceptibility to super-infections before sexual debut.
|
10.1371/journal.pntd.0006113 | Development of a Multilocus Sequence Typing (MLST) scheme for Treponema pallidum subsp. pertenue: Application to yaws in Lihir Island, Papua New Guinea | Yaws is a neglected tropical disease, caused by Treponema pallidum subsp. pertenue. The disease causes chronic lesions, primarily in young children living in remote villages in tropical climates. As part of a global yaws eradication campaign initiated by the World Health Organization, we sought to develop and evaluate a molecular typing method to distinguish different strains of T. pallidum subsp. pertenue for disease control and epidemiological purposes.
Published genome sequences of strains of T. pallidum subsp. pertenue and pallidum were compared to identify polymorphic genetic loci among the strains. DNA from a number of existing historical Treponema isolates, as well as a subset of samples from yaws patients collected in Lihir Island, Papua New Guinea, were analyzed using these targets. From these data, three genes (tp0548, tp0136 and tp0326) were ultimately selected to give a high discriminating capability among the T. pallidum subsp. pertenue samples tested. Intragenic regions of these three target genes were then selected to enhance the discriminating capability of the typing scheme using short readily amplifiable loci. This 3-gene multilocus sequence typing (MLST) method was applied to existing historical human yaws strains, the Fribourg-Blanc simian isolate, and DNA from 194 lesion swabs from yaws patients on Lihir Island, Papua New Guinea. Among all samples tested, fourteen molecular types were identified, seven of which were found in patient samples and seven among historical isolates or DNA. Three types (JG8, TD6, and SE7) were predominant on Lihir Island.
This MLST approach allows molecular typing and differentiation of yaws strains. This method could be a useful tool to complement epidemiological studies in regions where T. pallidum subsp. pertenue is prevalent with the overall goals of improving our understanding of yaws transmission dynamics and helping the yaws eradication campaign to succeed.
| Yaws is a neglected treponemal infection that is often transmitted among children in developing countries. Eradication programs in the 1940–50’s significantly reduced the incidence of yaws, but the disease has resurged. The World Health Organization has proposed to eliminate yaws by 2020, and mass treatment trials are underway in a number of countries. To assist in investigating the molecular epidemiology of yaws, we propose a new method for differentiating strains of the causative agent, Treponema pallidum subsp. pertenue. Using this typing method, we identified seven molecular types in yaws patients from a small island in Papua New Guinea. This method may prove useful in clarifying reinfection vs. relapse, in detecting cases newly imported into a village, in tracking the development of macrolide resistant strains, and in helping to define the transmission of yaws strains within a region.
| Yaws is a highly contagious treponemal infection caused by the bacterium Treponema pallidum subsp. pertenue (T.p. pertenue). It is transmitted by direct skin contact and is symptomatic predominantly in children <15 years of age, usually manifesting as chronic ulcers on the extremities. Latent, or inapparent, infection can persist for decades, often re-emerging as skin lesions or causing painful bone and joint damage [1,2]. Yaws continues to be endemic in a number of tropical countries, particularly in rural regions with lack of public health surveillance. In 2012, the World Health Organization (WHO) proposed a program to eradicate yaws by 2020 [3] using mass drug administration (MDA) with single dose azithromycin. To aid in post-MDA surveillance, a molecular typing scheme is needed to discriminate among T.p. pertenue strains, thus permitting investigators to track the movement of genetically distinct strains in populations and to identify strains newly introduced to already-treated populations. Careful molecular epidemiological studies using typing can assist in understanding the dynamics of disease transmission to improve control of future outbreaks.
T.p. pertenue is closely related to T. pallidum subsp. pallidum, the causative agent of venereal syphilis, which differs from pertenue by less than 0.2% of their genome sequences [4]. These subspecies are indistinguishable serologically and morphologically [1,2], but can be differentiated on the basis of molecular signatures [4–9]. For a number of years, molecular typing has been used worldwide for typing treponemes from syphilis patients. This method is based upon 1) the number of 60-base pair repeats in the acidic repeat gene (arp) gene (tp0433); 2) the restriction fragment length pattern of the Subfamily II Treponema pallidum repeat (tpr) E, G, and J genes (tp0313, tp0317, and tp0621, respectively) [10], and 3) is enhanced by inclusion of the sequence of a polymorphic 300 bp region of the tp0548 gene [11]. This typing scheme has been adopted globally in recent years to create a molecular epidemiology database for syphilis, and also to analyze linkage of specific T.p. pallidum molecular types to specific disease manifestations [11,12]. Nonetheless, the 1) well-recognized difficulty in amplifying the arp and tprE/G/J loci in samples where treponemal DNA is not abundant, 2) the concerns that amplification of the arp target might yield inconsistent results [13,14], and 3) the difficulty that sometimes arises in identifying unambiguously the tprE/G/J restriction patterns have prompted investigators to propose modifications to the typing approach. These include multilocus sequence typing (MLST) approaches with the capability of discriminating genetic differences among syphilis strains without the risk of ambiguous results. New target loci have included tp0136 [5,8,15–17] and tp0279 [18]. Compared to typing methods that rely on restriction fragment length polymorphisms or analysis of tandem repeats, a MLST of proven efficacy would also be more likely to be routinely adopted in research and clinical laboratories.
To provide a better understanding of the current yaws status and to guide control efforts, development of a molecular typing method for T.p. pertenue is highly desirable. Therefore, we sought a sequenced-based typing method using small gene regions that can readily be amplified even from clinical samples with low concentrations of T. pallidum DNA and whose analysis could unambiguously identify yaws isolates carrying different genetic signatures in these loci.
We propose a MLST method for differentiating T.p. pertenue strains using defined regions of tp0548, tp0136 and tp0326. Each of these genes codes for putative (tp0548) or bona fide (tp0136 and tp0326) treponemal surface-exposed proteins shown to be implicated in maintaining the homeostasis of the bacterial cell envelope (tp0326) [19,20], in mediating adhesion to host components (tp0136) [8,21], or hypothesized to mediate nutrient acquisition (tp0548). These typing targets yield a highly discriminating molecular method for distinguishing T.p. pertenue strains.
Historical T.p. pertenue isolates (Table 1) were propagated in New Zealand white rabbits by intratesticular inoculation as previously described [22]. DNA was extracted for PCR amplification using the QIAamp DNA Mini Kit (Qiagen, Valencia CA) following the manufacturer’s instructions, but adding 50 μl of proteinase K (100 mg/ml stock solution) instead of 20 μl and incubating the sample for 2 hours at 56°C. Samples were eluted in 200 μl of H2O and stored at -20C until used for PCR.
Swab samples containing T.p. pertenue were collected from study participants with exudative skin ulcers in Lihir Island, Papua New Guinea (PNG), during a yaws elimination campaign, between May 2013 and October 2016. Following baseline examination and sample collection, mass treatment with single dose azithromycin was administered. Treatment coverage was 84%. The population was re-examined at 6 month intervals for 42 months. At re-examination, swabs were collected from individuals with yaws-like ulcers, and targeted azithromycin treatment was provided to these persons and their family/childhood contacts. Details of the study have been published elsewhere [28,29]. Immediately after collection, the swabs were placed in 1 ml of 1x lysis buffer (10mM Tris-HCl, 0.1mM EDTA, 0.5% SDS), frozen, and transported to the University of Washington. DNA was extracted using the QIAamp DNA Mini Kit (Qiagen) according to the manufacturer’s instructions. Presence of T. pallidum DNA was assessed initially by PCR of the Tp47 gene (tp0574), which is conserved in all Treponema subspecies, including pertenue. Samples positive for T. pallidum DNA underwent amplification of the tp1031 (tprL) gene as previously described [30] and the amplicon size was used to determine the pertenue (vs. pallidum) subspecies. All T. pallidum-positive samples from Lihir Island (n = 232 with duplicates removed), corresponding to 30.7% of all lesion samples analyzed) were confirmed as T.p. pertenue and were used for molecular typing studies. Of these samples, 83.6% (n = 194) were fully typeable with our approach.
Participants with suspected yaws or, for young children, their parents or guardians provided written consent for inclusion in clinical surveys and etiological studies, including collection of swabs used in this study. The study was approved by the National Medical Research Advisory Committee of the Papua New Guinea Ministry of Health (MRAC no. 12.36). Only coded samples were sent to the University of Washington for testing.
Based upon published T.p. pallidum and pertenue genome sequences, we evaluated a number of genes that are polymorphic among strains; these included tp0136, tp0548, bamA (tp0326), tprC (tp0117), tprD (tp0131), and tp0619. Published full length sequences of these genes were initially examined from three human T.p. pertenue strains (Gauthier, CDC2, Samoa D; Table 1) and the Fribourg-Blanc simian isolate. The members of the tpr gene family (tprC and tprD) were not examined further due to the high homology between these two genes and other members of the tpr family, making specific amplification problematic. With the exception of tp0548, which has an already-identified region that is used for T. pallidum subsp. pallidum typing, we identified, within each of the remaining targets (tp0136, tp0326, and tp0619), regions containing polymorphisms potentially suitable to differentiate the strains. Primers were designed to amplify large regions of these genes for preliminary sequence analysis (Table 2). From 95 T.p. pertenue-positive PNG samples collected from baseline through 18 months of the study, we were able to successfully amplify and obtain good sequences from 66 (69%) samples for tp0619, 87 (92%) for tp0136, and 42 (44%) for tp0326. PCR amplifications for these targets were performed using genomic DNA in a 50-μl final volume containing 200 μM deoxynucleoside triphosphates, 1.5 mM MgCl2, 0.8 μM primers (Table 2) and 2.5 U of GoTaq DNA polymerase (Promega, Madison, WI). Cycling conditions for the tp0136 PCR were 95°C for 3 mins, then 45 cycles of 95°C for 1 min, 60°C for 2 min, 72°for 1 min; followed by 72°C for 10 mins. The conditions for tp0326 were 95°C for 3 mins, then 45 cycles of 95°C for 1 min, 56°C for 1 min, 72°for 1 min; followed by 72°C for 10 mins. For tp0619, cycling conditions were 95°C for 5 mins, then 45 cycles of 95°C for 1 min, 55°C for 1 min, 72°for 1 min; followed by 72°C for 10 mins.
Based on the alignments from the historical strains and 95 initial PNG samples, tp0619 proved not to be a suitable typing target: all historical strains had the same tp0619 sequence and there were 5 types identified in the amplified PNG samples (S1 Fig). In comparison, even from the low number of tp0326 sequences that we obtained with these initial primers, we were able to identify 8 tp0326 types. Thus, the three targets selected for further investigation were tp0548, tp0136, and tp0326. Based upon analysis of these large amplicons, we identified relatively short regions containing polymorphisms yielding the maximum number of unique “types” among the samples tested, and selected those as typing targets for our MLST protocol. Primers, amplicon size, and region identifications are shown in Table 3. Amplifications of these targets (tp0548, tp0136, and tp0326 gene fragments) were performed using genomic DNA in a 50-μl final volume containing 200 μM deoxynucleoside triphosphates, 1.5 mM MgCl2, 0.8 μM primers (Table 3) and 2.5 U of GoTaq DNA polymerase (Promega, Madison, WI). Thermocycling conditions for tp0548 have previously been described [11]. Conditions for the tp0136 PCR were 95°C for 3 mins, then 45 cycles of 95°C for 1 min, 59°C for 2 min, 72°for 1 min; followed by 72°C for 10 mins. The conditions for tp0326 were 95°C for 5 mins, then 45 cycles of 95°C for 1 min, 58°C for 1 min, 72°for 1 min; followed by 72°C for 10 mins. All amplified products were treated with ExoSAP-IT PCR Product Cleanup Reagent (Affymetrix, Santa Clara CA) for dye deoxy terminator sequencing in one direction. If ambiguities in base-calling were seen in the electropherograms, we repeated the sequencing in both directions, repeating the PCR when necessary. Further, all gene alleles described in our study were found in more than one clinical sample, thus providing confidence that the typing sequences are correct.
Sequence analysis and alignments of the six human yaws strains of T. pallidum subsp. pertenue (Gauthier, Ghana051, CDC1, CDC2, Samoa D, Samoa F), the Fribourg-Blanc strain, and the PNG samples were performed using Bioedit [31] (http://www.mbio.ncsu.edu/BioEdit/bioedit.html) and Clustal W (http://www.ebi.ac.uk/Tools/msa/clustalw2/). Phylogenetic analysis was conducted by first constructing multiple alignments using the Muscle algorithm implemented in Molecular Evolutionary Genetics Analysis (MEGA) version 7.0 software (http://www.megasoftware.net/) [32], and drawing phylogenetic trees using the Neighbor-Joining method and the number of differences model, with pairwise deletion of gaps and 1000 bootstrap repetitions.
GenBank accession numbers for the new tp0548 types are as follows: R, MF425823; S, MF425824; T, MF425825; U, HM585227; V, HM243495; W HM245777; X, CP020365. The authors that described tp0548 types M, N, P, and Q did not submit the sequences to GenBank, so no accession numbers are available The tp0136 types A-G are MF425826-MF425831 and MF425833. The tp0326 types 1–8 are MF425834-MF425836 and MF425838-MF425842.
Based upon our analyses of historical strains and a subset of PNG clinical samples, we chose tp0136, tp0548, and tp0326 as the most promising targets for use as a T.p. pertenue typing system. Primers were designed and tested for amplification of these regions and those with robust amplification were selected for the MLST scheme (Table 3) The targets, all of which are putative or bona fide outer membrane proteins, each contain small (300–600 nt) readily amplifiable regions with sequence heterogeneity among strains. While the selection of additional, or longer, targets could have increased discrimination, we weighed the resulting requirement for increased sample volume, cost, and time of adding more targets with the risk of losing the ability to fully type some samples. The proposed nomenclature for different T.p. pertenue strain types is expressed as two letters, representing tp0548 and tp0136 types, followed by a number, representing tp0326 types, e.g. JG8.
Using these new MLST primers, we attempted to amplify and sequence typing products from 232 T. pallidum–positive swab samples, collected over 42 months, from persons with chronic ulcers on Lihir Island; 194 (83.6%) samples could be completely typed.
The MLST typing approach based on the three chosen markers was then applied to T.p. pertenue isolates from PNG. Of the 232 total PNG T.p. pertenue-containing samples, 194 (83.6%) were successfully typed at all three loci; 22 (9%) could be partially typed; and 16 (7%) could not be typed at all. Only fully typed samples were further analyzed. During the 3.5 years of subsequent surveys and sampling, a total of seven types (dividing the samples into groups PNG 1 to PNG 7) were observed, with type JG8 (PNG 1) being predominant throughout that period (82%, Fig 8). The distribution of the molecular types during the course of the survey is discussed elsewhere (manuscript submitted).
A phylogenetic analysis of the final tripartite MLST system for T.p. pertenue, based upon the haplotypes (e.g. concatenated tp0548, tp0136, and tp0326 genotypes), is shown in Fig 9. This divided the haplotypes into two major clusters with high bootstrap values, with one containing the three haplotypes with the divergent tp0136 G genotype (PNG 1,3,5), and the other cluster comprising two minor clusters, one containing all historical isolates and the PNG 2 haplotype (SE7), and the other containing the haplotypes with the tp0136 D genotype (PNG 4,6,7).
Whole genome sequencing of the Samoa D, Gauthier, and CDC2 T.p. pertenue strains provided an excellent resource for beginning to develop a genotyping tool for yaws clinical samples [4]. For several years, a molecular typing method originally developed at the Centers for Disease Control [10] has been used to identify circulating strains of T. pallidum subsp. pallidum for epidemiological studies [10,11,16,18,40,41]. The enhanced typing method developed by Marra et al. built upon the earlier method, proved to provide greater discrimination, and has been widely adopted for typing syphilis strains [11,13,16,41]. Similarly, a typing scheme for yaws organisms could help to inform WHO’s yaws eradication program by permitting an examination of the diversity, stability, and movement of strains throughout a geographical area, and the importation of strains by travelers. The typing system will provide a tool to help to identify the resilience of a bacterial population (e.g. the emergence or importation of strains with enhanced virulence or drug resistance, or the occurrence of an outbreak). Also, the new strain-typing technique will help to improve the understanding of yaws transmission pathways, which will inform the development of improved management and preventative interventions. For example, this tool will help to determine the degree to which yaws cases are clustered within villages and districts; identifying the mechanisms for that clustering could contribute to determination of optimal implementation units for interventions. If inter-village yaws transmission were to be identified, public health officials might want to consider establishing larger implementation units. For evaluating clinical episodes, molecular typing may clarify whether repeated episodes of yaws are due to reinfection rather than relapse in patients in whom genotypically different strains of T.p. pertenue were detected from lesions during each of the separate episodes of ulcer.
Because of the significant difficulty inherent to the syphilis typing method, which relies heavily on analysis of restriction fragment length polymorphisms and of variable numbers of repeats, we sought to develop a multilocus sequence typing (MLST) approach for yaws samples that would be more straightforward and reliable to execute and would provide greater resolution while limiting ambiguous results. Based upon our analysis of sequenced yaws strain genomes and a subset of PNG samples, we chose fragments of the tp0136, tp0548, and tp0326 genes as the most promising targets for a T.p. pertenue typing system. Our selection was based primarily upon the level of strain discrimination afforded by the genes and the robustness of the PCR assay in samples containing low concentration of treponemal DNA. We weighed the increased cost and time of adding more targets with the risk of losing the ability to fully type some samples. We fully recognize that, by limiting the size of the gene fragments used in the typing system, we risk losing some discriminating capability. Our experience with typing clinical samples, often from distant locations where optimal handling of DNA is not practical, has convinced us however that the ability to derive a complete molecular typing designation from a high proportion of samples is preferable to a more discriminating system in which a lower percentage of samples can be fully typed. We do not exclude, however, that in the future additional targets might be added to our MLST. Preliminary evidence suggests, for example, that tp0488 might be a suitable typing targets for T. pallidum subsp. pertenue, and its use should be further evaluated.
Evidence for the utility of our novel T.p. pertenue typing system can be found by examining the strain types of the six historical yaws treponemes, which were collected from disparate geographical regions over nearly 3 decades, and could be divided into four molecular types based on our typing system. It was not unexpected to see that Samoa D and Samoa F, which were both isolated from children in Apia, Western Samoa, in January,1953 [25], had the identical molecular type, WB1. Typing and careful literature research can also lead to questioning of the origins of some DNA samples. We initially conducted typing analysis on DNA from two strains (called CDC2571 and Brazzaville) obtained from a laboratory in the Netherlands, and for which no known isolated strains exist. In carefully researching the origin of this DNA, we were unable to find published references describing the isolation of either strain by those names. In our typing analysis, we found that the Brazzaville strain had identical type sequences to the Gauthier strain (S2 Fig). The 1963 publication describing the isolation of the Gauthier strain [23] describes the collection of a sample from Nigeria in 1960 by a physician in Brazzaville; this publication names the sample “Gauthier, Eastern Nigeria”. We therefore suspect that the “Brazzaville strain” is actually the same as the Gauthier isolate.
Similarly, CDC2571 had the same type sequences as CDC1 and Ghana051 (S3 Fig). There is no known description of the isolation of CDC2575 which was provided to the Netherlands lab by Dr. Peter Perine [42]. The cited reference [24] for CDC2575 describes the isolation in hamsters of treponemes from three children with yaws; all hamster inoculations were conducted on the same date, and the children were residents of two towns in Ghana. Only two of the three strains were successfully transferred and propagated in subsequent animals, and these two are named CDC1 and CDC2; the third un-named strain was apparently lost. We therefore suspect that CDC2575 is actually strain CDC1.
The reference that is typically cited for strain Ghana051 [26] describes the 1988 isolation of the organism from a child who had recently emigrated from Ghana, although this publication does not name the strain. While this manuscript was under review, a publication from Strouhal et al. [43] described the genome sequences of CDC2575 and Ghana051, which were virtually identical. The existence of a description of the isolation of the Ghana051 strain and the clear difference in years of reported isolation suggests that Ghana051 (1988) is actually a different strain from CDC1 (1980) and CDC2575 (no description of isolation). The lack of published strain nomenclature for the 1988 isolate leaves the question open, however, as to whether strains were confused or mislabeled during passage or handling over the years. Even whole genome sequencing cannot always determine whether strain mislabeling has occurred.
The utility of strain typing is also apparent in the saga of the Paris case report by Grange et al. [36]. The penile lesion was initially thought to be caused by T.p. pallidum acquired by sexual contact in Pakistan, but the tp0548 sequence, named type J, suggested that it was T.p. pertenue. It was the astute observation of the unusual sequence, called type J, by Mikalova et al. [44] that suggested that the agent was not a pallidum subspecies. Subsequent more extensive analyses suggest that the treponeme present in this ulcer is actually most closely related to T. pallidum subsp. endemicum, the cause of bejel or endemic syphilis. It has been proposed by Mikalova et al. that the tp0548 sequence from this patient is the result of recombination between pertenue and endemicum subspecies [45]. Notably, tp0548 type J is the most prevalent type in the PNG samples that we examined, demonstrating that the tp0548 type J sequence is seen in modern T.p. pertenue strains, as well as in the putative hybrid T.p. endemicum strain that was presumably sexually acquired in Pakistan. The “Paris” sample also provides evidence that the oft-stated belief that only T.p. pallidum is sexually transmitted is not true. With more molecular analyses being conducted on pathogenic Treponema, we increasingly realize that the strict “distinctions” concerning the modes of transmission and, potentially, the clinical manifestations of the T.p. subspecies are becoming significantly blurred [2].
The overlap among subspecies in transmission and clinical manifestations is further suggested by the finding that the agent causing genital ulcerations (typically ascribed to the pallidum subspecies) in wild baboons [46] is most closely related to the yaws-causing pertenue subspecies. Subsequent analyses of the material from these animals revealed a pertenue-like lineage that was nonetheless distinct compared to the historical human yaws strains [47]. It is striking that analysis of DNA from flies associated with baboon lesions [33] revealed that some flies contained tp0548 sequences that clustered with the pertenue subspecies, while others contained Type J tp0548 sequences, discussed above as having been first identified in a T. pallidum subsp. endemicum human genital ulcer swab [36,44,45] and later found by us in the majority of samples from children with yaws (molecularly defined as pertenue) in Papua New Guinea. Molecular typing and gene sequencing has revealed the intersection of the subspecies [30,45,48].
This picture is further complicated by our finding that a majority of the PNG samples described in this study have a tp0136 allele that has previously been described only in Treponema paraluiscuniculi, which causes a venereal infection in wild rabbits and is thought not to be infectious for humans [49]. In other cases in which alleles thought to belong to one subspecies are found in another subspecies, it has been proposed that inter-subspecies recombination has occurred [45,48]. Might our finding represent an example of possible recombination between two treponemal species?
Aside from triggering deeper evaluations of the nature of T. pallidum subspecies discussed above, the establishment of a typing system for a pathogen might assist in assessing the association of a particular molecular type with a disease manifestation. If clear associations can be determined through careful epidemiological studies, typing could have a predictive value for regional clinicians and public health officials. For example, if a T.p. pertenue type strain associated with severe joint inflammation were found to be circulating in a community, local health workers could be on heightened alert for identifying and treating such cases. If associations are strong enough, it might justify the adoption of a typing system in routine surveillance programs or in clinical laboratories. Identification of links between genotype and clinical manifestations in yaws is speculative at this time and awaits further study, but a few studies have found associations of specific T. p. pallidum strain types and syphilis manifestations. For example, the 14D/f strain type of T.p. pallidum was significantly associated with neurosyphilis in a large prospective study [11]. In more recent studies, a cluster of T.p. pallidum type 8D/g strains was seen in cases of ocular syphilis in Seattle [12], and infection with the 14I/a type was found to be a significant predictor of serofast status among syphilis-infected patients [50].
With regard to yaws, infection is commonly believed not to affect the cardiovascular and central nervous systems, and not to be transmitted to the fetus during pregnancy. This oft-repeated “maxim” may reflect lack of extensive knowledge on the pathogenesis of yaws. Alternatively, there may be differences in strain invasiveness. Studies conducted by Edington identified syphilis-like aortitis as a major cause of death in people from Ghana where yaws is endemic [51], while Roman and Roman suggested that there is evidence in the literature to support not only neurological and cardiovascular involvement in yaws patients, but also vertical transmission of the pathogen [52]. In the future, discordant observations and conclusions concerning yaws pathogenesis and manifestations may be explained by genetic differences among strains, and with sufficient clinical data, our typing system might assist in linking genotype and phenotype in T.p. pertenue.
In summary, we have described a new sequence-based typing system for T. pallidum subsp. pertenue, based upon tp0548, tp0136, and tp0326 genes. The proposed method was developed to maximize the discriminating capability of the sequence target regions, balanced by the robustness of the PCR to amplify samples with limiting amounts of treponemal DNA. In this study, we limited our analysis to the aggregated typing results from clinical samples obtained during the 3.5 years of examinations of the population of Lihir Island. An analysis of the geographical clustering of the strain types across the island and the correlation of strain type with population migration or travel will provide critical information for developing protocols and monitoring progress of yaws eradication activities in the future. Those analyses are ongoing.
While this new typing system has been quite useful in examining strains circulating on Lihir Island, it is very important to assess its applicability to samples from yaws lesions from other geographical regions. It is fully expected that more strain types will be identified as the typing method is applied to more yaws-affected populations, and that modifications to the primer sets may be needed. It should also be remembered that no typing system will be universally sensitive, particularly for samples that cannot be collected, stored, or transported under optimal conditions. The discriminating ability of the typing system described here for historical T.p. pertenue isolates from Pacific Islands and Africa, as well as clinical samples, suggests however that it is a good prototype that will be readily applicable to the current WHO campaign to eliminate yaws.
|
10.1371/journal.pgen.1007008 | Analysis of nuclear and organellar genomes of Plasmodium knowlesi in humans reveals ancient population structure and recent recombination among host-specific subpopulations | The macaque parasite Plasmodium knowlesi is a significant concern in Malaysia where cases of human infection are increasing. Parasites infecting humans originate from genetically distinct subpopulations associated with the long-tailed (Macaca fascicularis (Mf)) or pig-tailed macaques (Macaca nemestrina (Mn)). We used a new high-quality reference genome to re-evaluate previously described subpopulations among human and macaque isolates from Malaysian-Borneo and Peninsular-Malaysia. Nuclear genomes were dimorphic, as expected, but new evidence of chromosomal-segment exchanges between subpopulations was found. A large segment on chromosome 8 originating from the Mn subpopulation and containing genes encoding proteins expressed in mosquito-borne parasite stages, was found in Mf genotypes. By contrast, non-recombining organelle genomes partitioned into 3 deeply branched lineages, unlinked with nuclear genomic dimorphism. Subpopulations which diverged in isolation have re-connected, possibly due to deforestation and disruption of wild macaque habitats. The resulting genomic mosaics reveal traits selected by host-vector-parasite interactions in a setting of ecological transition.
| Plasmodium knowlesi, a common malaria parasite of long-tailed and pig-tailed macaques, is now recognized as a significant cause of human malaria, accounting for up to 70% of malaria cases in certain areas in Southeast Asia including Malaysian Borneo. Rapid human population growth, deforestation and encroachment on wild macaque habitats potentially increase contact with humans and drive up the prevalence of human Plasmodium knowlesi infections. Appropriate molecular tools and sampling are needed to assist surveillance by malaria control programmes, and to understand the genetics underpinning Plasmodium knowlesi transmission and switching of hosts from macaques to humans. We report a comprehensive analysis of the largest assembled set of Plasmodium knowlesi genome sequences from Malaysia. It reveals genetic regions that have been recently exchanged between long-tailed and pig-tailed macaques, which contain genes with signals indicative of rapid contemporary ecological change, including deforestation. Additional analyses partition Plasmodium knowlesi infections in Borneo into 3 deeply branched lineages of ancient origin, which founded the two divergent populations associated with long-tailed and pig-tailed macaques and a third, highly diverse population, on the Peninsular mainland. Overall, the complex Plasmodium parasite evolution observed and likelihood of further host transitions are potential challenges to malaria control in Malaysia.
| Plasmodium knowlesi, a common malaria parasite of long-tailed Macaca fascicularis (Mf) and pig-tailed M. nemestrina (Mn) macaques in Southeast Asia, is now recognized as a significant cause of human malaria. A cluster of human P. knowlesi cases were reported from Malaysian Borneo in 2004 [1], but now human infections are known to be widespread in Southeast Asia [2,3], and have been reported in travellers from outside the region [2,4]. Clinical symptoms range from asymptomatic carriage to high parasitaemia with severe complications including death [5,6]. As rapid human population growth, deforestation and encroachment on remaining wild macaque habitats potentially increases contact with humans [7], in Southeast Asian countries P. knowlesi is now coming to the attention of national malaria control and elimination programmes that have hitherto focused on P. vivax and P. falciparum [2].
P. knowlesi commonly displays multi-clonality in humans and macaques, and analysis of microsatellite markers, csp, 18S rRNA, and mtDNA sequences indicates no systematic differences between human and macaque isolates from Malaysian Borneo [8]. Whole genome-level genetic diversity among P. knowlesi from human infections in Sarikei in Sarawak demonstrates substantial dimorphism extending over at least 50% of the genome [9]. This finding is supported by analysis of microsatellite diversity in parasites from Mf, Mn and human infections across Peninsular and Borneo Malaysia [10]. It also provides evidence that the two distinct genome dimorphs reflect adaptation to either of the two host macaque species, although no evidence of a complete barrier in primate host susceptibility was found [10]. A third genome cluster has been described from geographically distinct Peninsular Malaysia [11, 12, 13, 14].
Studies of mtDNA have revealed that ancestral P. knowlesi predates the settlement of Homo sapiens in Southeast Asia, the evolutionary emergence of P. falciparum and P. vivax, and underwent population expansion 30–40 thousand years ago [8]. Diversity at the genomic level is thus likely to reflect host- and geography-related partitioning during this expansion, as well as additional recent complexity due to contemporary changes in host and vector distributions during ongoing ecological transition in the region [15]. Several Anopheles species, all from the Leuchosphyrus group, are capable of transmitting P. knowlesi malaria, including A. latens and A. balbacensis in Malaysian Borneo [16, 17, 18], A. hackeri and A. cracens in Peninsular Malaysia [19] and A. dirus in southern Vietnam [20]. It is thus likely that patterns of genome diversity in natural populations of P. knowlesi reflect partitioning among both Dipteran and primate hosts occurring on varying time-scales through the evolutionary history of the species. Such partitioning can plausibly prevent or reduce panmictic genetic exchange.
Genomic studies of P. knowlesi to date have considered nuclear gene diversity and dimorphism among naturally-infected human hosts, and macaque-derived laboratory-maintained isolates from the 1960s [10, 12]. However, these studies did not consider non-nuclear organellar genomes in the mitochondrion and apicoplast of malaria parasites, which are non-recombinant and uniparentally inherited, and can provide evidence of genome evolution on a longer timescale [21]. Recombination barriers among insect and primate hosts may have less impact on sequence diversity in the organellar genomes of P. knowlesi. Utilising a new P. knowlesi reference genome generated using long-read technology [22] we performed a new analysis of all available nuclear and non-nuclear genome sequences. Patterns of polymorphisms were analysed to identify evolutionary signals of both recent and ancient events associated with the partitioning of the di- or tri-morphic genomes previously reported.
Raw short-read sequence data from all available P. knowlesi isolates (S1 Fig) were mapped to a new reference genome [22] from the human-adapted P. knowlesi line A1-H.1 genome [23], yielding an average coverage of ~120-fold across 99% of the reference genome (S1 Table), and 1,632,024 high quality SNPs. The high density of point mutations (1 every 15bp) in P. knowlesi compared to P. vivax and P. falciparum has been previously noted [10]. Seven macaque-derived isolates were found to have high multiplicity of infection (S2 Fig), and were excluded, leaving an analysis set of 60 isolates.
SNP-based neighbour-joining tree analysis revealed three subpopulation groups that coincide with isolates presenting the Mf-associated P. knowlesi genotype (Mf-Pk, Borneo Malaysia, Cluster 1), the Mn-associated P. knowlesi genotype (Mn-Pk, Borneo Malaysia, Cluster 2) [10, 11, 12, 14], and older Peninsular Malaysia strains (Cluster 3) (Fig 1A). Within Cluster 1 we observed two geographic sub-groups that coincide with Kapit and Betong regions in Malaysian Borneo. The samples from Sarikei region (DIM prefix), geographically located equidistant between Kapit and Betong, fall into either cluster (S3 Fig). Overall, the regional clusters from Kapit and Betong were more genetically similar to each other (mean fixation index FST 0.03, S4 Fig) than were the host-associated clusters (Cluster 1 vs. 2, mean FST 0.21). However, a significant chromosomal anomaly was identified that differentiated the Kapit and Betong Mf-Pk subgroups; this occurred in a multi-gene region on chromosome 8 (~500 SNPs with FST values >0.4; Fig 1B; S4 Fig).
To explore the anomaly in chromosome 8, individual haplotypes and neighbour-joining trees were constructed across several loci (Fig 1C and Fig 1D) revealing two very distinct patterns. The first pattern was observed in the chromosomal sections with low genetic diversity between the two Mf-Pk regional clusters (FST < 0.2, Fig 1B). The tree structure for these genomic regions (Fig 1D, 1st tree) mimics that of the genome-wide tree in Fig 1A. Strong haplotype differentiation between the host-associated Clusters 1 (Mf-Pk) and 2 (Mn-Pk) was confirmed in the SNP-based profiles (Fig 1C, 1st column).
A second pattern was observed in regions of chromosome 8 with distinct genetic differentiation between Kapit and Betong subgroups (FST > 0.4). Many Mf-Pk Betong subgroup isolates presented segments almost identical to chromosome 8 sequences of the Mn-Pk genotype from Cluster 2 (Fig 1D, 2nd, 3rd and 4th trees). This exchange is supported by the SNP-based haplotype patterns, where a distinct haplotype in the Betong samples is Cluster 2-like (Fig 1C, 2nd, 3rd and 4th columns, black arrows), suggesting the introgression of large chromosomal regions (up to 200Kb) between Mf-Pk (Cluster 1) and Mn-Pk (Cluster 2). This is consistent with a very recent event of natural genetic exchange between these subgroups of P. knowlesi recently isolated from human infections. The high frequency of the new haplotype (73%) in the Betong subgroup suggests that it is under (recent) strong selection pressure in this region. The presence of differences in extended haplotype homozygosity between the recombinant and non-recombinant regional Mf-Pk subpopulations provides additional evidence of recent positive selection (XP-EHH peak, P<0.0001) in a region of increased population differentiation (FST > 0.4, Fig 1B).
The functional nature of genes in chromosome 8 involved in these putative introgression events was investigated (FST > 0.4, Table 1), and found to include loci that are important in the vector component of the Plasmodium life cycle. For example, cap380 (PKNH_0820800, 101 SNPs with FST > 0.4) encodes a protein expressed in the external capsule of the oocyst. This gene is essential in the maturation from ookinete into oocyst in P. berghei, and is assumed to assist in evasion of mosquito immune mechanisms [24]. Another gene, PKNH_0826900 (19 SNPs) encodes for the circumsporozoite- and TRAP-related protein (CTRP), which has an established role in ookinete motility in P. berghei and is essential for binding to and invading the mosquito midgut [25]. Further, homologues of PKNH_0826400 (21 SNPs) display increased transcription levels in ookinete and gametocyte V sexual stages in both P. falciparum [26] and P. berghei [27] compared to the asexual ring stage (fold change of at least 2). The transcriptomic profiles of these strongly selected genes are shown in S5 Fig.
By applying a combination of neighbour joining trees and SNP diversity analysis across 50 Kbp windows, we identified that 33/60 isolates show clear evidence of genetic exchange between Clusters 1 and 2 (S2 Table). Regions involved in exchange (recombination) (137/494 regions, 86% contained an ookinete related gene) showed evidence of enrichment for ookinete-expressed genes compared to other (non-recombinant) chromosome regions (357/494 regions, 77% contained an ookinete related gene) (Chi Square P = 0.03). One such region in chromosome 12 included the Pf47-like (PKH_120710) gene, where the orthologue in P. falciparum is a known mediator of the evasion of the mosquito immune system [28]. Furthermore, it has been shown that a change in haplotype in this gene in a P. falciparum isolate is sufficient to make it compatible to a different mosquito species [28]. Nearly half (45%) of isolates from Betong presented with a recombinant profile in PKH_120710.
In general, the genetic exchanges generated differing levels of mosaicism in each population and among individual isolates across all chromosomes (S6 Fig). One isolate from Sarikei with the Mf genome dimorph type (DIM2) appeared to harbour Mn-type introgressed sequences in 8% of the genome, occurring across 6 chromosomes (6, 7, 8, 9, 11 and 12), including an almost complete Mn-type chromosome 8. Of the 33 samples with evidence of exchanges, 13 were from the Betong region, 14 from Kapit and 6 from Sarikei, which indicates that the events are not geographically restricted. Although, the majority of genetic exchange events involve the integration of Mn-type motifs into Mf-type genomes, introgression in the opposite direction was also observed, but on a smaller scale and at lower frequency.
The mitochondrial and apicoplast genomes of each P. knowlesi isolate was interrogated for signals of evolutionary history over longer time-scales, as in previous studies [21, 29, 30]. Combining the mitochondrial sequence data from the 60 P. knowlesi isolates from this study together with 54 previously published mitochondrial sequences including human and both Mn and Mf samples [9], we generated a phylogenetic tree (Fig 2). This tree shows four clades (shown in purple, red, blue and green). To interpret these clades, they were cross-referenced to the previously defined 3 nuclear genotypes (Clusters 1 to 3) and the host contributing the sample (human, macaque-type). The red and purple clades possess similar mitochondrial haplotypes as highlighted by their inter-cluster average FST (red vs. purple: average FST = 0.16), which is lower than comparisons including the other two clusters (red or purple vs. blue or green: average FST > 0.18). The purple clade consists of cultured isolates from Peninsular Malaysia, and is associated with the Peninsular nuclear genotype (Cluster 3). The red and green clades each contain a mixture of Borneo Malaysia samples from both humans and macaques with nuclear genotypes from Clusters 1 and 2. The green clade also includes the only sequence sourced from a M. nemestrina host. The blue clade contains samples from humans and macaques, all with Cluster 1 nuclear genotypes. The divergence of these mitochondrial clades from their common ancestor was estimated to be 72k years ago, and younger than the previous the estimate of 257k but within error [8]. Furthermore, the presence of monkey-derived sequences spread across the tree seems to indicate that none of the mitochondrial genotypic groups found is human-specific as all have also been observed in macaques, also consistent with previous findings [9].
Using the common SNPs (280/425 with MAF > 5%: apicoplast 252, mitochondria 28 SNPs) in the 60 isolates with the sequence data we confirmed that the organellar genomes are co-inherited (mean pairwise organellar linkage disequilibrium D’ = 0.99). SNP-based haplotype profile analysis (S7A Fig) revealed clustering that is consistent with the three main clusters seen in Fig 2. Similarly, a phylogenetic tree constructed using only apicoplast SNPs (S7B Fig) is congruent with the mitochondrial based tree (Fig 2). The presence of mismatched nuclear and organellar type genomes in two of the three clusters (black arrows in Fig 2) and the presence of such mismatched samples with little or no evidence of nuclear genome recombination suggests ancient genetic exchange events between distinct lineages. The nuclear footprints of such exchanges are likely to have been broken down by recombination over time. We observed a significant incongruence between the robust phylogenetic tree topologies based on organellar and nuclear genome SNPs (Shimodaira-Hasegawa test P = 0.001; Templeton test P = 0.003) (Fig 2). These results from organellar and nuclear genomes, in a small but geographically diverse set of P. knowlesi, indicate that there have been several genetic exchanges between the host-associated clusters in Malaysian Borneo.
P. knowlesi is now the major cause of malaria in Malaysian Borneo, but the biology of the parasite [15, 22, 23], host and vector interactions, and disease distribution and epidemiology [19, 31, 32] are not well understood. The availability of a new high-quality reference sequence and a more robust approach to MOI were used to re-evaluate the previously described peninsular and macaque-associated subpopulations of P. knowlesi parasites. We report two major new findings. First, clear evidence of natural genetic exchanges between the divergent Mf- and Mn-associated subpopulations of P. knowlesi, including a major segment of introgression on chromosome 8, is presented. Second, the presence of haplotype sub-divisions in the organellar genomes that do not map onto the subpopulations implied by nuclear genome analysis indicate that exchange events have previously occurred in non-recent history. A similar multi-tiered pattern of evolution among nuclear and organellar genomes has been found in Trypanosoma cruzi, an unrelated protozoan parasite with a mammalian host-insect vector life cycle [29, 30].
Unexpectedly, observed mosaicism and population differentiation signals were not encountered equally across the P. knowlesi nuclear genome, but were particularly prominent on chromosome 8, with genes expressed in mosquito stages over-represented. For example, the majority (73%) of Mf-associated isolates from Betong harboured the Mn-associated allele of the oocyst-expressed cap380 gene, which differs at 101 positions from the allele found in the Mf-associated cluster. This is essential for ookinete to oocyst maturation and therefore for the transmission of the parasite during the vector stage [24, 25]; here, we identify signals of recent selective pressure on this locus (Fig 1B). Other vector-related genes were identified within the introgressed segment, and point towards strong evolutionary selection pressure on the parasites driven by the transmitting Anopheles vector species. Such effects have been found in P. falciparum [28] and P. vivax genomes [33], and highlight the importance of understanding the distribution of the different Anopheles vector species, their host feeding preferences, and their interactions with the parasite in highly dynamic and complex environments such as the ecological niche of P. knowlesi.
Nearly 80% of Malaysian Borneo has undergone deforestation or agricultural expansion, which have driven habitat modification affecting both macaque and Anopheles host species, and the proximity of humans to both [8, 31]. Furthermore, studies have predicted that Mn predominantly inhabits forested areas while Mf reside in more cosmopolitan areas, which include croplands, vegetation mosaics, rubber plantations and forested areas [8, 34]. The main genomic exchange event on chromosome 8 involves essential vector-related genes and is pin-pointed geographically to the Betong area. This region has undergone significant forest degradation due to expansion of industrial plantations in the recent years [35]. These types of environmental changes have been previously related to alterations in the vector species distribution in Malaysia, leading to malaria epidemics [36]. Environmental changes also affect macaque habitats, and increase the opportunities for human-macaque interaction [31], but selection events highlighted in this study seem to primarily reflect adaptation of the parasite to changes in mosquito distribution or to recent changes in the vectorial capacity of the existing vectors. The depth, breadth and spread of the genetic exchanges observed in three different areas (Betong, Kapit and Sarikei) in Sarawak highlight the potential importance of these events for parasite adaptation in both vertebrate and invertebrate species.
Although, the level of genetic diversity between Mf- and Mn-associated P. knowlesi has some similarity to that observed between P. ovale curtisi and P. o. wallikeri, now considered separate species [37], the evidence of recombination and genetic exchanges observed in this study precludes species designation, as reproductive isolation is not complete. Nevertheless, better understanding of P. knowlesi population structure could aid future studies across the regions where human populations have been identified at risk of infection including both symptomatic and asymptomatic cases [4, 38, 39]. This would assist with characterising and tracking subpopulations and genetic exchanges, and provide a flexible framework for better understanding P. knowlesi diversity across the region.
Our work has provided insight into Plasmodium parasite evolution. It has been suggested that malaria parasites have survived using either adaptive radiation where host switching plays a key role [40], or alternatively adaptation to complex historical and geographical environments leading to speciation [41]. Plasmodium species in non-human natural conditions in the absence of drug selection pressure have a wide range of possible hosts [41, 42]. The P. knowlesi data has shown that geographical or ecological isolation of the different hosts over an extended time can generate subgroups of parasites with substantial genetic differentiation, but capable of recombining when in contact [12, 30, 31]. This pattern has a major impact on the parasite genome, as illustrated by the profound chromosome mosaicism observed among our study isolates. Our data suggest that the broad host specificity of some of the Plasmodium species are important drivers of parasite genomic diversity. In P. knowlesi this means that genetic divergence is enabled not only by long-term geographic isolation, as is the case between Peninsular and Bornean isolates, but also via the isolation afforded by extended transmission cycles within different primate hosts. The genetic trimorphism suggests that the separate macaque hosts provides sufficient genetic isolation to allow for host specific adaptations to occur, even within relatively small geographic areas. Furthermore, the possibility of recombination between partially differentiated parasite genomes increases opportunities for new adaptation, including further host transitions, and can only make malaria control more difficult. Genome-level studies on P. knowlesi isolates from Mf and Mn across the parasite’s geographic range are now needed to test the generalizability of this remarkable conclusion.
Raw sequence data were downloaded for 48 isolates from Kapit and Betong in Malaysian Borneo [11], 6 isolates from Sairikei in Malaysian Borneo (S1 Fig) [9] and 6 long-time isolated lines, maintained in rhesus monkeys sourced originally from Peninsular Malaysia and Philippines [11]. The sequence data accession numbers can be found in S1 Table. The samples were aligned against the new reference for the human-adapted line A1-H.1 (pathogenseq.lshtm.ac.uk/knowlesi_1, accession number ERZ389239, [22]) using bwa-mem [43] and SNPs were called using the Samtools suite [44], and filtered for high quality SNPs using previously described methods [45, 46]. In particular, the SNP calling pipeline generated a total of 2,020,452 SNP positions, which were reduced to 1,632,024 high quality SNPs after removing those in non-unique regions, and in low quality and coverage positions. Samples were individually assessed for detecting multiplicity of infection (MOI) using: (i) estMOI [47] software, and (ii) quantifying the number of positions with mixed genotypes (if more than one allele at a specific position have been found in at least 20% of the reads [46]). The measures led to correlated results (r2 = 0.8), which highlighted the robustness of these two methods. Samples were classified into three subcategories: (i) single infections (> = 98% genome showing no evidence of MOI and < = 1/10,000 SNP positions with mixed genotypes), (ii) low MOI (>85% genome showing no evidence of MOI and < = 4/10,000 SNPs positions with mixed genotypes); (iii) high MOI (<85% genome showing no evidence of MOI, and > 4/10,000 SNPs positions with mixed genotypes). Samples with high MOI were removed from subsequent analyses.
For comparisons between populations, we first applied the principal component analysis (PCA) and neighbourhood joining tree clustering based on a matrix of pairwise identity by state values calculated from the SNPs. We used the ranked FST statistics to identify the informative polymorphism driving the clustering observed in the PCA [48]. Finally, we created haplotype plots using only SNP positions with MAF > 0.05 over all the populations, and displayed each sample as a row to allow closer inspection of the chromosome regions where interesting recombination events are observed. The XP-EHH metric [49] implemented within the rehh R package was used to assess evidence of recent relative positive selection between regional clusters from Kapit and Betong. The results were smoothed by calculating means in 1 Kbp windows, where windows overlapped by 250bp. The raXML software (v.8.0.3, 1000 bootstrap samples) was used to construct robust phylogenetic trees (90% bootstrap values > 95) for nuclear and organellar SNPs. Estimates of divergence times for subpopulations was based on a Bayesian Markov Chain Monte Carlo (MCMC) (BEAST, v.1.8.1) approach applied to mitochondrial sequences, with identical parameters settings to those described elsewhere [8]. The Shimodaira-Hasegawa [50] and the Templeton [51] tests were used to detect incongruence between the tree topologies.
In order to identify regions that have undergone introgression we calculated the pairwise SNP diversity (π) of each sample against all the Borneo samples using a 50 Kbp sliding window. This window size was sufficient to include the required number of SNPs for the robust identification of introgression events. The average π in the M. nemestrina associated (Mn-Pk) and M. fascicularis associated (Mf-Pk) clusters was calculated, leading to two diversity values for each sample (Mfπ and Mnπ) and thereby a measure of genetic distance to the average of the two clusters. For Mf samples, an increase in the Mfπ and a decrease in Mnπ would mean the sample is more similar to the Mn-Pk cluster than the average; vice versa for the Mf samples. In order to avoid the identification of spurious events, we applied a threshold of a 0.001 increase in the deviation from the original cluster.
For P. knowlesi genes of interest, orthologues in P. falciparum and P. berghei genomes were identified using PlasmoDB (plasmodb.org). Gene expression data (including from the RNAseq platform) for these genes across different stages of the life cycle of the parasite were considered [26, 27]. In particular, we compared the average of the asexual blood stages and the sexual ookinete stage, highlighting the genes upregulated with a two-fold change (P<0.000001), for P. falciparum [26] and P. berghei [27].
|
10.1371/journal.pntd.0006950 | Safety of azithromycin in infants under six months of age in Niger: A community randomized trial | Mass azithromycin distribution reduces under-5 child mortality. Trachoma control programs currently treat infants aged 6 months and older. Here, we report findings from an infant adverse event survey in 1–5 month olds who received azithromycin as part of a large community-randomized trial in Niger.
Active surveillance of infants aged 1–5 months at the time of treatment was conducted in 30 randomly selected communities from within a large cluster randomized trial of biannual mass azithromycin distribution compared to placebo to assess the potential impact on child mortality. We compared the distribution of adverse events reported after treatment among azithromycin-treated versus placebo-treated infants. From January 2015 to February 2018, the caregivers of 1,712 infants were surveyed. Approximately one-third of caregivers reported at least one adverse event (azithromycin: 29.6%, placebo: 34.3%, risk ratio [RR] 0.86, 95% confidence interval [CI] 0.68 to 1.10, P = 0.23). The most commonly reported adverse events included diarrhea (azithromycin: 19.3%, placebo: 28.1%, RR 0.68, 95% CI 0.49 to 0.96, P = 0.03), vomiting (azithromycin: 15.9%, placebo: 21.0%, RR 0.76, 95% CI 0.56 to 1.02, P = 0.07), and skin rash (azithromycin: 12.3%, placebo: 13.6%, RR 0.90, 95% CI 0.59 to 1.37, P = 0.63). No cases of infantile hypertrophic pyloric stenosis were reported.
Azithromycin given to infants aged 1–5 months appeared to be safe. Inclusion of younger infants in larger azithromycin-based child mortality or trachoma control programs could be considered if deemed effective.
ClinicalTrials.gov NCT02048007.
| Trachoma control programs currently treat all adults and children age 6 months and older in communities endemic for trachoma. If shown to be safe, programs could consider inclusion of younger children in mass treatment programs. Here, we evaluated adverse events in infants aged 1–5 months who were participating in a placebo-controlled trial of mass azithromycin for the reduction of child mortality in Niger. Overall, there was no difference in the frequency of adverse events among children treated with azithromycin compared to placebo. Common adverse events in both arms included diarrhea, vomiting, and skin rash. Azithromycin distribution to children between 1 and 5 months of age appeared to be safe. Inclusion of younger children in azithromycin-based trachoma and child mortality programs could be considered.
| Mass azithromycin distribution has been a core component of the World Health Organization (WHO)’s trachoma control program, with over 700 million doses of azithromycin distributed to adults and children aged 6 months and older [1,2]. Mass azithromycin distribution dramatically reduces the prevalence of the ocular strains of Chlamydia trachomatis that lead to trachoma [3–7]. For most indications, azithromycin is approved by the Federal Drug Administration (FDA) for use in children over 6 months of age, and programmatic treatment of children under 6 months with azithromycin has been limited by lack of safety data [8]. Observational studies have documented an increase in infantile hypertrophic pyloric stenosis (IHPS) following the use of macrolides in a child’s first month of life, with the greatest risk associated with macrolide using during the first 14 days and with erythromycin in particular [9–11]. These studies are limited by confounding by indication, as infants receiving macrolides are generally sicker than their untreated peers and may have different indications for treatment than those receiving other antibiotic classes. If shown to be safe, treatment of children less than 6 months of age may be beneficial for trachoma control programs, as infants infected with C. trachomatis have been shown to have a higher chlamydial load [4]. Higher chlamydial loads have been shown to correlate with disease severity and children with higher loads may be more likely to transmit infection [12].
Recently, the MORDOR (Macrolides Oraux pour Réduire les Décès avec un Oeil sur la Résistance) trial found a 14% reduction in all-cause child mortality following four rounds of biannual mass azithromycin distribution to all children aged 1–59 months in Malawi, Niger, and Tanzania compared to biannual placebo [13]. In MORDOR, the largest effects were seen in children under 6 months of age, with nearly 1 in 4 deaths averted in azithromycin-treated communities. MORDOR enrolled children as young as 1 month (>28 days) of age. The primary MORDOR trial was a large simple trial [14], and as such was not designed to efficiently evaluate adverse events other than mortality. Therefore, 30 communities were randomly selected from each country for more intensive monitoring, including active adverse event assessments. Here, we present adverse event data from children aged 1 to 5 months from the Niger site of the trial, to establish the safety of provision of azithromycin to infants under 6 months of age.
MORDOR was a community randomized, placebo-controlled trial conducted in Mangochi, Malawi, Boboye and Loga, Niger, and Kilosa, Tanzania (ClinicalTrials.gov NCT02048007) that compared biannual azithromycin distribution compared to biannual placebo distribution for the prevention of childhood mortality. The current study is restricted to the Niger study site. Methods for the trial have been previously reported [13]. Eligible communities had a population between 200 and 2,000 inhabitants on the most recent national census. Children were eligible for treatment if they were between 1 and 59 months of age and weighed at least 3,800 grams at the time of treatment. In each country, 30 of the randomized communities (15 per arm) were randomly selected to participate in a morbidity sub-study. The morbidity communities included additional assessment of nutritional status as well as rectal swabs, nasal swabs, nasopharyngeal swabs, and dried blood spot collection in a random sample of 50 children in each community. The study intervention (biannual azithromycin or placebo) and census were identical in morbidity and mortality communities. The infant adverse event study was conducted in morbidity communities in Niger following each treatment round among all infants aged 1 to 5 months per the most recent census.
Morbidity communities were randomized in a 1:1 fashion using R (R Foundation for Statistical Computing, Vienna, Austria). Participants, observers, investigators, and those performing data cleaning were masked to treatment arm. The placebo was identical in appearance and in packaging to the oral azithromycin suspension.
A door-to-door census was conducted prior to each treatment round. All children aged 0–59 months and pregnant women were enumerated. Vital status was assessed (dead, alive, unknown) at each follow-up census. Children who were aged between 1 and 5 months during each census round in the morbidity communities were eligible for the infant adverse event survey.
Every child aged 1–59 months at the most recent census was offered a single dose of directly observed oral azithromycin or placebo (both provided by Pfizer, Inc, New York City). Each child was given a volume of suspension equivalent to 20 mg/kg as estimated by height stick approximation (per Niger’s trachoma guidelines) or by weight for those unable to stand. Treatment was given after each examination round, with a treatment coverage target of 80%.
Following each treatment round, the caregivers of infants aged 1–5 months were interviewed regarding adverse events since the last treatment, with a goal of interviewing caregivers within 2 weeks of treatment. A list of all infants aged 1–5 months based on the most recent census in each morbidity community was generated, and we attempted to interview caregivers of all children. Caregivers were asked if their child was treated as part of the study, and for those treated, if the child had a health problem in the two-week period following treatment and if the child was brought to a health clinic for treatment. Only caregivers of children who received the study treatment were asked about health problems to estimate the incidence of health problems in treated children, as inclusion of untreated children could have biased estimates towards the null. To estimate the intention-to-treat effect, all caregivers, regardless of whether or not the mother reported that the child was treated, were asked if the child had any of the following symptoms since the last time the study team visited the child’s community: abdominal pain, vomiting, nausea, diarrhea, dyspepsia, constipation, hemorrhoids, or skin rash.
The sample size for the morbidity communities was based off the primary morbidity outcome, which was macrolide resistance in Streptococcus pneumoniae. We assumed 12% baseline resistance (based on previous studies) and an ICC of approximately 0.051 (based on the Trachoma Elimination Follow-up study [15]). We estimated that inclusion of 30 villages (15 per arm) and 10 samples per community would yield approximately 80% power to detect a difference in prevalence of resistance of 18% (e.g., 12% versus 30%) assuming 80% carriage of S. pneumoniae. For the infant adverse event survey, the sample size was limited by the number of 1 to 5 month old children residing in the 30 communities during the study period.
Descriptive characteristics were calculated with medians and interquartile ranges (IQR) for continuous variables and proportions for categorical variables. Generalized linear models were used to compare 1) if the child had a health issue within two weeks of treatment and 2) if the caregiver sought medical care for the child within two weeks of treatment between the azithromycin and placebo arms. Because the survey restricted questions related to health issues following treatment to children who had received treatment, models were restricted only to children who received treatment per caregiver report. A repeated measures model was used to assess whether there was an overall difference in the distribution of adverse events in azithromycin- versus placebo-treated infants, with a random effect for each child and study community. To estimate risk ratios for any adverse event and each adverse event individually in azithromycin-treated infants compared to placebo-treated infants, we used generalized linear models with a binomial distribution and log link, with standard errors clustered by the community of residence of the infant (the unit of randomization). All analyses were conducted in Stata version 14.2 (StataCorp, College Station, TX) and R version 3.4.3 (The R Foundation for Statistical Computing).
Of 2,056 eligible infants, caregivers of 1,712 (83.3%) were interviewed between June 2015 and February 2018 (Fig 1). The median time between each community’s treatment and the caregiver survey was 34 days (IQR 21 to 61 days). There was no difference in the time between treatment and caregiver survey between azithromycin and placebo-treated communities (P = 0.78). Table 1 shows baseline characteristics of communities and infants included in this study. Approximately half of the children were female (48.5%) and median age at the time of the census was 2 months (IQR 1 to 4 months). Caregivers reported that 70.2% (N = 1,201) of infants received study treatment, with no differences in treatment between arms (P = 0.22).
Among infants for whom the caregiver reported receiving study treatment, there was no difference in reports of any health problems in the two-week period following treatment (34.1% azithromycin versus 40.3% placebo, P = 0.24, Table 2). Similarly, there was no difference in visiting a health clinic for a health problem among those who received treatment (P = 0.27).
There was no difference in the overall distribution of adverse events in children between the two study arms (P = 0.43, repeated measures model). Overall, caregivers of 32.7% of all infants reported at least one adverse event (Table 3). The most commonly reported adverse events in the period following treatment included diarrhea (25.2%), vomiting (19.3%), and skin rash (13.1%). There was no difference overall in report of any adverse event (RR 0.86, 95% CI 0.68 to 1.10, P = 0.23). Infants in the azithromycin-treated arm had reduced risk of diarrhea (RR 0.68, 95% CI 0.49 to 0.96, P = 0.03) and hemorrhoids (RR 0.27, 95% CI 0.08 to 0.87, P = 0.03). The distribution of all other adverse events was similar between treatment arms. There were no reported cases of IHPS.
Trachoma control programs currently only treat children aged 6 months and older, with younger children receiving topical tetracycline. The majority of FDA-approved indications for azithromycin include children aged 6 months and older. Here, we were unable to find any evidence of an increase in adverse events in a sample of infants receiving azithromycin versus placebo as part of a community randomized trial. These results suggest that azithromycin treatment may be safe in infants under 6 months of age, and expansion of indications for azithromycin in public health programs such as trachoma control to include children under 6 months of age could be considered.
Previous evaluation of the safety of azithromycin in children under 6 months of age have consisted of large epidemiologic studies or small randomized controlled trials of azithromycin use among very low birth weight neonates for the prevention of bronchopulmonary dysplasia [8–10,16–19]. Epidemiologic cohorts have shown no increase in risk of IHPS in children over 6 weeks of age compared to untreated children, but may be subject to confounding by indication. In the general population, the vast majority of IHPS cases are diagnosed during the first 12 weeks of life, with a sharp decline in incidence after the 5-6th week of life [20]. Population-based estimates of IHPS in sub-Saharan Africa are rare, but IHPS is thought to be less common in sub-Saharan Africa than in other regions, potentially due to differences in practices that have been shown to increase risk, such as bottle and formula feeding [21–23]. Although no cases of IHPS were reported, the present study was underpowered to assess IHPS, given its rarity particularly among children over one month of age [9,20]. Projectile vomiting is the most common symptom of IHPS [24], however the risk of vomiting in the present study was lower in azithromycin-treated infants compared to placebo-treated infants, suggesting that azithromycin did not lead to IHPS in this population. While future studies using azithromycin in children under 12 weeks of age should remain vigilant in screening for IHPS, the results of this study suggest that the risk of IHPS is likely rare in this population.
Trachoma control programs distribute azithromycin to children and adults aged 6 months and older in communities with endemic trachoma. Although infants under 6 months are thought to have lower infection rates than older children [25], earlier treatment of infants for trachoma may reduce community prevalence of C. trachomatis as infants may have a higher chlamydial load if infected [4]. Caregivers of infants under 6 months of age are given topical tetracycline ointment with instructions to apply the ointment daily for 6 weeks, however completion of the regimen is generally thought to be poor [25–28]. The ability to expand azithromycin distribution to children as young as one month of age could potentially contribute to reductions in trachoma in endemic regions if treatment of children under 6 months of age with azithromycin is shown to be effective for trachoma control.
A subgroup analysis of the MORDOR study demonstrated a nearly 25% reduction in mortality among infants under 6 months of age compared to an overall decrease of 14%, generating the hypothesis that the largest effects of azithromycin for prevention of child mortality may be in the youngest age groups [13]. Previous studies have shown a significant decrease in child mortality among children aged 6–59 months in the context of azithromycin distribution for trachoma control [29–31]. Younger children are at higher risk of mortality compared to older children [32]. The potential for benefit from a mortality-reducing intervention, such as azithromycin, may be greater in this age group than in all children under the age of 5. In the parent study, which included more than 300,000 person-years at risk, few adverse events were reported, although active surveillance was not undertaken. The parent study was a large simple trial designed specifically to evaluate the effect of azithromycin on mortality, which can be considered the most serious adverse event. However, this design was not efficient for evaluation of other adverse events due to the sample size required for the study to be adequately powered for the mortality outcome, given that mortality is a rare event. Active adverse event monitoring was therefore only conducted in the smaller morbidity study, which included more intensive monitoring of study participants.
The results of this study must be considered in the context of several limitations. MORDOR did not enroll children under one month of age, and thus we cannot comment on the safety of azithromycin in neonates. Active surveillance was conducted via caregiver report, which could be subject to social desirability or recall biases. Due to inability to link study records to clinic records, we did not attempt to validate caregiver responses against health post or hospital records. Estimates may therefore be an over- or underestimate of the true burden of adverse events. Although we planned to survey caregivers within two weeks of treatment, due to logistical challenges the survey was often conducted several weeks following treatment. A longer duration between treatment and the survey could increase the likelihood of misclassification. However, due to the use of masked placebo, any misreporting is unlikely to be differential with respect to study arm. In addition, a longer duration between treatment and the survey could increase the number of events that occurred with decreased probability that they were related to study treatment, which could bias results towards the null. However, there were no significant differences in the effect of azithromycin versus placebo by timing of the survey on any adverse event reporter in this study. Although we attempted to interview the caregiver of each child, only 83% of eligible children’s caregivers were interviewed. Due to the placebo-masked nature of the study, differential response by arm is unlikely, however it is possible that caregivers of children who died were less likely to be interviewed. The probability of mortality is very low, and thus unlikely to substantially bias results. This study was conducted in one of three of the MORDOR trial sites, in a region of the Sahel with very high child mortality and infection rates. The results of this study may only be generalizable to other regions with similar distributions of childhood infection.
The results of MORDOR suggest there may be a large reduction in mortality in the 1 to 5 month age group with the use of azithromycin, but as a large simple trial, the trial was not ideal for quantifying common adverse events. In this ancillary study, we were unable to find a difference in adverse events in infants aged 1 to 5 months participating in a large community-randomized trial of biannual mass azithromycin distribution for prevention of child mortality. Currently, several guidelines indicate azithromycin for use in children over 6 months of age. These results suggest that azithromycin could be considered in infants over 1 month of age, and their inclusion in various public health programs using azithromycin may be appropriate.
|
10.1371/journal.ppat.1004490 | Densovirus Is a Mutualistic Symbiont of a Global Crop Pest (Helicoverpa armigera) and Protects against a Baculovirus and Bt Biopesticide | Mutualistic associations between symbiotic bacteria and their hosts are common within insect systems. However, viruses are often considered as pathogens even though some have been reported to be beneficial to their hosts. Herein, we report a novel densovirus, Helicoverpa armigera densovirus-1 (HaDNV-1) that appears to be beneficial to its host. HaDNV-1 was found to be widespread in wild populations of H. armigera adults (>67% prevalence between 2008 and 2012). In wild larval populations, there was a clear negative interaction between HaDNV-1 and H. armigera nucleopolyhedrovirus (HaNPV), a baculovirus that is widely used as a biopesticide. Laboratory bioassays revealed that larvae hosting HaDNV-1 had significantly enhanced resistance to HaNPV (and lower viral loads), and that resistance to Bacillus thuringiensis (Bt) toxin was also higher at low doses. Laboratory assays indicated that the virus was mainly distributed in the fat body, and could be both horizontally- and vertically-transmitted, though the former occurred only at large challenge doses. Densovirus-positive individuals developed more quickly and had higher fecundity than uninfected insects. We found no evidence for a negative effect of HaDNV-1 infection on H. armigera fitness-related traits, strongly suggesting a mutualistic interaction between the cotton bollworm and its densovirus.
| The old world cotton bollworm, Helicoverpa armigera, is one of the most significant pests of crops throughout Asia, Europe, Africa and Australia. Herein, we report a novel densovirus (HaDNV-1) which was widely distributed in wild populations of H. armigera and was beneficial to its host by increasing larval and pupal development rates, female lifespan and fecundity, suggesting a mutualistic interaction between the cotton bollworm and HaDNV-1. The cotton bollworm is currently widely controlled by the biopesticides Bacillus thuringiensis (Bt) toxin and the baculovirus HaNPV. It is therefore important to estimate the risk that the symbiotic virus will negatively impact on the efficiency of these biopesticides. Field and laboratory results suggest that HaDNV-1 infection significantly increases larval resistance to HaNPV and Bt toxin. These results have important implications for the selection of biopesticides for this species, and highlight the need for greater research into the elegant microbial interactions that may impact host individual and population dynamics.
| The interactions between symbiotic species and their hosts are becoming increasingly understood within insect systems [1], [2], [3]. Symbionts form diverse evolutionary relationships that influence the life history of their host, from mutualistic, by protecting them from natural enemies or increasing their host's fitness though a variety of means [1], [4], [5], [6], [7], [8], to parasitic, either by decreasing their resistance to harmful microorganisms or their tolerance to environmentally harmful factors, or by killing them directly [9], [10], [11]. There is a growing literature on the mutualistic interactions between intracellular bacterial symbionts, such as Wolbachia and their insect hosts, in which the symbionts spread through the host population by increasing the fitness of infected hosts [1], [6], [12], [13]. However, viral mutualistic symbioses have rarely been reported. This may be because, as obligate symbionts, viruses have long been considered harmful to their host and are usually isolated from cadavers killed by the virus. Moreover, until relatively recently, laboratory techniques only had the capacity to shed light on overtly pathogenic viruses, and not covert beneficial ones [14], [15], [16]. The development of molecular and sequencing technology facilitates the discovery and analysis of non-pathogenic virus species, using techniques such as suppression subtractive hybridization (SSH) and RNA-seq [17], [18]. Generally, viruses isolated from healthy individuals may be conditionally beneficial to their hosts. Recently, these ‘good viruses’ have attracted more attention, largely due to the prospect of using them in applications such as gene therapy and as tools for gene manipulation [2], [19]. As defined by Roossinck, there are few examples of viral mutualistic symbioses in insects (identified as conveying benefit to the host without any detectable fitness costs) [2].
The cotton bollworm moth, Helicoverpa armigera, is a major migratory pest of cotton and other economically-important crops throughout Asia, Africa, Europe and Australasia [20], [21], [22]. In China, the introduction of Bt-cotton in the 1990s has seen a dramatic decline in the H. armigera moth population. However, there are signs of Bt-resistance emerging [23], [24], fueling renewed interest in other forms of biological pest control, including the use of host-specific viral pesticides, derived from densoviruses [25], small RNA viruses [26] and baculoviruses [27], [28], [29], [30], [31]. Previously, we reported a novel densovirus (HaDNV-1, from the family Parvoviridae) in H. armigera moths that possesses a monosense genome that is 4926 nucleotides in length and clustered with the members of the genus Iteravirus in phylogenetic analysis [32]. This has allowed further investigation into the interactions between HaDNV-1 and its host H. armigera, which we report here. The main objective of this study was to establish the ecological significance of this virus within the migratory H. armigera system. Specifically, we undertook experiments to determine the transmission strategies of HaDNV-1, the impact of HaDNV-1 infection on host fitness, including its capacity to modulate resistance to potentially lethal biopesticides, and the prevalence of HaDNV-1 in field populations of H. armigera. Our results show that HaDNV-1 can be both horizontally- and vertically-transmitted in H. armigera; that HaDNV-1 infection increases host-fitness by increasing larval/pupal development rate, female lifespan and egg/offspring production; and that it also enhances larval resistance to H. armigera nucleopolyhedrovirus (HaNPV), a widely-used biopesticide. Resistance to Bt Cry1Ac protoxin was also enhanced, but only at relatively low toxin concentrations. Overall, we found no evidence for a negative effect of densovirus infection on H. armigera fitness-related traits, strongly suggesting a mutualistic interaction between the cotton bollworm and HaDNV-1.
To establish the modes of transmission of the densovirus HaDNV-1, we first produced an uninfected laboratory colony from a single breeding pair of H. armigera (NONINF strain). An infected strain (INF strain) was subsequently produced using neonate larvae from the NONINF strain, dosing them with either purified HaDNV-1 (108/µl; method 1, see Materials and Methods) or filtered liquid from infected individuals (108/µl; method 2, see Materials and Methods). Thus, our results indicated that HaDNV-1 could efficiently infect larvae by oral ingestion. The efficiency of infection with filtered liquid was higher than that of the purified virus (Table 1, Fig. S1A, S1B), suggesting that the purification process might have inactivated the virus in some way. We also found that individuals artificially infected with HaDNV-1 via peroral infection could efficiently transmit the viral infection to their offspring (Fig. S1C), and the same was true for naturally infected individuals (Fig. S1D), suggesting vertical transmission of the virus.
HaDNV-1 was capable of being vertically-transmitted from both infected females and infected males, but transmission-efficiency was higher from infected females than males (Table 1, Fig. S1E, S1F, S1G). With qPCR, we tested whether vertical transmission of HaDNV-1 was due to virus contamination on the surface of the eggs (transovum), or whether the virus was transmitted within the egg itself (transovarial). HaDNV-1 titers were not significantly different between sodium hypochlorite-treated and non-treated eggs (t = 1.296, d.f. = 6, P = 0.24) (Fig. 1), suggesting that transovarial transmission was occurring.
To examine the possibility of horizontal transmission through ingestion of contaminated foodplant (as would be a possibility in wild populations), we placed uninfected neonate larvae in diet cells that had previously housed infected insects (n = 8). Our results indicated that horizontal virus transmission did not occur in this manner, despite our previous experimental evidence that larvae could be orally infected. To examine this further, we used a range of HaDNV-1 concentrations to infect larvae and subsequently examined virus intensity in host frass (faeces). As expected, larval infection rate was positively related to the magnitude of the HaDNV-1 oral challenge, with low infection rates at doses less than 106/µl (Table 2); but even for larvae challenged with large viral doses, their frass contained only very low levels of HaDNV-1, with only 3 out of 20 samples containing more than 1×105/mg and none with more than 5×105/mg. Therefore, while we cannot exclude the possibility that horizontal transmission of HaDNV-1 may occur via the oral-fecal route, the viral levels in frass were very low and may not be sufficient for oral infection.
HaDNV-1 distribution was quantified within different host body tissues using qPCR. In both larvae and adults, HaDNV-1 titers were significantly higher in the fat body than in all other tissues: larvae: F = 11.098, d.f. = 5,36, P<0.0001 (Fig. 2A); adult females: F = 26.601, d.f. = 5,30, P<0.0001 (Fig. 2B); adult males: F = 44.560, d.f. = 5,30, P<0.0001 (Fig. 2C).
Using H. armigera as a control, we tested four other species of lepidopterans for their potential to act as alternative hosts for HaDNV-1, by attempting oral inoculation in Spodoptera exigua, Spodoptera litura, Agrotis segetum and Agrotis ipsilon. Results indicated that while oral inoculation with HaDNV-1 could successfully infect H. armigera, none of the four other species tested positive (Fig. S2A). We also tested field-captured adults of the closely-related species H. assulta but failed to find any HaDNV-1 positive individuals (n = 9; Fig. S2B). Based on these available data, it appears that infection with HaDNV-1 is host-specific to H. armigera.
To quantify the impact of HaDNV-1 infection on H. armigera development, a number of bioassays were performed using neonate larvae orally inoculated with filtered liquid from either HaDNV-1 infected (DNV+) or non-infected (DNV−) individuals (Fig. S3A, S3B). Both male and female DNV+ individuals developed significantly more quickly than the control individuals in both the larval (female: t = 2.732, d.f. = 312, P = 0.0067, male: t = 4.147, d.f. = 379, P<0.001) (Fig. 3A) and pupal stages (female: t = 5.100, d.f. = 312, P<0.001, male: t = 4.057, d.f. = 379, P<0.001) (Fig. 4A). Between 7–11 days post-hatch (approximately 3rd–5th instar) DNV+ larvae weighed significantly more than DNV- larvae by an average of ∼20% (GLMM with larval identity as a random term and log10-transformed larval weight as the dependent variable: Age (days): F = 2386.8, d.f. = 1,127, P<0.0001; HaDNV-1 infection status (+ve or −ve): F = 27.25, d.f. = 1,36, P<0.0001) (Fig. 3B, Fig. S4). However, their growth rates over this period did not differ, suggesting that densovirus effects on larval growth rate occurred prior to day 7 post-hatch (GLMM: interaction between infection status and age: F = 0.01, d.f. = 1,126, P = 0.91) (Fig. S4). A chloroform-wash assay indicated that at 9 days old, DNV+ larvae contained more lipid than DNV− individuals, measured as either lipid mass (t = 2.045, d.f. = 50, P = 0.046) or as a percentage of the whole body (t = 2.342, d.f. = 50, P = 0.023) (Fig. 3C, 4D). Larval mortality of DNV+ was significantly lower than DNV− (Table 3). However, there was no significant difference in pupal weight between DNV+ and DNV− insects (GLM: densovirus infection status: F = 0.99, d.f. = 1,692, P = 0.329; Sex: F = 41.08, d.f. = 1,693, P<0.0001; interaction term: F = 0.064, d.f. = 1,691, P = 0.80; female: t = 0.96, d.f. = 312, P = 0.34, male: t = 0.481, d.f. = 379, P = 0.63) (Fig. 4B), or pupation rate or eclosion rate between HaDNV-1 positive and HaDNV-1 negative insects (Table 3).
To determine the effect of HaDNV-1 infection on adult life-history traits, we used individuals from the non-infected (NONINF) and infected (INF) strains; and their infection status was confirmed by PCR (Fig. S3C, S3D). Infected INF strain moths produced significantly more eggs (t = 2.172, d.f. = 93, P = 0.032; Fig. 4C) and more neonates (t = 3.026, d.f. = 93, P = 0.0032; Fig. 4D) than individuals from the uninfected NONINF strain. Egg viability (hatch-rate) was significantly higher in the INF strain than in the NONINF strain (Table 3). The life-span of densovirus-infected females was significantly longer than that of females that were virus-free (χ21 = 13.5, d.f. = 1, P = 0.0002; Fig. 4E), but the longevity of males was not significantly different between the two strains (χ2 = 1.64, d.f. = 1, P = 0.2; Fig. 4E).
In larval field-collections, there was a non-random association between the two viruses (Chi-square test with Yates' correction: χ2 = 35.63, d.f. = 1, P<0.0001). Thus, there were relatively fewer larvae infected with both HaDNV-1 and HaNPV than would be expected by chance alone (14% versus 20%). When split by year, this effect was significant in 2012, when the overall HaNPV prevalence was 61% (χ2 = 19.75, d.f. = 1, P<0.0001; proportion infected with both viruses = 20% observed versus 26% expected), but not in 2013, when HaNPV prevalence was just 4% (χ2 = 0.82, d.f. = 1, P = 0.36; 2% observed versus 2% expected) (Table S1).
In adult field-collections, the prevalence of HaDNV-1 infection was uniformly high each year between 2008 and 2012 - 87%, 81%, 77%, 68% and 67%, respectively (Fig. S5). However, there was evidence for a significant decline in densovirus prevalence over the five years (GLMM with location as a random effect: χ21 = 39.06, P<0.0001). Despite high levels of baculovirus being observed in the larval field populations, we failed to detect any HaNPV-positive individuals in a random selection adult moths collected from four geographically diverse sites (n = 361 samples).
To determine the interaction between the densovirus HaDNV-1 and the baculovirus HaNPV, we first confirmed individuals from NONINF strain were NPV-free using PCR with specific primers. Then, NONINF strain neonates were inoculated with either HaDNV-1 (DNV+) or water (DNV− controls), and infections verified using PCR. Survival to pupation in larvae not exposed to HaNPV did not differ between DNV+ (95%) and DNV− (92%) larvae (χ2 = 0.27, d.f. = 1, P = 0.60). However, for those larvae exposed to the baculovirus, there was a significant difference between DNV+ and DNV− larvae in their susceptibility to HaNPV (GLM: HaDNV-1 infection-status: χ2 = 4.04, d.f. = 1, P = 0.044, parameter estimate ± standard error = 0.4645±0.2319), with densovirus-infected larvae suffering lower mortality rates for a given virus dose (GLM: log10 virus dose: χ21 = 98.56, P<0.0001; LC50s = 3.13×107 versus 9.10×107 OB per ml, for DNV− and DNV+ larvae, respectively; Fig. 5A); the interaction between viral dose and infection status was marginally non-significant (dose*status: χ2 = 3.72, d.f. = 1, P = 0.054).
We tested the differences of HaNPV replication between HaDNV-1 positive and negative individuals by repeating the HaNPV bioassay with 108 OBs/ml. The baculovirus bioassay indicated that there was no HaNPV-induced mortality in the control larvae that were exposed to water only, and that most mortality in the HaNPV-challenged larvae started at day 5 (120 h post-inoculation) (Fig. 5B). In NPV-challenged larvae, those carrying HaDNV-1 suffered significantly lower mortality overall than HaDNV-1 negative insects (Likelihood-ratio test: χ2 = 23.24, d.f. = 1, P<0.0001; linear coefficient (95% confidence interval) = 0.248 (0.134, 0.457)). Therefore, we collected samples before day 5 post-challenge to estimate HaNPV viral loads using qPCR. As would be expected, HaNPV titers (log-transformed) increased over time post-challenge and the rate of HaNPV titer increase was lower for HaDNV-1 positive larvae than in larvae lacking HaDNV-1, as indicated by a significant interaction term (linear model: Time post-challenge: F = 27.02, d.f. = 1,112, P<0.0001; DNV infection status: F = 5.69, d.f. = 1,112, P = 0.019; Time* DNV status interaction: F = 8.69, d.f. = 1,112, P = 0.0038; Fig. 5C). However, HaNPV titers were not directly correlated with HaDNV-1 titers in HaDNV-1 positive individuals (r = 0.066, n = 58, P = 0.623). These results suggest that HaDNV-1 protected H. armigera from HaNPV, possibly by slowing the accumulation of HaNPV.
A similar bioassay using the Bt toxin Cry1Ac instead of the baculovirus generated consistent results. As expected, larval development score increased over time and declined with increasing Bt dose (linear mixed-effects model with larval identity as a random term: Day: F = 18147.38, d.f. = 1,4172, P<0.0001; Log2Btdose: F = 1335.48, d.f. = 1,4172, P<0.0001). However, development was also influenced by the interaction between DNV infection status and the dose of Bt administered (DNV status: F = 120.21, d.f. = 1,4172, P<0.0001; DNV status * Bt dose interaction: F = 111.81, d.f. = 1,4172, P<0.0001), with the enhanced development of HaDNV-1 positive larvae at low Bt concentrations declining as Bt dose increased, such that mean development rate was independent of DNV infection status as Bt concentrations above 1.6 µg/g (Fig. 6). We also performed the bioassay with Bt cotton. As expected, there was a significant effect of Bt cotton on larval development rate, with development being significantly stunted in larvae exposed to the Bt plants (linear model: Diet: F = 63.74, d.f. = 1,476, P<0.001; mean score ± s.e.: Bt cotton = 1.717±0.153; non-Bt cotton = 3.529±0.167). However, whilst DNV positive larvae tended to have slightly higher development scores than DNV negative larvae (2.754±0.176 versus 2.492±0.164), this difference was non-significant and the interaction between DNV status and Bt exposure was also non-significant (DNV status: F = 1.336, d.f. = 1,476, P = 0.24; DNV status * Diet interaction: F = 0.0084, d.f. = 1,476, P = 0.93).
To date, viral mutualistic symbioses have attracted little attention and are rarely reported, most likely due to a lack of obvious pathogenicity within their insect hosts. In our study system, SSH was previously used to detect and isolate a novel densovirus (HaDNV-1) from healthy migratory cotton bollworms, H. armigera [32]. To date, most reported DNVs have been pathogenic to their hosts, even resulting in mortality, and as a result DNVs have been considered as potential biological control agents of insect pests [33], [34], [35], [36]. However, in our present study, for the first time, we show a mutualistic relationship without any detectable negative interactions between a DNV and its host.
Although endosymbionts of insects do have the capacity for horizontal transmission, they are usually transmitted via maternal inheritance [1], [3]. However, viral symbionts can be efficiently transmitted both vertically and horizontally [37], [38], [39], [40], [41], [42], [43], [44]. We found that HaDNV-1 was efficiently vertically-transmitted via both the paternal and maternal lines. This was most likely via transovarial infection, with the efficiency of transmission being higher from infected females than males. The results presented here also suggest that HaDNV-1 can be horizontally-transmitted to H. armigera by peroral infection of larvae, in a dose-dependent manner. However, we failed to detect horizontal transmission by diet contamination, suggesting that although larvae can be infected orally, peroral infection may only be possible at very high HaDNV-1 concentrations. Indeed, infection rate and intensity were both positively correlated with the magnitude of the HaDNV-1 challenge, and the frass of larvae contained only very low levels of HaDNV-1. This suggests that in the field, HaDNV-1 is likely to be almost exclusively transmitted vertically from parents to offspring. Previous studies suggest that DNVs may vary in their host ranges, for example Junonia coenia densovirus (JcDNV), Mythimna loreyi densovirus (MlDNV) and Periplaneta fuliginosa densovirus (PfDNV), all infect several host species, whereas Galleria mellonella densovirus (GmDNV) infects only one species [33]. Our results suggest that HaDNV-1 is also strongly host-specific following oral exposure, only infecting H. armigera.
Certain bacterial beneficial symbionts have been reported to benefit their hosts by shortening host development time and increasing host fecundity [1], [45]. However, evidence of viruses increasing host fecundity has rarely been reported. One exception is in a vector-virus complex in the whitefly Bemisia tabaci: a plant virus transmitted by B. tabaci was found to accelerate the population growth rate of its insect host [46]. In our system, HaDNV-1 infection intensity was greatest in the host fat body, suggesting that the virus might play a role in the development of H. armigera. Indeed, the significantly shortened development time and faster growth rate of H. armigera infected with HaDNV-1 could be mediated by the virus promoting the accumulation of fat body by the host. Our results showed that at 9 days old, HaDNV-1-infected larvae contained more lipid than uninfected larvae. The positive effect of the HaDNV-1 on these life-history traits, including egg/offspring production, suggests a possible mutualistic relationship. Taken together with the results of the baculovirus bioassay, these results suggest that HaDNV-1 benefits H. armigera, but is not an obligate microbe required by the host to survive.
The baculovirus HaNPV is a large double-stranded DNA virus, which was first isolated in China in 1975 and has since become an important biopesticide for a number of agricultural pests [27], [28], [29], [30], [31]. To determine the interaction between HaDNV-1 and HaNPV in H. armigera, we collected samples of larvae and adults from the field to determine the natural infection rates of HaDNV-1 and HaNPV. Most significantly, we found that there was a clear negative interaction between the two viruses across larval populations, with there being more insects infected with one or other of the viruses than would be expected by chance alone, and fewer with both viruses or neither. One possible explanation for this observation is that there is a negative interaction between the two viruses: perhaps HaDNV-1 increases susceptibility to HaNPV disease, resulting in those individuals with both viruses being more likely to die, as seen in larvae of the African armyworm moth, Spodoptera exempta, co-infected with Wolbachia and the baculovirus SpexNPV [9]. However, our results from the HaNPV-HaDNV-1 bioassay suggest the opposite, with HaDNV-1 infected larvae being significantly more resistant to HaNPV than those not carrying the densovirus. Therefore, it is likely that fewer than expected HaNPV-HaDNV-1 co-infected individuals were detected in field populations because HaDNV-1 protects its host against HaNPV infection. Our qPCR assay supported this hypothesis: HaNPV was found to accumulate in HaDNV-1 infected larvae at a slower rate than in uninfected larvae. Another possibility to explain the dearth of co-infected individuals is that rather than there being a direct interaction between the viruses, the interaction is indirect. Baculoviruses only infect the larval stages of Lepidoptera and early larval instars are generally more susceptible to viral infection (via oral ingestion) than older larvae, possibly because they slough virus-infected midgut cells at a slower rate [47]. If a larva can grow more quickly than its peers in the same cohort, then it will be less susceptible to virus infection and potentially “escape” disease (via this developmental resistance mechanism). Consistent with this, we found that HaDNV-1-positive larvae developed faster than HaDNV-1-negative larvae (Fig. 3) and accumulated HaNPV at a slower rate (Fig. 5C). In field populations of adults, the infection rate of HaDNV-1 remained high from 2008 to 2012 (more than 67%). However, we failed to detect any HaNPV baculovirus in any of the 361 adults sampled. Only the larval stage is susceptible to baculovirus infection and so one possible explanation for this is that most of the baculovirus-infected individuals are lost from the system before adulthood due to increased larval mortality, abnormal pupation, or unsuccessful eclosion [48]. Alternatively, enhanced resistance to HaNPV in the adult stage may effectively clear all viral infections gained in the larval stage.
Theory suggests that the presence of a beneficial symbiont should result in a high frequency of infection, spreading rapidly through a population until reaching infection fixation [1]. However, our data from adult moths suggest that although there was a high frequency of HaDNV-1 infection, there was also, perhaps unexpectedly, a steady decline in prevalence from 2008 to 2012, which would suggest an unidentified cost of DNV infection. One possible explanation for this decline is that the prevalence of HaDNV-1 is related to the recent widespread introduction to China of genetically-modified Bacillus thuringiensis (Bt) cotton [49], [50], [51]. For example, it might be that selection for Bt-resistance has selected against densovirus infection. If this was the case, then we might expect to observe a negative association between HaDNV-1 infection and resistance to Bt. However, in our laboratory experiment with Bt protoxin and artificial diet, HaDNV-1-positive larvae showed significantly higher resistance to Bt than HaDNV-1-negative larvae at low Bt concentrations (≤0.8 µg/g), while no significant difference was observed at high Bt concentrations (≥1.6 µg/g). Interestingly, the bioassay with Bt cotton plants showed that although HaDNV-1 positive larvae developed faster than negative ones, the difference was not statistically significant, possibly because the leaves of the Bt cotton used (at the seedling stage) contained a high concentration of Bt protein (about 1 µg/g) [52]. A related possibility is that densovirus prevalence is positively associated with the size of the H. armigera population in the wild, which has markedly declined since Bt-cotton was introduced [51], perhaps because horizontal transmission of the densovirus is enhanced at high population densities. The possibility of unknown competitive factors, including other microorganisms, can also not be excluded. Therefore, despite some evidence suggesting that HaDNV-1 could impact the population dynamics of H. armigera, our data are currently not comprehensive enough to explain the long-term dynamics of HaDNV-1, and more monitoring of field populations will be required to answer some of these intriguing questions.
In conclusion, our studies to date suggest a mutualistic relationship between the cotton bollworm and HaDNV-1, in which the cotton bollworm appears to benefit from HaDNV-1 infection, with all host fitness parameters so far tested (larval growth rate, larval and pupal development rate, fertility, adult female lifespan, and resistance to baculovirus and low doses of Bt toxin) enhanced at no detectable cost. The study of beneficial viruses in both vertebrate and invertebrate systems has only relatively recently attracted researchers' attention [2], predominantly due to the explosion of new technologies that now make the detection of such organisms possible. It should be noted that the coevolution between viral mutualistic symbionts and their hosts could be an important factor to consider when studying the adaptability of insect host species. Illuminating the function of such viral symbionts may offer novel insights for future pest management strategies.
Cotton bollworms (H. armigera) were reared using artificial diet [53] at 25±1°C with a 14:10, light:dark photoperiod. Adult moths were provided with 10% sugar and 2% vitamin complex. The colony was established from thirty breeding pairs captured at Langfang (Hebei province, China) in 2005. Individuals successfully producing offspring were tested for the presence of HaDNV-1, using the methods described below. Offspring from a single uninfected breeding pair were reared to produce the NONINF strain (uninfected) laboratory culture.
HaDNV-1 virus was isolated from migrating H. armigera adults captured in 2010 and 2011 using a vertical-pointing trap, and stored in liquid nitrogen [20]. Briefly, DNA was extracted from host tissues (except for the abdomens) of each individual, and PCR undertaken to detect the presence of HaDNV-1. Subsequently, the abdomens of positive individuals were divided into two groups: one group was used to purify the HaDNV-1 using the method described by La Fauce et al. (method 1) [54]; the other group was used to prepare a filtered liquid, containing an unpurified form of virus (method 2). Briefly, this second method involved grinding four abdomens under liquid nitrogen and transferring to 1 ml PBS buffer (0.01M, pH 7.4). The homogenate was centrifuged at 6500×g for 15 min at 4°C, and the liquid supernatant subsequently filtered with Sartorius Minisart 0.2 µm PES (Invitrogen, Grand Island, USA). The abdomens of negative individuals were filtered using the same method. Quantification of the viruses was performed using the qPCR method described below. All the samples were stored at −20°C.
To detect the existence of HaDNV-1 in H. armigera, specific primers amplifying a 496 bp fragment, DVVPF/DVVPR (Table S2) were designed according to the genomic sequence of HaDNV-1. The PCR program was as follows: 30 s at 94°C, 30 s at 55°C, and 30 s at 72°C for 40 cycles. For detection of H. armigera nucleopolyhedrovirus (HaNPV), a pair of specific primers amplifying a fragment of 445 bp, NPVF/NPVR, were designed according to the open reading frame 14 (ORF14) of the genomic sequence of HaNPV. The PCR program was as follows: 30 s at 94°C, 30 s at 57°C, and 30 s at 72°C for 40 cycles.
For quantifying the copy numbers of HaDNV-1 and HaNPV, an absolute quantification qPCR methodology using a standard curve was performed [55]. Fragments containing the primers and probes of HaDNV-1 and HaNPV were amplified with our de novo primers (PF/PR for HaDNV-1, NPVF/NPVR for HaNPV) using the program: 30 s at 94°C, 30 s at 53°C, and 60 s at 72°C for 40 cycles, and cloned into the pEASY-T Cloning Vector (TransGen, Beijing, China). These plasmids were subsequently used for the quantification standard curve assay. qPCR was carried out with the TaqMan method in 20 µl reaction agent comprised of 1 µl of template DNA, 2×Premix Ex Taq (Takara, Japan), 0.2 µM of each primer and 0.4 µM probe, using a 7500 Fast Real-time PCR System (Applied Biosystems). Thermal cycling conditions were: 45 cycles of 95°C for 15 s, 60°C for 34 s. The DNA sample of each group was replicated three times. All primers used in this study were shown in Table S2. The equation of y = −1.052x+42.327 (y = the logarithm of plasmid copy number to base 2, x = Ct value, R2 = 0.9997) and y = −0.9861x+44.647 (y = the logarithm of plasmid copy number to base 2, x = Ct value, R2 = 0.9999) were used to calculate the copy number of HaDNV-1 and HaNPV, respectively.
We constructed an infected line (INF strain) of H. armigera by orally infecting NONINF strain larvae with HaDNV-1 (from filtered liquid, method 2 - see above) and maintained them by vertical transmission of the virus, using the primers DVVPF/DVVPR to confirm successful establishment of HaDNV-1 infection. Subsequently, individuals from both NONINF strain and INF strain were used to determine the transmission modes of HaDNV-1. For vertical transmission, ♀+/♂−, ♀−/♂+, ♀+/♂+ and ♀−/♂− pairs were crossed and DNA from 3rd instar offspring larvae used to probe for HaDNV-1.
For the diet contamination assay, (to determine horizontal transmission efficiency), infected individuals from the INF strain were reared in diet cells until the start of the 3rd instar and then removed. Uninfected NONINF strain neonates were then placed in the vacated cells and reared to the pupal stage. DNA was extracted from the adults and probed for HaDNV-1 infection using PCR. Horizontal transmission of HaDNV-1 was determined using PCR with adult DNA as temples and different concentrations of the densovirus: 108, 107, 106, 105, 104/µl. The frass of larvae from HaDNV-1 positive individuals were also quantified by qPCR, as described above.
To examine virus infection in different body tissues, DNA was extracted from body parts of infected individuals (both larval and adult stages) and the copy numbers of HaDNV-1 were quantified by qPCR. To account for individual variation, we first calculated the copy numbers per milligram of tissue and then summed all the copy numbers from different tissues from the same individual and the percentage of each tissue was statistically analyzed (larvae: n = 7; adult males: n = 6; adult females: n = 6).
To further establish the role of vertical transmission in the life-cycle of the densovirus, we quantified HaDNV-1 infections in H. armigera eggs, primarily to distinguish between transovarial and transovum infection routes. Eggs from INF strain breeding pairs, which both of females and males were infected by HaDNV-1, were submerged in 1% sodium hypochlorite for 10 minutes. They were then filtered through a damp cloth, thoroughly rinsed, and allowed to dry. Four groups of hypochlorite-treated eggs (n = 50 eggs per group) were tested against non-treated eggs (control) and HaDNV-1 infections tested by qPCR.
To test the impact of HaDNV-1 infection on the life table parameters of its host, neonate NONINF strain larvae were first orally inoculated with either filtered-liquid containing HaDNV-1, or filtered-liquid from uninfected individuals (control). One hundred NONINF strain neonates were placed in each treatment Petri-dish for 2 days to ensure that larvae ingested the treated diet. They were then transferred to a 24-well plate (one individual per well: diameter = 1.5 cm; height = 2 cm) until the 5th larval instar; larvae were then individually reared in glass tubes until eclosion (diameter = 2 cm; height = 7.5 cm) (Fig. S6). The status of individuals was checked every day at 9:00 am. The weight of larvae from the 7th to 11th day post hatch, and the pupa on the 3rd day were recorded. Fifth-instar larvae were randomly selected to estimate the infection rate of HaDNV-1 during the experiment. This bioassay was replicated twice (n = 288 and n = 168 individuals, respectively). Individuals dying within 24 hours of the experimental set up were considered handling deaths, and excluded from the analysis.
In addition, newly eclosed adults from both the HaDNV-1 negative NONINF strain and HaDNV-1 positive INF strain were used to determine longevity, egg production and hatch rate. Three pairs of adults were put in each plastic cup (diameter = 8.5 cm; height = 10 cm) (Fig. S6). The experimental replicates were 3×77 for NONINF strain and 3×60 for INF strain, respectively. We recorded the number of eggs and newly hatched larvae every day. After death, individuals were used to detect HaDNV-1 via PCR. Data from failed matings were excluded.
To quantify the impact of HaDNV-1 infection on host growth, we measured relative lipid mass within larvae of H. armigera. Larvae 9 days post-hatch were chosen to compare the lipid content between HaDNV-1 positive (n = 19) and HaDNV-1 negative (n = 33) individuals. The protocol was undertaken as Clissold et al. [56]. Briefly, the larval samples were freeze-dried, weighed, chloroform-extracted 3 times, dried again and weighed. The lipid mass was calculated by subtracting the post-chloroform-wash mass from the pre-chloroform-wash mass.
To assess the capacity of HaDNV-1 to act as a beneficial symbiont, we quantified the interaction between HaDNV-1 and the common baculovirus pathogen HaNPV, via a series of laboratory bioassay studies. As previously described, neonate larvae were first treated with HaDNV-1 filtered liquid (either from HaDNV-1 infected or HaDNV-1 negative individuals). Two-day old larvae were then transferred to a 24-well plate and maintained on diet until the 9th day after hatching. Individuals weighing between 5–11 mg (early third-instar stage) were chosen for the HaNPV bioassay. Purified powder of HaNPV at a concentration of 5×1011 occlusion bodies (OBs) per g was generously provided by Dr. Qilian Qin in the Institute of Zoology, Chinese Academy of Science, Beijing, China. Larvae were orally dosed with 4 treatments of HaNPV (30 larvae per treatment at: 0 (control), 1×106, 1×107, 1×108, and 1×109 OBs/ml). Only larvae that ingested all the NPV within a 24 h period were used for the bioassay. Larvae were subsequently monitored daily for NPV mortality until pupation, and all viral deaths stored at −20°C. PCR with specific primers was used to test for NPV in dead larvae with non-obvious symptoms.
To assess HaNPV infection levels in HaDNV-1 positive and negative individuals, we performed a separate HaNPV bioassay with 108 OBs/ml. There were 24 individuals in each replicate and three replicates per treatment. Only larvae that ingested all the NPV within a 24 h period were used for the bioassay. The absolute quantification qPCR methodology was used to quantifying the copy numbers of HaNPV as described above. Survival analysis was conducted using Cox's proportional hazards model.
For the Bacillus thuringiensis bioassays, various concentrations of the Bt Cry1Ac protoxin were added and thoroughly mixed with standard artificial diet to obtain the desired concentrations (0 (control), 0.4 µg/g, 0.8 µg/g, 1.6 µg/g and 3.2 µg/g). After mixing, the diet solidified and solid 1 mg pieces were placed into each well of a 24-well plate and two-day old larvae infected or uninfected by HaDNV-1 were then transferred to each well (Fig. S6). There were 24 individuals in each replicate and three replicates per treatment. We graded the larvae from day 4 to day 9 after hatching according to the development rate: death = 0, early first instar stage = 1, middle first instar stage = 2, last first instar stage = 3, early second instar stage = 4, middle second instar stage = 5, last second instar stage = 6, early third instar stage = 7, middle third instar stage = 8, last third instar stage = 9, early fourth instar stage = 10, middle fourth instar stage = 11 [57].
At seedling stage with 5 leaves, we chose the new cotton 33B with Cry1Ac (Monsanto Company, Bt cotton) using Shi Yuan 321 (Shijiazhuang Acadamy of Agricultural Sciences, NonBt cotton) as control to perform the bioassay. Two-day old larvae infected or uninfected by HaDNV-1 were transferred to a 24-well plate with Bt-cotton or NonBt-cotton. There were 40 individuals in each replicate and three replicates per treatment. We graded the larvae after 7 days according to the development rate.
Samples of larvae were collected at 7 locations in 2012 (Jinan, Dezhou and Taian, Shandong province; Cangzhou, Heibei province; Tianmen and Qianjing, Hubei province; Maanshan, Anhui province) and 6 locations in 2013 (Luohe, Luoyang, Yuanyang and Nanyang, Henan province; Langfang and Cangzhou, Hebei province). The infection rate of HaDNV-1 and HaNPV was determined using the PCR method described as above.
Samples of adults were collected at fifteen locations from 2008 to 2012: A = Xinxiang, Henan province; B = Dezhou; C = Langfang; D = Yantai Shandong province; E = Yancheng, Jiangsu province; F = Handan, Shandong province; G = Changde, Hunan province; H = Tianmen, I = Qianjiang; J = Maanshan; K = Taian; L = Luohe; M = Weinan, Shanxi province; N = Shihezi, O = Kashi, Xinjiang province. We also randomly selected four places to detect HaNPV in the populations, including site 1 in 2010 (54 samples), site 2 in 2010 (103 samples), site 4 in 2012 (104 samples) and site 13 in 2011 (100 samples).
Using the same oral inoculation method as previously described (section 2.5), we chose four species of Lepidoptera (Spodoptera exigua, Spodoptera litura, Agrotis segetum, Agrotis ipsilon) to determine the host range of HaDNV-1 infection. We also collected nine adults of H. assulta from field populations, and PCR was used to detect HaDNV-1 infection.
Statistical analyses were conducted using STATA v.9.0 and R v3.0.1 [58]. Student's t-test or ANOVA with Tukey were used to determine the level of significance in the relative levels of HaDNV-1. Egg hatch rates and larval/pupal mortality, pupation and eclosion rates were determined using generalized linear models (GLMs) with binomial errors. Analysis of the NPV and Bt bioassay data was also conducted using GLMs with binomial errors. A generalised linear mixed effects model (GLMM) with binomial errors was used to determine temporal variation in HaDNV-1 infection rates. A GLMM with Gaussian errors was used to quantify variation in larval growth rates with larval identity included as a random term. Development following exposure to Bt toxin in artificial diet was analyzed using linear mixed effects models using the lme function in R, with larval identity as a random term to account for the repeated measures data structure.
The GenBank accession number of genomic sequence of HaDNV-1 and HaNPV were HQ613271 and AF303045, respectively.
|
10.1371/journal.pntd.0007127 | Community-level chlamydial serology for assessing trachoma elimination in trachoma-endemic Niger | Program decision-making for trachoma elimination currently relies on conjunctival clinical signs. Antibody tests may provide additional information on the epidemiology of trachoma, particularly in regions where it is disappearing or elimination targets have been met.
A cluster-randomized trial of mass azithromycin distribution strategies for trachoma elimination was conducted over three years in a mesoendemic region of Niger. Dried blood spots were collected from a random sample of children aged 1–5 years in each of 24 study communities at 36 months after initiation of the intervention. A multiplex bead assay was used to test for antibodies to two Chlamydia trachomatis antigens, Pgp3 and CT694. We compared seropositivity to either antigen to clinical signs of active trachoma (trachomatous inflammation—follicular [TF] and trachomatous inflammation—intense [TI]) at the individual and cluster level, and to ocular chlamydia prevalence at the community level.
Of 988 children with antibody data, TF prevalence was 7.8% (95% CI 6.1 to 9.5) and TI prevalence was 1.6% (95% CI 0.9 to 2.6). The overall prevalence of antibody positivity to Pgp3 was 27.2% (95% CI 24.5 to 30), and to CT694 was 23.7% (95% CI 21 to 26.2). Ocular chlamydia infection prevalence was 5.2% (95% CI 2.8 to 7.6). Seropositivity to Pgp3 and/or CT694 was significantly associated with TF at the individual and community level and with ocular chlamydia infection and TI at the community level. Older children were more likely to be seropositive than younger children.
Seropositivity to Pgp3 and CT694 correlates with clinical signs and ocular chlamydia infection in a mesoendemic region of Niger.
ClinicalTrials.gov NCT00792922.
| Trachoma programs currently use the clinical sign of trachomatous inflammation-follicular (TF) to guide community treatment decisions and evaluate response to mass drug administration with azithromycin. These programs rely on clinical grading that poorly correlates with infection with the causative agent of trachoma, Chlamydia trachomatis (Ct), in low prevalence areas. Serologic measures of Ct may provide additional information about exposure and transmission patterns. Here, we evaluated the relationship between serologic markers of Ct, infection, and TF at the individual and community levels to evaluate the utility of serology for measuring trachoma in a mesoendemic region of Niger. We found that serologic markers correlated with both infection and TF, indicating that inclusion of serologic markers may be useful to guide trachoma decision making.
| Trachoma, caused by repeated ocular infection with Chlamydia trachomatis (Ct), has been targeted by the World Health Organization (WHO) for elimination as a public health problem by the year 2020. As part of the strategy to achieve elimination, WHO recommends annual mass drug administration (MDA) of azithromycin in endemic districts [1]. Program targets related to MDA focus on the district-level prevalence of trachomatous inflammation—follicular (TF) amongst children aged 1–9 years. To monitor progress towards elimination, population-based impact surveys are recommended to evaluate whether a district has reached the threshold of less than 5% TF prevalence in 1–9-year-olds and can cease azithromycin distribution. Two years after cessation of MDA, a surveillance survey to ensure that district-wide TF prevalence in 1–9-year-olds remains below 5% is conducted prior to the validation process. Currently, there are no guidelines for post-validation surveillance.
These surveys rely on a clinical grading scheme that is relatively inexpensive and simple to perform, but is poorly correlated with ocular Ct infection in low-prevalence settings [2]. Following MDA, the clinical sign trachomatous inflammation—intense (TI) has been shown to correlate better with infection than TF does [3]. However the measurement of clinical signs is subject to inter-grader variability and lack of real-time auditing since grading is performed in the field and thus can only later be validated or audited if images are taken. As trachoma elimination programs stand to benefit from an accurate, reproducible assessment of trachoma prevalence, other testing methods may be useful to help guide program decisions. These include tests of infection (polymerase chain reaction [PCR] testing of ocular swabs) and antibody-based testing [4–7].
Antibodies to Ct antigens may act as markers of cumulative exposure to Ct. Two previously described Ct antigens, Pgp3 and CT694, have been shown to be reactive against sera in young children living in trachoma-endemic communities [4,7,8]. At the individual level, antibodies to these proteins demonstrate high sensitivity to ocular infection and high specificity against non-endemic control specimens [8–10]. However, individual associations may not always hold at the community level, and trachoma elimination programs treat ocular Ct infection on a population level. Additionally, as antibody markers are not yet widely used to assess for Ct prevalence, better characterization of how seropositivity compares to other methods of assessing trachoma prevalence is necessary. Here, we evaluate the association between seropositivity, PCR positivity, and clinical signs of active trachoma (TF and TI) at the individual and community level in a region of Niger where some trachoma transmission is occurring (TF prevalence approximately 25% at baseline). Data were collected during the final follow-up visit of the Partnership for the Rapid Elimination of Trachoma (PRET)-Niger trial, in which communities were randomized to receive annual or biannual oral azithromycin for 3 years in order to assess the impact of treatment frequency on ocular chlamydia infection [11].
The study methods have been previously reported in detail elsewhere [11–13]. Briefly, a cluster randomized trial of annual versus biannual mass azithromycin distribution for trachoma control was conducted in the Matameye district of the Zinder region of Niger from May 2010 until August 2013 [4–6]. Data on active trachoma and ocular infection were collected biannually on children aged 0–5 years; dried blood spots for serological analysis were collected only at the 36-month time point and only from children aged 1–5 years. Dried blood spots were shipped to CDC at ambient temperature and tested for antibodies from July to August 2014.
Communities were chosen from among six different catchment areas for primary health care facilities and were eligible for inclusion if they met the following criteria: (1) contained a population between 250 to 600 persons, (2) were located more than 4 kilometers from the center of any semi-urban area, and (3) had a prevalence of active trachoma more than 10% in children aged 0–5 years [11]. 235 communities in the 6 health centers were deemed eligible, of which 48 were randomly selected for inclusion in the trial. Children aged 1–5 years were included in this analysis, due to the inability of antibody tests to differentiate between maternal-child antibodies in <1–year-olds.
48 communities were randomly divided into 4 treatment arms in a 2x2 factorial design (12 communities per arm), comparing two azithromycin coverage targets (standard versus enhanced coverage) and annual versus biannual treatment. Randomization of communities to treatment arms was done using RANDOM and SORT functions in Microsoft Excel (Version 2003). Only communities from the enhanced coverage arms were included in testing for antibodies (N = 24 communities) for logistical reasons.
Trained study health workers conducted a full household census in all communities prior to the initial survey visit. During the baseline visit, adults in the household consented to census data collection, and study personnel recorded the name, sex, and age or date of birth for all individuals in the household.
After consent was obtained, study participants were examined for the presence of TF and TI. Clinical grading of each everted superior tarsal conjunctiva was performed using a 2.5x binocular loupe and a torch light, if necessary, per the WHO grading system. Clinical grading was performed according to the WHO simplified grading system of TF being the presence 5 or more follicles >0.5 mm in diameter and TI as inflammation severe enough to obscure 50% of deep tarsal vessels in one or both eyes [14]. Prior to swabbing, a trained photographer took at least 2 photographs of the right eyelid of all participants using a Nikon D-series camera and a Micro Nikon 105 mm; f/2.8 lens (Nikon, Tokyo, Japan). After conjunctival examination, a Dacron swab was passed 3 times over the right upper tarsal conjunctiva, rotating the swab approximately 120 degrees between each pass. All samples were placed immediately on cold packs in the field and transferred to -20°C within 10 hours, then shipped on cold packs to University of California, San Francisco, CA, USA where they were stored in -80°C freezers until processing.
PCR testing was performed for children aged 0–5 years. Samples from the same village, age in years, and visit were randomly pooled into groups of five for group testing, with a possible remainder pool of one to four samples [15]. Pooled samples were tested for the presence of Ct DNA using the Roche Amplicor qualitative PCR assay (Roche Molecular Systems, Indianapolis, IN, USA). Community prevalence was estimated from the pools as previously described [11,15].
In communities randomized to annual treatment, study participants age 6 months of age and older received a directly observed dose of oral azithromycin (20 mg/kg up to a maximum dose of 1 g in adults). In biannually treated communities, only children up to 12 years of age were offered treatment. Children under 6 months of age in all communities were offered topical tetracycline ointment (1%) to be applied to both eyes twice a day for six weeks. Pregnant women in the annual arm and individuals allergic to macrolides were offered topical tetracycline. All communities were visited up to four days in order to achieve 90% treatment coverage [12,13].
Children under 5 years of age were selected randomly in each village for blood sample collection via finger stick or heel stick, with a goal of 50 children per village. Blood spots were analyzed for antibody to Ct antigens Pgp3 and CT694 using a multiplex bead array assay on a Luminex 200 platform, as previously described [7]. Results were reported as median fluorescence intensity minus background (MFI-BG) where background is the signal from beads run with buffer only. Positivity cut-off for Pgp3 was greater than or equal to 1083, and CT694 cutoff was greater than or equal to 496 as determined by receiver operator characteristic (ROC) curve analysis from a pediatric U.S.-based negative panel (N = 117) and Tanzania based positive panel from children with ocular Ct infection (N = 40) [7].
Data were entered into a customized database (Microsoft Access v2007) developed at the Dana Center, Johns Hopkins University. To estimate associations between seropositivity, clinical trachoma, and age at the individual level, we used generalized linear models with a binomial distribution and log link to estimate prevalence ratios (PR). All standard errors were clustered at the community level, which was the randomization unit of the study. As individual-level PCR data were not available, associations between seropositivity to the Ct antigen and ocular chlamydia infection were conducted only at the community level. We additionally analyzed the association between seropositivity and clinical trachoma at the cluster level. We used linear regression models to evaluate relationships between trachoma indicators at the community level. All analyses were conducted in Stata 14.1 (StataCorp, College Station, TX).
All procedures and protocols for this study were approved by the Committee for Human Research of UCSF, and le Comité Consultatif National d’Ethique du Minstère de la Santé Publique, Niger (Ethical Committee, Niger Ministry of Health). The study’s Data Safety and Monitoring Committee observed the study implementation during annual reviews of quality assurance, as appointed by the PRET study Executive Committee. All village leaders of communities within the study agreed to participate in the trial with written (thumbprint) consent. For children under the age of 16, consent was given by a parent or a guardian. All persons participating in the trial were given the opportunity to be treated according to their community’s random treatment assignment. Communities not included in the study were offered treatment through the national treatment program. CDC personnel did not have access to personal identifying information and were determined to be non-engaged in the study.
At the 36-month follow-up, 988 1–5-year-old children from 24 communities had clinical trachoma and serology data available. TF prevalence was 7.8% (95% CI 6.1 to 9.5%) and TI prevalence was 1.6% (95% CI 0.9 to 2.6%) (Table 1). The overall prevalence of antibody positivity to Pgp3 was 27.2% (95% CI 24.5 to 30%), and to CT694 was 23.7% (95% CI 21 to 26.2%). Community prevalence of ocular chlamydia in 0-5-year-olds was 5.2% (95% CI 2.8 to 7.6%).
Children with antibodies to either Pgp3 or CT694 were more likely to have TF (PR, 1.90, 95% CI 1.49 to 2.43, P<0.001) and also TI, although the latter relationship was not statistically significant (PR 1.65, 95% CI 0.99 to 2.75, P = 0.06).
Fig 1 shows the community-level prevalence of TF and ocular chlamydia compared to seropositivity to Pgp3 and/or CT694. At the community level, Pgp3 and/or CT694 seroprevalence was significantly correlated with ocular chlamydia infection (linear regression coefficient 0.19, 95% CI 0.08 to 0.29, P = 0.001), TF prevalence (linear regression coefficient 0.25, 95% CI 0.13 to 0.36, P<0.001), and TI prevalence (linear regression coefficient 0.07, 95% CI 0.004 to 0.14, P = 0.04).
The probability of having an antibody response to Pgp3 or CT694 increased with increasing age (PR 1.25 per one-year increase in age, 95% CI 1.17 to 1.35, P-trend<0.001; Fig 2). Older age was not significantly associated with a diagnosis of TF (PR 1.02 per year, 95% CI 0.89 to 1.16, P-trend = 0.82) or TI (PR 0.87, 95% CI 0.60 to 1.25, P-trend = 0.45).
There was no significant difference between study arms in the percentage of children antibody-positive to Pgp3 (PR 0.72 biannual versus annual, 95% CI 0.39 to 1.33, P = 0.29) or CT694 (PR 0.81 biannual versus annual, 95% CI 0.51 to 1.28, P = 0.36; Table 2).
The presence of antibodies to Ct antigens was correlated with both TF and ocular Ct infection following a 36-month annual or biannual mass azithromycin distribution program in a trachoma-endemic region of Niger. However, there was no difference in serologic outcomes by study arm, consistent with clinical data suggesting that biannual treatment did not significantly alter transmission of ocular chlamydia compared to annual treatment [11]. That serologic outcomes were consistent with other trachoma indicators supports the finding that antibodies to Ct antigens are correlated with TF and ocular Ct infection and provide complementary information. Future work evaluating serologic outcomes in a trial with a significant effect on ocular Ct infection or TF would provide additional evidence about the relationship between these indicators.
Currently, decision-making for MDA in trachoma elimination programs relies solely on clinical grading of TF. Grading of clinical trachoma is subjective, and prevalence surveys have demonstrated that there is poor agreement between clinical disease and ocular Ct infection, particularly after multiple rounds of antibiotic treatment [2,3,14,16]. Additionally, TF may be observed in the absence of infection, either as residual inflammation from the etiologic agent Ct[17,18], or due to other non-chlamydial bacteria [19]. Point prevalence of TF, or a test of infection, reflects disease or infection state at a current point in time, whereas serologic patterns may allow for identification of longer-term patterns in Ct transmission [7,8]. Anti-Ct antibody responses increased with age in this study, whereas TF prevalence was not significantly different across age groups, suggesting that antibody positivity rates represent the pool of exposed individuals rather than currently or recently-infected ones.
While PCR assessment of ocular Ct infection and clinical grading for TF and TI represent cross-sectional prevalence of trachoma, age-seroprevalence curves may provide additional insight to changes in transmission of ocular Ct over time. PCR or NAAT testing has historically been too costly for use in program settings[20] but cost-effective PCR tests are now being evaluated in program contexts [6]. In ocular swab specimens, PCR tests for Ct infection are a more specific indicator for the causative agent of trachoma than antibody testing in sera, as antibody responses to the antigens we studied cannot differentiate between exposure to ocular Ct and other chlamydial infections. Perinatal transmission of Ct from mothers with urogenital chlamydia could potentially lead to seropositivity among young children, as could sexual exposure in individuals after the age of sexual debut. However, focusing on the younger children and on the age-seroprevalence curve rather than absolute rates of seropositivity may allow for distinction between ongoing ocular Ct transmission and a single exposure. For example, in the Solomon Islands, a lack of increase in seropositivity to Pgp3 in 1–9-year-olds correlated with low infection rates, despite the 26% prevalence of TF, contrasting to the steep increase in seropositivity with age observed amongst 1–9-year-olds in Kiribati where TF prevalence was 28% and infection prevalence was 24% [21,22].
Antibody testing for chlamydial antigens has been conducted in a number of trachoma program settings. In treatment-naïve communities, the slope of the age seroprevalence curve increased with increasing community TF prevalence [4,21]. A significant decline in antibody responses has been shown following mass azithromycin distribution compared to pre-treatment levels in a cross-sectional study [8]. In this mesoendemic region of Niger, we noted more than 30% prevalence of antibodies to Pgp3 and CT694, consistent with results from mesoendemic communities in Tanzania [4]. In settings where surveillance surveys for trachoma elimination have been conducted, age seroprevalence curves corresponded to decreases in trachoma transmission [23]. Seroprevalence of antibodies to Pgp3 in children ages 1 to 9 years in these surveys ranged from less than 2% in some surveys[24,25] to as high as 7.5%[26] but without a steep increase in age seroprevalence curves seen in settings with ongoing transmission. The current data add to this body of knowledge by evaluating antibody responses at impact surveys after 3 rounds of MDA, a program setting for which limited data exist.
TF prevalence in this study was 7.5% after 3 years of MDA and thus antibody data do not come from a setting in which elimination thresholds have been achieved. Furthermore, the communities included in this analysis were treated with an enhanced coverage target, with up to four days of treatment. Antibiotic coverage may have been higher than is seen in programmatic or trial settings with lower coverage targets. The antibody responses and curves therefore may not be representative of what would be seen in a previously-endemic setting that has reached the elimination threshold. Trachoma programs typically include children up to age 9 in monitoring. Other studies have shown further increases in antibody responses in children aged 6–9 years[21,26], and inclusion of those ages here may have improved our ability to draw inferences from the shape of the curve. Since ocular swabs were pooled for PCR analysis, we were unable to obtain individual-level correlations between PCR and antibody positivity.
WHO currently recommends use of TF for deciding programmatic endpoints. Laboratory testing, including Ct serology, could be used as a supplement or replacement for TF when conducting surveillance after validation of the elimination of trachoma as a public health problem, given serology is generally inexpensive, objective, and provides estimates of exposure over time.[26] The elimination thresholds do not require a complete absence of ocular Ct infection, and therefore infection may still be present in communities that have reached the district-wide elimination threshold, as was seen in Tanzania where infection was present but did not lead to re-emergence.[27] Having a test of exposure, or of repeated infection, would allow more complete evaluation of the history of exposure to ocular Ct in children.
The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the Centers for Disease Control and Prevention.
|
10.1371/journal.pntd.0004274 | Pregnancy Outcomes after a Mass Vaccination Campaign with an Oral Cholera Vaccine in Guinea: A Retrospective Cohort Study | Since 2010, WHO has recommended oral cholera vaccines as an additional strategy for cholera control. During a cholera episode, pregnant women are at high risk of complications, and the risk of fetal death has been reported to be 2–36%. Due to a lack of safety data, pregnant women have been excluded from most cholera vaccination campaigns. In 2012, reactive campaigns using the bivalent killed whole-cell oral cholera vaccine (BivWC), included all people living in the targeted areas aged ≥1 year regardless of pregnancy status, were implemented in Guinea. We aimed to determine whether there was a difference in pregnancy outcomes between vaccinated and non-vaccinated pregnant women.
From 11 November to 4 December 2013, we conducted a retrospective cohort study in Boffa prefecture among women who were pregnant in 2012 during or after the vaccination campaign. The primary outcome was pregnancy loss, as reported by the mother, and fetal malformations, after clinical examination. Primary exposure was the intake of the BivWC vaccine (Shanchol) during pregnancy, as determined by a vaccination card or oral history. We compared the risk of pregnancy loss between vaccinated and non-vaccinated women through binomial regression analysis. A total of 2,494 pregnancies were included in the analysis. The crude incidence of pregnancy loss was 3.7% (95%CI 2.7–4.8) for fetuses exposed to BivWC vaccine and 2.6% (0.7–4.5) for non-exposed fetuses. The incidence of malformation was 0.6% (0.1–1.0) and 1.2% (0.0–2.5) in BivWC-exposed and non-exposed fetuses, respectively. In both crude and adjusted analyses, fetal exposure to BivWC was not significantly associated with pregnancy loss (adjusted risk ratio (aRR = 1.09 [95%CI: 0.5–2.25], p = 0.818) or malformations (aRR = 0.50 [95%CI: 0.13–1.91], p = 0.314).
In this large retrospective cohort study, we found no association between fetal exposure to BivWC and risk of pregnancy loss or malformation. Despite the weaknesses of a retrospective design, we can conclude that if a risk exists, it is very low. Additional prospective studies are warranted to add to the evidence base on OCV use during pregnancy. Pregnant women are particularly vulnerable during cholera episodes and should be included in vaccination campaigns when the risk of cholera is high, such as during outbreaks.
| Pregnant women are at high risk of complications and fetal deaths when ill with cholera. However, they have been excluded in most cholera vaccination campaigns because of the lack of safety data on oral cholera vaccines during pregnancy. This study aimed to determine if the risk of pregnancy loss changed after the administration of the oral cholera vaccine in Guinea in 2012. We visited all households in Boffa and Koba sub-prefectures, where the vaccination campaign took place, and enrolled a total of 2,493 women in the study. In this large retrospective cohort, we found no association between fetal exposure to the cholera vaccine and the risk of pregnancy loss or malformation. Pregnant women are particularly vulnerable during a cholera episode and should be included in vaccination campaigns when the risk of cholera is high, such as during the outbreaks.
| Cholera represents a risk of complications for pregnant women and their fetus. Published literature reports fetal loss rates during cholera episodes of between 2% and 36% [1–7]. However, comparison of pregnancy outcomes among different reports is difficult, due to differences in inclusion criteria, treatment provided, and access to care. Although the exact cause of fetal death during a cholera episode has not yet been identified, several studies suggest an association between fetal loss and the degree of dehydration and hypovolemia [2,4–7].
In cholera-endemic countries, the World Health Organization (WHO) recommends vaccination “for groups that are especially vulnerable to severe disease and for which the vaccines are not contraindicated, such as pregnant women and HIV-infected individuals” [8]. WHO has prequalified two oral cholera vaccines (OCV), both consist of killed whole-cells of V. cholerae. One consists of several strains of V. cholerae O1 and a recombinant B subunit of the cholera toxin (WC-rBS, marketed as Dukoral); the other contains strains from both serogroups O1 and O139, but no component of the cholera toxin (BivWC, marketed as Shanchol) [8]. According to the package inserts, neither vaccine is contraindicated in pregnant women, but only recommended when the potential benefits are considered higher than the risk. Inactivated OCVs are unlikely to have a harmful effect on fetal development as the killed bacteria in the vaccine do not replicate, the vaccine antigens act locally in the gastrointestinal mucosa, are not absorbed and do not enter the maternal or fetal circulation. In addition, the vaccines do not trigger systemic reactions (e.g. fever) associated with miscarriage in early pregnancy [9].
Pre-licensure studies and post-marketing surveillance suggest that Dukoral has a good safety profile when used during pregnancy [4] and inadvertent vaccination of pregnant women with the vaccine during a mass vaccination campaign in Zanzibar in 2009 was not associated with any harmful effects [9]. However, pregnant women have been excluded systematically from most other cholera vaccination campaigns because of the weak data on safety during pregnancy for Dukoral and the absence of safety data during pregnancy for Shanchol [10]. Shanchol has several advantages compared with Dukoral for public health use. The vaccine is cheaper, has a lower storage volume and does not require water for administration. Thus, understanding the safety of BivWC during pregnancy will provide essential information for its future use throughout the cholera-endemic world.
The Ministry of Health and Public Hygiene (MHPH) of Guinea, with the support of Médecins Sans Frontières (MSF), carried out mass OCV campaigns using BivWC in 2012 in Boffa and Forécariah Prefectures as part of a comprehensive response to a cholera epidemic that was spreading in remote rural areas with limited access to health facilities [11,12]. These campaigns targeted all people aged one year and above living in the target areas [11,12]. Pregnant women were not excluded from the target population.
In order to assess whether there was a difference in pregnancy outcomes between women who exposed their fetus to OCV and those who did not, we report the results of a retrospective cohort study, which compared the incidence of pregnancy losses (miscarriages and stillbirths) and malformations between these two groups.
The study took place in Boffa Prefecture of Guinea where six sub-prefectures bordering the ocean were targeted for cholera vaccination campaigns. All residents one year of age and above were offered a first dose from 18 to 23 April and a second dose from 9 to 14 May 2012 (Fig 1). The retrospective cohort study was conducted in two of these sub-prefectures (Koba and Boffa), since the association between vaccine exposure and pregnancy outcomes was assumed independent of the sub-prefecture.
Women were included in the study if they were residents of the Koba and Boffa subprefectures, were 15 to 49 years old, were pregnant in 2012 (i.e., conception and/or birth occurred that year) and if they (or their guardians for minor participants) provided informed consent. Exclusion criteria were non-residence in Boffa prefecture at the time of the vaccination campaign, absence from the home after two visits, lack of knowledge of their vaccination status, and refusal to participate.
Based on published literature [13–16], we assumed a 10% incidence of pregnancy loss, an unexposed/exposed ratio of 0.3 (based on 77% of pregnant women vaccinated in the vaccination coverage survey), an alpha error of 0.05, and a statistical power of 0.8. Thus, 1,200 vaccinated pregnant women and 360 non-vaccinated pregnant women were necessary to estimate a 1.5 increase in the risk of pregnancy loss among vaccinated women.
All interviewers and supervisors were recruited locally and received theoretical and practical training. They visited all households (defined as a group of individuals living under the same roof and regularly sharing the same meals). Interviewers revisited households later in the day where no one was at home. If there was no response the second time, the household was skipped. Interviewers asked the head of household for the number of women between 16 and 50 years old living in the household, and the number of women who were pregnant in 2012, irrespective of pregnancy outcome. They obtained written informed consent from the women who were pregnant in 2012 and conducted face-to-face interviews in the local language.
A standardized pre-tested questionnaire was used to collect inclusion criteria, socio-demographic data, information about the pregnancy, pregnancy history and other risk factors for pregnancy loss. Vaccination status was assessed at the end of the questionnaire. Interviewers also completed a questionnaire to determine the health condition of live-born babies. Mothers and children were referred to a pediatrician if the questionnaire elicited concerns. The pediatrician completed a clinical examination and determined if the child was ill or presented any malformation. The medical team was also in charge of patient management (i.e. ambulatory treatment or transfer to hospital), if needed.
The primary outcome of the study was the incidence of pregnancy loss, defined as any loss of a product of conception after the woman recognizes she is pregnant. Secondary outcomes included the incidence of miscarriage, stillbirth and malformation in live children. A miscarriage was defined as a loss of a clinically recognized pregnancy before the end of the fifth month of gestation and a stillbirth as the delivery of a dead fetus (without pulse) after the end of the fifth month of gestation. These outcomes were reported orally by the mother and verified by documentation when possible. A malformation was defined as a physical defect in a live infant that was identified by the study pediatrician.
Primary exposure was defined as the intake of OCV during pregnancy. Participants were asked whether they had been vaccinated and, if so, to show their vaccination cards. A fetus was considered exposed if the mother was pregnant during the campaign, received at least one dose of OCV (card-confirmed or reported orally), and at least one dose was received after the estimated date of conception and before the date of birth or fetal loss. Date of birth was reported orally and verified by documentation when possible. The date of conception was calculated by subtracting the duration of the pregnancy (reported orally or confirmed by documentation) from the date of birth or fetal loss. When date of birth or fetal loss was unknown, the mother was asked if she was pregnant during the vaccination campaign.
The primary data analysis included women who were pregnant during the mass vaccination campaign. Descriptive analysis of these women was stratified by their vaccination status. Qualitative and quantitative variables were compared, respectively, through Fisher and Wilcoxon tests. The fetus was then considered as the unit of analysis since some women had multiple pregnancies. We calculated crude cumulated incidence of pregnancy loss as the number of pregnancy losses divided by the number of conceived fetuses. We compared the risk of pregnancy loss through a binomial regression. Possible confounders were variables for which p-values were less than 0.20 in the bivariate analysis. We obtained an adjusted estimate of relative risk (aRR) of pregnancy loss and its 95% confidence interval (95%IC) according to OCV exposure using a forward stepwise procedure. The interaction between trimester of the pregnancy on 18 April 2012 and primary exposure was tested. All covariates significantly associated with the risk of a pregnancy loss (p-value <0.05) or those improving model fit (based on Bayesian Information Criterion) were retained in the final model. Women with missing data were excluded from the analysis.
In a secondary analysis, the same procedure was applied to other negative outcomes (miscarriages, stillbirths and malformations). Fetuses born to mothers who had been pregnant for more than five months on 18 April 2012 were excluded from the analysis of the risk of miscarriage. Fetuses who did not complete five months of gestation were excluded from the analysis of the risk of stillbirth. Children who were not alive at the time of the survey (fetal or perinatal deaths) were excluded from the analysis of the risk of malformations.
A bias-indicator analysis of fetuses conceived in 2012 after the second vaccination round was conducted to assess bias from possible misclassification of the women vaccination status or fetal outcome. This analysis again compared pregnancy outcomes of woman who had been vaccinated during the campaign with women who did not receive the vaccine. Since OCV intake before conception is not supposed to have an effect on pregnancy outcome, this analysis provides information about possible information bias.
Since the exact dates of vaccination, conception and fetal lost were mainly estimates, we conducted sensitivity analyses by excluding all fetuses born or lost within seven days of the first round of the vaccination campaign and those whose estimated date of conception was within two weeks following the first round of the campaign.
Data entry was performed using EpiData 3.1 (EpiData Association, Denmark) and data analysis was performed using Stata 12.0 (College Station, USA).
This study was conducted according to the ethical principles for research on human subjects, described in the Helsinki Declaration, and in accordance with international principles and guidelines for biomedical research involving human subjects, published by the Council for International Organizations of Medical Sciences. The study protocol was approved by National Ethics Committee of the Republic of Guinea and the Médecins Sans Frontières Ethics Committee.
Each woman (or her legal representative) received the information on the methods and potential risks and benefits of the study. The participant or her representative signed an informed consent form after being informed that participation in the study was voluntary and that she could withdraw from the study at any time. Anonymity and confidentiality of collected data were ensured throughout the study. If there was any suspected illness in the live-born babies, they were referred to the pediatrician, were treated or referred and hospitalized, if needed. All treatment was provided free of charge.
From 11 November to 4 December 2013, 10,211 households were visited; 315 were absent (3.1%) and 13 refused to participate (0.1%). A total of 15,732 women 16 to 50 years old were asked about their pregnancy status and 3,177 (20.2%) reported a pregnancy in 2012 (Fig 2). After applying the exclusion criteria, 2,724 women pregnant in 2012 were enrolled; however, 231 were excluded at the time of the analysis (Fig 2). One woman was pregnant twice in 2012. A total of 2,494 pregnancies were therefore included in the analyses; 1,543 in the primary analysis and 951 in the bias-indicator analysis.
Overall, 84.8% [95%CI: 83.0–86.6%] of the women pregnant during the campaign received at least one dose of OCV and could therefore have exposed their fetus to the vaccine. Vaccine coverage was significantly higher among women who were pregnant during the vaccination campaign (primary analysis) than those who became pregnant after the campaign (bias-indicator analysis), both for the first round (81.1% [95%CI: 79.2–83.1%] vs 76.1% [95%CI: 73.4–78.8%], p-value = 0.003) and the second round (64.0% [95%CI: 61.6–66.4%] vs 55.5% [95%CI: 52.4–58.7%], p-value<0.001). Vaccination status was confirmed by vaccination card in 24% of the cases.
Women vaccinated during their pregnancy were not significantly different from those not vaccinated in terms of socio-demographic variables, pregnancy history, pregnancy status and practices, aside from owning a television (p = 0.033) and an oven (p<0.001) (Table 1). Vaccinated and non-vaccinated women included in the bias-indicator analysis were also similar in their baseline characteristics (Table A in S1 Appendix).
Most (84.3%) of the women pregnant during the vaccination campaign presented a child health record booklet. The percentage of women who received antenatal care services and who delivered in a health facility was higher among those who received the vaccine during their pregnancy than those who did not, though the differences were not statistically significant (Table 1).
A total of 1,584 fetuses whose mother was pregnant during the campaign were included in the primary analysis; 1,312 (82.8%) were exposed to the vaccine (Table 2). A total of 56 fetuses were classified as lost. There was no difference in the crude cumulative incidence of pregnancy loss between fetuses exposed to the vaccine and those who were not (p = 0.350).
The adjusted risk ratio for pregnancy loss (aRR) was 1.13 [95%CI: 0.54–2.38, p-value = 0.738] (Table 2). The risk of pregnancy loss was found to be higher among fetuses of mothers who reported a cholera episode in 2012 than those who did not in the adjusted analysis (aRR = 3.18 [95%CI: 1.56–6.48], p-value = 0.002) (Tables B-D in S1 Appendix). The interaction between the trimester of pregnancy on April 18, 2012 and the primary exposure was not significant (p = 0.465) (Table G in S1 Appendix). In the bias-indicator analysis, the risk of pregnancy loss was not associated with the vaccination status (aRR = 1.19 [95%CI: 0.47–3.00], p-value = 0.717).
A total of 1,263 fetuses exposed to the vaccine and 265 non-exposed fetuses were born alive. Among them, 18 exposed (1.4% [95%CI: 0.7–2.1%]) and five non-exposed (1.9% [95%CI: 0.2–3.5%]) babies died before the survey. This difference was not statistically significant (p-value = 0.577). In addition, 133 children (8.8%) were referred to the study pediatrician among those screened in the primary analysis, as were 87 children (9.4%) in the bias-indicator study.
After the pediatrician’s clinical examination, seven vaccine-exposed children and three non-exposed children were considered to have a malformation (Table 2). Malformations were mainly from limbs (five from lower limbs and two from hands) (Table E in S1 Appendix). There was no statistically significant increase in the risk of malformation for fetuses exposed to OCV in the primary analysis (p-value = 0.314) (Table 2). After adjusting for other factors, the risk of malformation was significantly associated with the mother’s profession (p-value = 0.008) (Table F in S1 Appendix). In the bias-indicator analysis, the risk of malformation was not associated with vaccination status (aRR = 0.51 [95%CI: 0.13–2.02], p-value = 0.341).
These are the first estimates of the risk of pregnancy loss following vaccination of pregnant women with the bivalent, whole-cell only oral cholera vaccine. Exposure of the fetus to this vaccine was not significantly associated with the risk of pregnancy loss and malformation in this study. Vaccine coverage among pregnant women was high (83%) and similar to the overall vaccination coverage of the campaign [11]. This suggests that pregnant women who were offered OCV during the campaign chose to participate rather than forego vaccination. Vaccination coverage was higher among women who were pregnant during the campaign than among those who become pregnant after the campaign. Pregnant women may have been better informed about the vaccination campaign, less occupied by outside activities on the day of vaccination, and more willing to follow the advice of the Ministry of Health to get the vaccination than non-pregnant women. Overall, vaccinated and non-vaccinated women had similar baseline characteristics, both in the primary and in the bias-indicator analyses. Vaccinated pregnant women included in the primary analysis were more likely to attend antenatal care services and delivered more frequently in health facilities than those not vaccinated, which could be the result of a greater interest and awareness of preventive activities during pregnancy.
The lack of association between the exposure of the fetus to OCV and pregnancy loss in both the crude and the adjusted primary analysis is consistent with the findings with Dukoral in Zanzibar (aRR = 1.62 [0.76–3.43], p-value = 0.21) [9]. In the present study, the exposure of the fetus to OCV was not significantly associated with miscarriage or stillbirth. In the Zanzibar study, analysis of pregnancy loss was not broken down by miscarriage or stillbirth, although the crude incidence of stillbirths was slightly higher among vaccinated women (4.6% versus 2.1%) [9].
Another key finding in this study is that women who reported having had cholera in 2012 while they were pregnant were at six times higher risk of miscarriage and three times higher risk of having a stillborn child than women who did not report having had cholera. Although consistent with the literature [1–7], biological confirmation of cholera cases and determination of the date of onset of the illness would have strengthened the causal link between cholera episodes and pregnancy loss. The number of reported cholera episodes was lower among vaccinated versus non-vaccinated women who were pregnant during the campaign. This is in line with the vaccine effectiveness (86%) reported following the campaigns in Guinea [17].
The main reason newborns were referred to the pediatrician for clinical examination was illness rather than malformation. Malformations were detected mainly on upper and lower limbs. After adjusting on other factors, exposure to OCV was not statistically associated with malformation.
There are several important limitations of note in this study. First, the incidence of pregnancy loss was lower than expected both in vaccinated and non-vaccinated women, especially in the first trimester. Pregnant women may not have reported, or been aware of, pregnancy losses during the study period. Conversely, some women could have falsely reported pregnancies or loss of pregnancies, since few pregnancy losses could be verified on official documentation. Since the number of pregnancy losses is low, this possible information bias could affect our point estimates, though it is difficult to determine in which direction. Second, less than 25% of the women could present a vaccination card, leading to potential misclassification of their vaccination status. In order to minimize this potential bias, we reminded participants about the way the vaccination campaigns were organized and the route of administration. To understand further the potential presence of information bias, we conducted a bias-indicator analysis to estimate the risk of pregnancy loss among women who were pregnant after the vaccination campaign. As in the primary analysis, the risk of pregnancy loss in the bias-indicator analysis was slightly but not significantly higher among vaccinated women.
Another possible bias influencing our results is the presence of a seasonal component in pregnancies and pregnancy losses (Fig A in S1 Appendix). When comparing non-vaccinated women, the incidence of pregnancy loss was higher among women who were pregnant during the campaign than among women who become pregnant afterwards. We could therefore not consider fetuses conceived after the vaccination campaign as controls in the primary analysis, reducing the power of our study.
Lastly, as previously discussed, the number of negative events was lower than expected and the vaccine coverage was higher than expected, leading to a low number of non-exposed fetuses with negative events. This reduced the power of our analysis to detect statistical differences.
In conclusion, we found no association between fetal exposure to OCV and risk of pregnancy loss or malformation. Despite the weaknesses of a retrospective design and a decreased statistical power due to the low number of fetuses not exposed to the vaccine, we can conclude that if there is a risk of poor pregnancy outcomes from taking OCV during pregnancy, it is likely to be very small. Further studies are needed to confirm these results and provide further evidence about the risks and benefits of OCV for pregnant women and their fetus. As far as possible, these studies should be prospective cohort studies to reduce the likelihood of misclassifying negative pregnancy outcomes or exposure to the vaccine.
It is also important to note that any small potential risk of pregnancy loss could be offset by the possible benefit of vaccination. During preventive campaigns in non-epidemic periods, if the risk of infection is low, vaccination of pregnant women could be delayed, notably for women who have other risk factors for pregnancy loss. However, during epidemics, when the risk of cholera infection is high, vaccination should be offered to all pregnant women, since they are at particularly high risk of losing their fetus if they become ill with cholera.
|
10.1371/journal.pbio.1001200 | Fossilized Biophotonic Nanostructures Reveal the Original Colors of 47-Million-Year-Old Moths | Structural colors are generated by scattering of light by variations in tissue nanostructure. They are widespread among animals and have been studied most extensively in butterflies and moths (Lepidoptera), which exhibit the widest diversity of photonic nanostructures, resultant colors, and visual effects of any extant organism. The evolution of structural coloration in lepidopterans, however, is poorly understood. Existing hypotheses based on phylogenetic and/or structural data are controversial and do not incorporate data from fossils. Here we report the first example of structurally colored scales in fossil lepidopterans; specimens are from the 47-million-year-old Messel oil shale (Germany). The preserved colors are generated by a multilayer reflector comprised of a stack of perforated laminae in the scale lumen; differently colored scales differ in their ultrastructure. The original colors were altered during fossilization but are reconstructed based upon preserved ultrastructural detail. The dorsal surface of the forewings was a yellow-green color that probably served as a dual-purpose defensive signal, i.e. aposematic during feeding and cryptic at rest. This visual signal was enhanced by suppression of iridescence (change in hue with viewing angle) achieved via two separate optical mechanisms: extensive perforation, and concave distortion, of the multilayer reflector. The fossils provide the first evidence, to our knowledge, for the function of structural color in fossils and demonstrate the feasibility of reconstructing color in non-metallic lepidopteran fossils. Plastic scale developmental processes and complex optical mechanisms for interspecific signaling had clearly evolved in lepidopterans by the mid-Eocene.
| Biological structural colors are generated when light is scattered by nanostructures in tissues. Such colors have diverse functions for communication both among and between species. Structural colors are most complex in extant butterflies and moths (lepidopterans), but the evolution of such colors and their functions in this group of organisms is poorly understood. Fossils can provide insights into the evolution of biological structures, but evidence of structurally colored tissues was hitherto unknown in fossil lepidopterans. Here, we report the preservation of structurally colored scales in fossil moths with striking metallic hues from the ∼47-million-year-old (Eocene) GrubeMessel oil shales (Germany). We identify the color-producing nanostructure in the scales and show that the original colors were altered during fossilization. Preserved details in the scales allow us to reconstruct the original colors and show that the dorsal surface of the forewings was yellow-green. The optical properties of the scales strongly indicate that the color functioned as a warning signal during feeding but was cryptic when the moths were at rest. Our results confirm that structural colors can be reconstructed even in non-metallic lepidopteran fossils and show that defensive structural coloration had evolved in insects by the mid-Eocene.
| Structural color has long been of interest to biologists. It is phenotypically significant in many organisms [1], forms the basis of diverse inter- and intra-specific communication strategies [2], and is implicated in pivotal evolutionary transitions [3]. Evidence of structural color has been reported from some fossil biotas [3]–[6], but has received little attention. This limits our ability to reconstruct the origins of activity patterns, habitat preferences, and social and sexual signaling mechanisms [7]. This is particularly problematic in the case of Lepidoptera (butterflies and moths), which exhibit the most complex and diverse structural colors of any living group of organisms [8]. Structural colors in extant lepidopterans are generated by modification of one or more components of the basic scale architecture (longitudinal ridges and transverse crossribs upon a basal lamella that is supported by columnar trabeculae in the scale lumen) into a biophotonic nanostructure of chitin and air [9]. Such color-generating multilayer structures can arise via specialization of the ridges and their ridge-lamellae, crossribs, or the scale lumen; the lumen can also exhibit various other modifications, including complex three-dimensional photonic crystals. The various color-producing nanostructures in lepidopteran scales may be related developmentally [9] and all generate color via interference of scattered light [1], although the overall visual effect can be influenced by other optical mechanisms at the level of ultra- and macrostructure [10]. Attempts to reconstruct the evolution of color-producing nanostructures using phylogenetic and/or structural evidence have hypothesized that multilayer structures in the scale lumen are evolutionarily primitive [11],[12], but these conclusions are not widely accepted [1],[10].
Fossils provide direct evidence of stages in the evolution of biological structures and can be used to test evolutionary hypotheses. The lepidopteran fossil record extends from the Early Jurassic to the Recent and includes representatives of numerous extant lepidopteran families [13],[14]. Fossil specimens of adult macrolepidopterans often exhibit light- and dark-toned areas on their wings [14] and can retain ultrastructural details of their scales [15]; preservation of pigmentary or structural colors has not been reported. Most fossil lepidopterans occur as inclusions in amber and within fine-grained sediments [14]; Baltic and Dominican amber (Eocene-Oligocene), the lacustrine sediments of Florissant (Eocene, Colorado), and the offshore marine sediments of the early Palaeocene Fur Formation (Denmark) are especially rich sources [14]. Fossil lepidopterans have also been reported (but not described) from the mid-Eocene Messel oil shale of Germany, which is celebrated for preserving a diverse paratropical ecosystem with remarkable fidelity [16]; the biota includes mammals, reptiles, amphibians, abundant fish and insects, and plants [17], the last represented by leaves, fruits, and seeds [18]. Messel fossils are typically well preserved: animals are often well-articulated and many show evidence of soft tissues (including stomach contents); insects (especially beetles) may exhibit metallic coloration [16]. Here we use scanning- and transmission electron microscopy (SEM and TEM), reflectance micro-spectrophotometry, and 2-D discrete Fourier analysis [1],[19] to demonstrate that metallic color in the fossil lepidopterans from Messel is structural in origin and to reconstruct their original color.
The fossils (Table S1) occur as isolated individuals (Figure 1A, Figure S1) and in coprolites (Figure S1) but have not been described [20]. Wing venation patterns indicate that specimens are possibly extinct representatives of the Zygaenidae (burnet and forester moths), in particular Procridinae (forester moths; see Text S1). Two taxa are represented; specimens of the smaller taxon are more complete, and therefore the focus of this study. Electron dispersive X-ray analyses demonstrate that the fossil scales are organically preserved: they comprise predominantly carbon and there is no evidence for replacement of the preserved tissue by authigenic minerals.
Brilliantly colored scales cover the dorsal surface of the forewing except for a thin brown (non-metallic) zone along the outer margin (Figure 1A–C, Figure S1); they are restricted to basal and discal zones of the ventral surface (Figure S1). The dorsal surface of the hindwing is predominantly brown but exhibits metallic colors apically (Figure S1); the ventral surface is not visible in any specimen. Metallic scales also occur on the body of the insect. Specimens in glycerine exhibit predominantly yellow colors in basal and discal to postdiscal zones of the wing; the color grades to green and then blue in postdiscal to submarginal wing zones, and is brown along the outer wing margin (Figure 1A–C; Figure S1a–c). Scales on the abdomen typically exhibit yellow to orange colors in glycerine. The observed color varies when a fossil is placed in media of different refractive indices (Figure S2) in a fashion characteristic of many structurally colored materials [21].
The gross morphology of the scales is difficult to determine as they overlap and are typically fractured. Ultrastructural evidence demonstrates that four types are present (Figures 1, 2, Figures S3, S4). Type A scales, the most common, are the primary contributor to the observed color. They are cover scales and occur over the dorsal and ventral surface of the forewing (Figures 1D–J, 2). The abwing surface of these scales, as in extant lepidopterans [22], exhibits prominent longitudinal ridges connected by orthogonal crossribs (typical spacing 1.8–2.5 µm and 510–600 nm, respectively) (Figure 1D). The ridges are up to 1 µm high (Figure 2C). They comprise overlapping lamellae (each 1.2–3.1 µm long and 110–150 nm wide) inclined at 10–12° to the scale surface and exhibit short lateral microribs (typical spacing 122–170 nm) (Figure 1D–G). The ridges and crossribs frame a series of windows that are typically perforated (Figure 1E–G); the lamina perforation factor (p) [23] increases from the proximal (p = 0.05) (Figure 1F) to distal (p = 0.32) parts of a scale (Figure 1G).
The scale lumen contains 3–5 perforated internal laminae that differ in their structure and thickness (Figures 1E,H, 2). The uppermost lamina (93–124 nm thick) exhibits densely packed, bead- to rod-like spacers (60 nm wide and 60–500 nm long) (arrow in Figure 1E). The next two to four laminae exhibit less densely packed, bead-like, spacers (typically 60 nm×60 nm) (Figure 1I) and decrease progressively in thickness (from 74–110 nm to 55–63 nm) towards the adwing scale surface (Figure 2A,B). In the proximal parts of a scale, the lowermost lamina in the stack is the basal lamina of the scale. In medial and distal parts of a scale, however, the stack is supported by additional pillar-like trabeculae (each 0.6–1 µm high) above a lamina with a distinctive reticulate texture (55–65 nm thick), which forms the base (Figure 1J). The ultrastructure of Type A scales (including the spacing of laminae, which is known to control color in living lepidopterans) varies according to their color and location on the wing (see Table S2). Similar variation occurs in extant lepidopterans [22],[24]. Even non-metallic brown scales in the fossils preserve ultrastructural details, including the laminar ultrastructure in the scale lumen (Figures S3, S5). Scale types B, C (“satin-type” [9]), and D are rare and do not contribute significantly to the observed color in the fossils (see Text S1).
Three-dimensional structures in fossils are vulnerable to compaction during burial of the host sediments. It is not assumed a priori that the preserved structure of the laminar array in the fossil lepidopteran scales is identical to that in vivo, especially as the trabeculae are typically fractured and now orientated parallel to, and superimposed upon, the basal lamina of the scale (Figure 2G). There is, however, no evidence that the laminar array has been similarly affected. Successive laminae are not superimposed and the vertical spacers between them are neither fractured nor flexed. Preferential fracturing of the trabeculae may have been promoted by their wider spacing and greater height. There is no evidence (e.g., dessication cracks) that the geometry of the laminar array was affected by shrinkage of the scales during diagenetic dehydration of the organic tissue. Nor is there evidence for diagenetic expansion of the scale structure: spacers are continuous between adjacent laminae. Collectively, these observations indicate that diagenetic processes had little or no impact upon the preserved structure of the laminar array. The preserved ultrastructure is therefore considered to be extremely similar, if not identical, to that originally present in vivo.
The ultrastructure of the laminar array was not modified during fossilization and is therefore a reliable basis for reconstructing the original colors of the fossil scales. 2-D Fourier analysis of longitudinal TEM images of the ultrastructure in the lumen of scales from the basal part of the dorsal forewing reveals two points of high values aligned above and below the origin (Figure 3A,B). The dominant periodicity is in the vertical direction; that is, the preserved structure is highly laminar. Fourier power spectra of transverse TEM images show a wider distribution of Fourier power peaks above and below the origin (Figure 3C,D). This results from the concave geometry of the laminar array in transverse section and consequent increase in the range of angles over which the observed color maintains the same peak hue [1]. Radial averages of the Fourier power spectra demonstrate that the preserved laminar nanostructure is a multilayer reflector: the peak spatial frequencies in refractive index lie within the range capable of producing visible colors by scattering of light (Figure 3E).
The visual properties of extant lepidopteran scales can be influenced by scale tilt (the angle between the scale and the wing membrane) [10], scale curvature [2], the number and thickness of laminae [25], the degree of overlap of ridge lamellae [21], the spacing of the ridges [21], microribs [21], and crossribs [26], and the lamina perforation factor [23]. Scale tilt and curvature are not preserved in the fossils. The number (up to five) of laminae and their different thicknesses indicate that the fossil multilayer reflector is non-ideal (i.e. reflects much less than 100% of incoming light) [27]. The ridge lamellae in the fossils do not overlap sufficiently [28] to have a significant impact on the observed color. Closely spaced microribs or crossribs in satin-type scales can also generate diffraction [26] and play a secondary role in the generation of blue and violet colors in lepidopteran scales [19],[23]. The spacing of the microribs (140 nm) in the fossil satin-type scales, however, is significantly less than the wavelength of visible light (approx. 350–700 nm); conventional diffraction theory indicates that zero-order diffraction (i.e. specular (directional) reflectance) will be produced. Further, the satin-type scales are restricted to the inner margin of the forewing in the fossils, precluding their having a significant impact on the observed color. Collectively, these observations indicate that the primary color-producing nanostructure in the fossils is the multilayer reflector in the scale lumen.
The optical properties of the fossil scales are, however, influenced by the perforation factor and concave geometry of the laminae, and by the spacing of the ridges. In extant lepidopterans, iridescence, spectral bandwidth, and total reflectance are reduced at higher perforation factors (between 0.2 and 0.4) relative to scales with lower perforation factors [23]; this generates a purer (albeit less intense) color that is visible over a wider range of angles. In the fossils the exposed (medial and distal) parts of the overlapping scales typically have perforation factors of 0.32. Concave distortion of laminar arrays also reduces iridescence [1]; the arcuate geometry of the fossil multilayer reflector in transverse section would have enhanced the iridescence-reducing effect of the perforated laminae. Multilayer reflectors typically generate directional (specular) color that flashes at specific observation angles [29]; this effect can be modified by diffraction. Ridge periodicities of between 0.85 µm and 4 µm generate diffraction [21],[30],[31]; a strong diffractive effect has been reported for periodicities of ca. 1.3 µm [31] and 1.7 µm [30]. The ridges in the fossils are spaced 1.8–2.5 µm apart and therefore probably constitute diffraction elements that render the color generated by the multilayer reflector visible over a wide range of observation angles, but do not contribute to the observed hue [31].
Scales from the dorsal surface of the basal part of the forewing exhibit a measured reflectance peak of 473 nm (Figure 3F) that corresponds to their blue color in air. The predicted peak of reflectance (with λmax = 565 nm) calculated from the radial averages (using refractive index values of 1.56 and 1.0 for the high- and low-index layers, respectively), however, indicates that the dorsal surface of the basal part of the forewing was originally yellow-green. The color in air today and the measured reflectance peak are artefacts, probably a result of alteration of the biomolecular composition of the scale cuticle, and thus its refractive index, during fossilization; most fossil arthropod cuticles are chemically altered during diagenesis [32]. Furthermore, recent experiments using extant butterfly scales demonstrated that alteration of the original organic material results in a shift in the reflectance peak without altering the scale ultrastructure significantly [33]. Calculation of reflectance peaks for scales from other parts of the wings (see Figure S5) allows the original colors of the fossil lepidopterans to be reconstructed (Figure 4; Table S3). Scales in postdiscal to submarginal wing zones have predicted reflectance peaks λmax≈515 nm and λmax≈440 nm, respectively; scales along the wing margins have a predicted reflectance peak λmax≈750 nm, and scales from the abdomen have a predicted reflectance peak λmax≈550 nm (Figure S5). The fossil lepidopterans therefore originally exhibited yellow-green hues in basal and discal to postdiscal zones of the wing; the color graded to green-cyan and then blue in postdiscal to submarginal wing zones, and was brown along the outer wing margin. Scales on the abdomen were yellow to yellow-green.
Structural colors in extant butterflies function primarily in species and mate recognition [29]; the function of structural colors in extant moths, however, has not been investigated. The fossil moths described here are colored most highly on the dorsal surfaces of the forewings (the surfaces which are exposed in most extant moths, including zygaenids, when they are at rest [34]), suggesting that the Eocene moths, like extant zygaenids, were diurnal. The visual ecology of the structural color in the fossil moths can therefore be compared with those in extant diurnal lepidopterans. The fossil moths were characterized by a yellow-green dorsal coloration that was visible over a wide range of angles but not highly reflective. The visual signal lacked certain properties, e.g. strong iridescence, brightness, and color contrast within the wing, that are important in conspecific communication [30]. Instead, the optical characteristics of the fossil scales, notably their original yellow-green hue and suppression of iridescence, indicate a primary defensive function. In extant lepidopterans, reduced iridescence enhances presentation of visual signals for protective purposes [35]. Structural green coloration functions cryptically in extant butterflies [24],[29] and beetles [36],[37]. In particular, a combination of a structural green hue with reduced iridescence provides particularly efficient color matching with a diffuse leafy background [36],[37].
A cryptic function for the structural color in the fossil lepidopterans is consistent with the ecology of extant zygaenid moths: many Procridinae species with green scales are cryptic except when feeding on flowers [38], when they can be highly conspicuous (Gerhard Tarmann (Tirolier Landesmuseen, Austria), personal communication) [39]. The latter feature is inconsistent with cryptism: high chromatic contrast with the background environment is characteristic of an aposematic (warning) signal [40]. However, an aposematic function for the structural color while feeding does not necessarily conflict with a cryptic function in a foliaceous environment: dual-purpose visual signals are known in extant lepidopterans [41]–[43]. The visual signal generated by the structurally colored scales in the fossil lepidopterans probably served two functions: cryptic when specimens were at rest, and aposematic during nectaring. It is possible that this dual function is evolutionarily conserved in Procridinae and that aposematism and diurnality are ancestral traits of zygaenids. Further, defensive behavior in the fossil moths is consistent with the use of chemical defense: extant zygaenids, including taxa that are largely cryptic [28], can synthesize cyanide for defense by enzymatic breakdown of cyanoglucosides [24],[44],[45].
The discovery of structural color in Messel lepidopterans constrains the timing of the origin of several important evolutionary novelties. Different scale types in extant lepidopterans arise via subtle modifications of a common membrane-folding developmental process dominated by self-assembly [9],[22],[46]. The presence of different scale types in the fossils confirms that such plastic developmental processes had evolved in moths by the mid-Eocene. The complexity of the iridescence-reducing nanostructure in the fossil moths indicates that sophisticated optical mechanisms for interspecific signaling were in use at this time. Predator-prey interactions are recognized as a major stimulus in insect evolution [47]; the use of cryptic and aposematic signals by the fossil moths described here supports the evidence of other fossils from Messel [48] that sophisticated mechanisms for avoiding detection by visually hunting predators had evolved in insects by the mid-Eocene. The striking resemblance of the fossil moths to some extant zygaenids and the cryptic/aposematic function of their structural color suggest that dual-purpose visual signals, and especially aposematism, may be evolutionarily conserved in this group of moths, originating early in the history of the group and persisting to the present day. Preservation of ultrastructural detail in all scales in the fossils, even non-metallic brown examples, offers the possibility of reconstructing the original colors and patterning of even lepidopteran fossils that lack obvious structural color.
Specimens are held by the Senckenberg Forschungsinstitut und Naturmuseum, Forschungsstation Grube Messel, Germany.
Small (2–3 mm2) tissue samples were removed using sterile tools and, for TEM, placed in the following ethanol∶glycerine mixtures, each for 24 h under rotation: 10%, 25%, 50%, 75%, 100% ethanol. For SEM, samples were dehydrated using HMDS or under vacuum, mounted on aluminum stubs, carbon- or gold-coated, and examined using a FEI XL-30 ESEM-FEG microscope equipped with an EDAX energy disperse X-ray spectrometer. Observations were made at an accelerating voltage of 15 kV, with acquisition times of 60 s for EDS spectra of carbon-coated samples. For TEM, samples were washed in propylene oxide twice, each for 1 h, and impregnated with Spurr's resin under vacuum in the following resin∶ethanol mixtures, each for 24 h: 10%, 20%, 30%, 40%, 50%, 60%, 70%, 80%, 90%, 100% resin. To ensure optimal orientation for sectioning, a 10 mm3 block of resin containing the sample was extracted, re-orientated, and re-embedded in 100% resin. Ultrathin (80–90 nm thick) microtome sections were placed on formvar-coated Cu grids, stained using uranyl acetate and lead citrate, and examined using a Zeiss EM900 TEM at 80 kV with an objective aperture of 90 µm diameter.
Reflectance spectra were recorded from samples in 100% glycerine, 100% ethanol, and in air (the latter from only the basal part of the dorsal forewing to minimize damage due to drying) using an epi-illumination Nikon Optiphot 66 microscope, an Ocean Optics HR2000+ spectrophotometer, and a tungsten-halogen light source; spectra were collected from a 70 µm spot. All recorded spectra were normalized against the spectrum of the light source recorded from a white standard.
Nanoscale spatial periodicity in the refractive index of a material results in constructive interference of scattered light; structural color is generated where such scattering occurs in the visible part of the spectrum. Herein we use an established analytical method [13] of analyzing the periodicity and optical properties of structurally colored biological tissues using the discrete Fourier 2-D transform. Digital TEM micrographs of scales from the fossil lepidopterans were analyzed using MATLAB (version 7.11.0) and a 2-D Fourier tool freely available as a series of MATLAB commands (http://www.yale.edu/eeb/prum/fourier.htm). Variation in the refractive index of nanostructures in the fossil scales was analyzed using the procedure described in ref. [1].
The reconstruction of the original colors of the fossil lepidopterans is based upon the preserved ultrastructure of the multilayer reflector in the scale lumen and the assumption that the original refractive index of the fossil scale cuticle was similar to that in modern lepidopterans (i.e. ∼1.56). Only the original colors of the dorsal surface of the specimens were reconstructed: (1) only this surface is exposed at rest and (2) the ventral surface of the hindwing is not visible in any specimen. TEM images of the multilayer reflector preserved in Type A scales of different colors were analyzed using 2-D Fourier analysis. Predicted wavelength values scales from different locations on the wing are based on 2–4 replicate analyses. Wavelength data for predicted reflectance peaks were converted to RGB values. Calculation of precise RGB values for a specific wavelength, however, is difficult [49]. RGB values for predicted wavelength data were therefore calculated using three different methods: (1) using the “Wavelength to RGB” application available from http://miguelmoreno.net/sandbox/wavelengthtoRGB/ (downloaded December 28, 2010), (2) using the “Spectra” application available from www.efg2.com/lab (downloaded December 28, 2010), and (3) using the “Wavelength to RGB” converter available online at www.uvm.edu/~kspartal/Physlets/Lecturedemo/LambdaToRGB.html (accessed December 28, 2010). The three methods yield similar RGB values; the colors depicted in the reconstruction are based on the averages of the values obtained.
|
10.1371/journal.pcbi.1006643 | Detection and analysis of spatiotemporal patterns in brain activity | There is growing evidence that population-level brain activity is often organized into propagating waves that are structured in both space and time. Such spatiotemporal patterns have been linked to brain function and observed across multiple recording methodologies and scales. The ability to detect and analyze these patterns is thus essential for understanding the working mechanisms of neural circuits. Here we present a mathematical and computational framework for the identification and analysis of multiple classes of wave patterns in neural population-level recordings. By drawing a conceptual link between spatiotemporal patterns found in the brain and coherent structures such as vortices found in turbulent flows, we introduce velocity vector fields to characterize neural population activity. These vector fields are calculated for both phase and amplitude of oscillatory neural signals by adapting optical flow estimation methods from the field of computer vision. Based on these velocity vector fields, we then introduce order parameters and critical point analysis to detect and characterize a diverse range of propagating wave patterns, including planar waves, sources, sinks, spiral waves, and saddle patterns. We also introduce a novel vector field decomposition method that extracts the dominant spatiotemporal structures in a recording. This enables neural data to be represented by the activity of a small number of independent spatiotemporal modes, providing an alternative to existing dimensionality reduction techniques which separate space and time components. We demonstrate the capabilities of the framework and toolbox with simulated data, local field potentials from marmoset visual cortex and optical voltage recordings from whole mouse cortex, and we show that pattern dynamics are non-random and are modulated by the presence of visual stimuli. These methods are implemented in a MATLAB toolbox, which is freely available under an open-source licensing agreement.
| Structured activity such as propagating wave patterns at the level of neural circuits can arise from highly variable firing activity of individual neurons. This property makes the brain, a quintessential example of a complex system, analogous to other complex physical systems such as turbulent fluids, in which structured patterns like vortices similarly emerge from molecules that behave irregularly. In this study, by uniquely adapting techniques for the identification of coherent structures in fluid turbulence, we develop new analytical and computational methods for the reliable detection of a diverse range of propagating wave patterns in large-scale neural recordings, for comprehensive analysis and visualization of these patterns, and for analysis of their dominant spatiotemporal modes. We demonstrate that these methods can be used to uncover the essential spatiotemporal properties of neural population activity recorded by different modalities, thus offering new insights into understanding the working mechanisms of neural systems.
| Recent advances in brain recording techniques have led to a rapid influx of high spatial- and temporal-resolution datasets of large neural populations [1–4]. One of the major challenges in modern neuroscience is to identify and extract important population-level structures and dynamics from these datasets [5,6]. Traditionally, neural population activity has been mainly studied from the perspective of temporal synchrony or correlation, and relating correlated neural activity to brain functions has been the major focus of many studies in neuroscience during the past two decades [7,8].
However, growing evidence indicates that population-level brain activity is often organized into patterns that are structured in both space and time. Such spatiotemporal patterns, including planar traveling waves [9–11], spiral waves which rotate around a central point [12–14], source and sink patterns which expand or contract from a point [13,15], and saddle patterns which are formed by the interaction of multiple waves [13], have been observed at different neural levels within multiple recording techniques, including multi-electrode arrays [13,16–18], voltage sensitive dye (VSD) imaging [9,12,19], and electroencephalography (EEG), electrocorticography (ECoG), magnetoencephalography (MEG) and functional magnetic resonance imaging (fMRI) [20–24].
The functional role of these spatiotemporal patterns is a subject of active research: In spontaneous activity, propagating patterns have been shown to follow repeated temporal motifs instead of occurring randomly [13,15], and are postulated to facilitate information transfer across brain regions [10,17] and carry out distributed dynamical computation [25]. In sensory cortices, stimuli can elicit repeatable propagating patterns [9,10,19,26,27], and the properties of these waves can be linked to stimulus features. For instance, the phase and amplitude of traveling waves in the motor cortex and visual cortex correlate with reach target location [17] and with saccade size [18], respectively, and the propagation direction of moving patterns in the visual cortex is sensitive to visual movement orientation [28]. These studies thus indicate that the ability to detect and analyze these patterns is essential for uncovering the principled dynamics of neural population activity and for understanding the working mechanisms of neural circuits [15,26,29,30].
In this study, to detect changes of neural signals happening across both space and time, we introduce velocity vector fields which represent the speed and direction of local spatiotemporal propagations. These vector fields allow us to make a novel conceptual link between spatiotemporal patterns in neural activity and complex patterns such as vortices or eddies found in the field of fluid turbulence [31–33], in which these patterns are similarly characterized by using velocity fields of the underlying moving molecules. Velocity vector fields in our methods are computed by adapting optical flow estimation methods originally developed in the field of computer vision [34]. Optical flow techniques have previously been implemented to analyze brain activity [13,26–28], but here we extend these methods to consider the amplitude and phase of oscillatory neural signals, allowing for a comprehensive analysis of neural spatiotemporal patterns. When constructed from oscillation phase, velocity vector fields are conceptually similar to phase gradient vector fields as often used in previous studies [15,18]. However, velocity vector fields provide a conceptual basis for us to adapt methods from turbulence to develop a unified methodological framework for analyzing neural spatiotemporal patterns.
We show that by examining the critical points in a velocity vector field (also called “stationary points” or “singularity points”), where the local velocity is zero [35], different types of spatiotemporal patterns including spiral waves (“foci”), source/sink patterns (“nodes”) and saddles can be detected. In addition to these complex wave patterns, neural systems can exhibit widespread synchrony and planar travelling waves. These types of activity are common to many physical and biological systems, and can be detected by introducing global order parameters calculated from velocity vector fields [36]. These methods thus enable the automatic detection of a diverse range of spatiotemporal patterns after user-defined parameters have been chosen; these parameters are discussed in detail in Methods and Materials.
Aside from detecting these patterns, our methods can provide systematic analysis of pattern dynamics including their evolution pathways and their underlying spatiotemporal modes that exhibit intrinsic and inseparable spatial and temporal features, thus providing a novel alternative to existing dimensionality reduction techniques which instead separate space and time components [6]. We validate the effectiveness of all methods and their implementation in the toolbox through multiple approaches. Using synthetic data with known pattern activity, we show that spatiotemporal pattern detection is accurate and reliable even in noisy conditions. We then analyze local field potentials from multi-electrode arrays in marmoset visual cortex and whole-brain optical imaging data from mouse cortex to test our methodological framework across different recording modalities, species, and neural scales. We find that pattern properties including location and propagation direction are modulated by visual stimulus, and that patterns evolve along structured pathways following preferred transitions.
Here we outline a methodological framework for detecting and analyzing wave patterns in neural recordings using velocity vector fields. These methods can be applied to any recording methodology with high spatial and temporal resolution, including multi-electrode LFPs, VSD and optical imaging, ECoG, EEG, and MEG. Some of these methods have been briefly described in our previous work [13], but in this paper we examine them in more depth and show how they can be uniquely combined with new techniques to form a systematic framework for pattern detection and analysis. We also discuss their implementation into a freely available MATLAB toolbox, the NeuroPatt toolbox. Neural data from a two-dimensional spatiotemporal recording are represented by a four-dimensional matrix, zx,y,t,p, where z is the recorded signal with regular spatial coordinates x,y, time t, and trial presentation p. Recordings without repeated trials or averaged across all trials have p=1, although we caution that trial-averaged signals typically do not capture the spatiotemporal patterns present in single trials [24,28].
Neural activity, although appearing highly disordered at the single-neuron level, can form dynamical coherent structures such as propagating waves at the population level [37]. There are many other complex systems that display similar emergent pattern dynamics, including fluid turbulence, in which coherent flows and vortices emerge from interacting molecules that behave irregularly [31,32]. Velocity vector fields, which represent the direction and speed of fluid motion, are essential mathematical tools for detecting and analyzing coherent activity patterns embedded in turbulent flow [32]; studies using this approach typically separate activity patterns at different scales, independently detecting both large-scale flows and small-scale eddies [33]. In turbulent flow, velocity vector fields are typically measured by following the movement of tracer particles within the fluid [38]. Here, we introduce a method for calculating analogous velocity vector fields in neural signals, representing the local direction and speed of propagating activity at each recording site. As for in studies of coherent structures in turbulence, these velocity fields obtained in neural data provide a powerful conceptual framework for analyzing a diverse range of propagating wave patterns in the brain.
For a data sequence D(x,y,t), which may represent the raw recorded signal z or the amplitude A or phase θ of an oscillatory neural signal, extracted by using either the Hilbert transform [39] or complex Morlet wavelets [40] (see Oscillatory data filtering in Methods and Materials), the velocity vector field w(x,y,t)=(u(x,y,t),v(x,y,t)) represents the velocity in x- and y-directions at each location between time t and t+δt, where δt is the time step specified by the sampling frequency. If data contain multiple trials, the velocity vector field is computed iteratively for each trial.
To calculate the velocity field w(x,y,t), we adapt optical flow estimation techniques from the field of computer science, which were first developed by Horn and Shunck to track the motion between successive frames of a sequence of images [34]. In their original formulation, the optical flow is calculated by solving two constraints. The first is the data constancy constraint,
D(x+u,y+v,t+δt)−D(x,y,t)=0,
(1)
which specifies that the same data is present at time t and time t+δt, only shifted in space. To first order, this can be expressed as
Ed=Dxu+Dyv+Dt≈0,
(2)
where Ed is the error in the data constancy, Dx and Dy denote spatial derivatives, and Dt denotes the temporal derivative at the point x,y,t. The second is the spatial smoothness constraint, which specifies that the computed velocity fields contain smooth and continuous motion where possible. This constraint can be expressed as
Es2=|∇u|2+|∇v|2,
(3)
where Es quantifies the overall departure from smoothness, and ∇≐∂x,∂y denotes the gradient operator. The velocity vector field can uniquely be defined by minimizing both these error terms:
minu,v{∬[ρ(Ed2)+αρ(Es2)]dxdy},
(4)
for some regularization parameter α and penalty function ρ. Horn and Shunck used a quadratic penalty, ρx2=x2, but this can lead to inaccuracies if the underlying data contains hard edges and adjacent regions moving in different directions [41]. More accurate velocity vector fields can be obtained by using the Charbonnier penalty, ρ(x2)=x2+β2, for a small positive constant β [42]. Eq 4 can be solved by linearizing its corresponding Euler-Lagrange equations, creating a unique velocity vector field w (see Solving optical flow equations in Methods and Materials).
Turbulence studies typically separate activity at different scales based on velocity fields [33]. We similarly implement independent procedures to detect global patterns (plane waves and synchronous activity), which are active across the whole recording area, and complex wave patterns, including source, sink, spiral and saddle patterns, which are characterized by local activity around their central points. Complex spatiotemporal wave patterns, which are analogous to eddies, are organized around critical points where the velocity field has zero magnitude [35]. These complex wave patterns generate distinctive dynamics around their central critical points; in our methodology, we exploit this dynamical property to automatically detect and classify such patterns. In velocity fields, we identify critical points as locations where both x- and y-components of the velocity are zero by finding intersections of the bilinearly interpolated zero-level contours lines of u and v [43]. Each critical point is then categorized by the Jacobian matrix,
J=(∂u∂x∂u∂y∂v∂x∂v∂y),
(5)
which is estimated at the critical point using bilinear interpolation from the corners of the surrounding 4-element cell of recording sites. Depending on the trace (τ) and determinant (Δ) of the Jacobian, critical points are classified as a node (∆>0 and τ2>4∆), focus (∆>0 and τ2<4∆), or saddle (∆<0), and node and focus points are further classified as stable (τ>0) or unstable (τ<0).
These classes of critical points correspond to different types of wave patterns (Fig 1): Nodes expand or contract from a critical point, forming sources or sinks, respectively; saddles have one stable axis and one unstable axis, and are typically formed through interactions between different waves; and foci rotate around the critical point, thereby corresponding to spiral waves. In addition to their rotating motion, foci can also involve expansion or contraction from the critical point, forming spiral-out or spiral-in wave patterns, respectively. However, previous studies of spiral waves have not distinguished between spirals-out and spirals-in [12,14,44]. In our methods and toolbox, these patterns can therefore optionally be combined to facilitate direct comparison with other published results.
Although complex wave patterns are classified only by the local properties of their central critical point, they can spread over larger regions of space. We thus develop a method for characterizing the spatial extent of complex wave patterns by using the winding number (Poincaré index), which has a value of +1 for node and focus patterns and -1 for saddle patterns for all closed paths within the pattern’s extent around the location of the critical point [43]. We create approximately circular paths around the location of the critical point, and the winding number is estimated in each of these paths as
windingnumber=12π∑k=1n(θk+1−θk),
(6)
where θk is the angle of the k-th vector around a closed counter-clockwise path with n points, angles are subtracted circularly, and where θn+1=θ1. We compute the winding number in paths of expanding size around a pattern’s center, and its spatial extent is defined by the largest area within which all computed winding numbers are consistent with the critical point type. This procedure therefore provides an efficient estimate how far wave patterns spread across the cortex, an important property of neural oscillatory activity.
We next introduce methods for detecting and analyzing simple, large-scale patterns such as synchrony and planar waves by defining order parameters based on the velocity vector fields. We detect planar waves using an order parameter defined as the average normalized velocity [36]:
φ(t)=||∑x,yw(x,y,t)||∑x,y||w(x,y,t)||
(7)
This statistic is equivalent to phase gradient directionality [17] except it uses velocity vector fields instead of the phase gradient. Normalized velocity ranges from zero to one, with φ→1 as velocity vectors align to one direction, reflecting coherent motion across the recording area. Plane wave activity is therefore detected at times when φ is greater than some threshold value Tpw, which should be close to one (Tpw=0.85 by default in the toolbox).
If data has been band-pass filtered to extract the oscillation phase θ, we also detect large-spread synchronous activity using another order parameter, which is defined as the resultant vector length of phase across the recording area [45,46],
R(t)=1N|∑x,yeiθ(x,y,t)|,
(8)
where N is the number of spatial recording sites in phase maps. The resultant vector also ranges from zero to one, with R→1 as the phase of oscillations at all recording sites align to the same value, reflecting wide-spread synchrony. We note that the order parameter as defined in Eq 8 is similar to that used to characterize global synchrony in coupled phase oscillators [47], and 1-R is commonly defined as the circular variance [45]. Synchronous activity is therefore detected at times when R is higher than another threshold value Tsyn, the default of which is Tsyn=0.8 in the toolbox.
To test the performance of our pattern detection methods, we generated artificial data sets with simultaneous source and sink patterns active at the same frequency, located at random positions and propagating in random directions within a 12×12 spatial grid (see Simulated data in Methods and Materials). We then added Gaussian white noise, band-pass filtered the signal, calculated velocity vector fields, searched for complex wave patterns in the velocity fields, and compared the detected pattern centers with their true locations. An example of this procedure is shown in Fig 2, which shows calculated velocity fields and pattern centers between two frames of a simulated data set (Fig 2A). The velocity fields depend on two parameters in the optical flow estimation procedure (Eq 4): The smoothness regularization parameter α, and non-linear penalty constant β.
The smoothness regularization parameter α determines the weighting of the smoothness constraint compared to the data constraint, and thereby the overall smoothness of the velocity fields. Small values of α generate velocity fields that primarily capture local changes and are therefore sensitive to added noise, potentially leading to the detection of spurious, noise-driven patterns (Fig 2B, left column, α = 0.1). Large values of α are more robust to noise, but can over-smooth the data, creating mostly uniform flow fields that do not capture the underlying dynamics (Fig 2B, right column, α = 1.5). Reasonable values for α can range from ~0.1 to ~20, depending on the size of the data, the dynamics of the propagations, and the level of noise; for example, reducing the spatial sampling frequency of a dataset reduces the number of grid spaces between complex patterns, typically requiring a lower value of α to effectively resolve individual patterns. The non-linear penalty constant β determines the degree of non-linearity of the penalty functions, with large values β≫1 resulting in a quadratic penalty and small values β≪1 in a more robust non-linear penalty. Small values of β give more accurate velocity vector fields for any regions with discontinuous motion in the underlying data [41,48], but we find that such discontinuities are rare in neural recordings, so using large values of β generally gives similar results (Fig 2B). In addition, when β is large and the equations are effectively quadratic, the optical flow procedure can typically converge much faster.
The choice of appropriate values for α and β cannot be fully automated for a real dataset without making assumptions about the dynamics of the data. However, pattern detection accuracy can be evaluated in simulated datasets with specified properties and pattern dynamics, which can then be used to guide parameter choices in real data. Fig 3 illustrates the effectiveness of the pattern detection algorithm for one such set of properties and dynamics (see Simulated data in Methods and Materials). The detected spatial position of patterns is most accurate when using small values of α (α≤1, Fig 3A). Using a quadratic penalty function (β=10) generally gives more consistent results across a range of α values than a non-linear penalty function (β=0.01) and results in fewer missing patterns (Fig 3B), but using the non-linear penalty can give a lower false positive rate (Fig 3C). Plotting the true positive rate against the false positive rate provides a clear way to examine the effectiveness of pattern detection across a range of parameters (Fig 3D). We generally recommend using large values of β when examining new data sets, as this provides more reliable performance and faster processing overall. Additionally, pattern detection is largely unaffected by noise if the variance of the noise is equal or less than the variance of the pattern oscillations and remains fairly accurate for significantly greater noise levels (Fig 3E–3G).
To validate our methods and test for wave pattern activity in real neural data from different scales and imaging techniques, we examined previously published LFP recordings from the MT area of anaesthetized marmosets [49] and optogenetic voltage imaging recordings from a complete cortical hemisphere in awake mice [50] (see Experimental recordings in Methods and Materials). Using our methodology, we searched for wave patterns within the phase and amplitude of oscillations across a range of frequency bands. Both datasets exhibited a rich repertoire of wave patterns which were successfully detected. Some examples of common pattern activity for each modality are shown in Fig 4.
In delta-band phase of the marmoset data, complex waves were commonly present across the whole 16 mm2 recording area, including saddle (Fig 4A) and spiral-out (Fig 4B) patterns. We also observed multiple complex wave patterns active simultaneously in different areas of the cortex, as shown for sink and saddle patterns in Fig 4C. In the mouse data, complex waves were present in the phase of slow (4 Hz) oscillations, and these waves sometimes spread across the whole cortical hemisphere, including sink (Fig 4D) and spiral-in (Fig 4E) patterns. Large-scale propagating patterns were also present in the amplitude of 10 Hz oscillations, where multiple spreading patterns often interacted to form saddles (Fig 4F). These examples demonstrate that complex wave patterns are present at multiple scales of brain activity, and that these patterns can be detected and quantified through our methodology.
Having presented our pattern detection methods, we now demonstrate how these techniques can be used to examine the properties and dynamics of waves patterns in more detail, and how these properties can be further related to brain function. Directly tracking simple and complex wave patterns allows their location, movement direction, prevalence, duration and other properties to be collated across many occurrences. To validate the results of the pattern detection procedure, the properties of patterns detected in a real dataset can be compared to those of patterns detected in a surrogate dataset comprised of noise with similar characteristics to the original data (see Pattern detection parameters and result validation in Methods and Materials). The processes of band-pass filtering and velocity vector field estimation can smooth data and may therefore generate spurious wave patterns in noise-driven surrogates. However, these patterns in surrogate data are typically more localized and transient than real neural wave patterns and can therefore be mostly removed if the minimum pattern spatial extent and duration parameters are sufficiently large. In neural recordings with genuine wave pattern activity, all pattern types will typically occur more frequently (Fig 5A), be present for a larger proportion of recording time (Fig 5B) and last longer per occurrence (Fig 5C) than equivalent patterns in noise-driven surrogates.
The properties of wave patterns can vary depending on brain state, recording location, or cognitive task, revealing relevant dynamical changes in the recorded neural system. An example of this is shown in Fig 6, which compares properties of patterns in spontaneous and stimulus-evoked phase velocity fields from the same animal. During ongoing activity (sustained blank screen stimulus), plane waves were active for much of the recording time and propagated in a range of directions (Fig 6A, mean resultant vector length 0.28). Complex wave patterns were also common and did not form randomly in space. Instead their central critical points were clustered around preferred locations (Fig 6B), which were situated at different points in the recording array for node and saddle patterns. When relevant stimulus was presented (coherently propagating dot fields turned on and off every two seconds), relatively fewer plane waves were active overall, but their propagations were more tightly distributed around one preferred direction (Fig 6C, mean resultant vector length 0.42). The presence of stimulus also affected the overall prevalence of critical point patterns, increasing the number of stable and unstable nodes and decreasing the number of saddles, and changed their patterns of distribution across space (Fig 6D). Our methods can therefore be used to quantify changes in spatiotemporal pattern dynamics driven by different stimuli, cognitive tasks or behavioral states.
Detected wave patterns can also be processed to reveal their temporal evolution dynamics. Brain activity evolves between different activity patterns in a complex and non-random way, but the mechanisms of these transitions are not well-understood [13,15]. Our methodology provides an ideal framework for exploring such dynamics: Once all patterns in a recording have been identified, common pattern transitions and motifs can easily be identified. We demonstrate some of these evolution dynamics in stimulus-evoked LFP recordings (see Pattern evolution dynamics in Methods and Materials). Patterns were typically active for tens to hundreds of milliseconds, often then transitioning into a different pattern (Fig 7A). The total number of transitions between all pairs of pattern types were counted across a recording, and the significance of these observed counts was established by comparison to the expected number of counts if all patterns began and ended at random times (Fig 7B). Using this simple analysis, we observed that periods of plane wave and synchronous activity were usually interspersed by other pattern types, synchronous activity was highly likely to transition to or from all other pattern types, and patterns commonly evolved from sources to sinks and vice versa. This analysis illuminates the temporal dynamics of the spatiotemporal activity patterns present in the recording and provides quantitative measurements which can be linked to cognitive tasks or used to constrain models of cortical dynamics. Similar analyses can also facilitate tracking the movement of neural structures of interest across brain regions [15], detecting repeated motifs in pattern dynamics [13], or examining gradual changes in pattern dynamics corresponding to changes in brain states [50].
In neural recordings, amplitude and phase data at the same frequency reflect different properties of brain activity, with amplitude representing a combination of the coherence and overall activity of a local ensemble and phase representing the timing of its oscillations. Accordingly, these signals typically contain different spatiotemporal patterns, and both phase and amplitude patterns can be relevant and informative. Fig 8 illustrates the spatiotemporal profile of raw LFP data, filtered LFP data, and the amplitude and phase of filtered LFP data, again taken from marmoset visual area MT. The spatiotemporal dynamics in the raw data (Fig 8A) primarily reflect those in the oscillations with greatest power, but also contain a large amount of noise from other frequencies. Filtering the data to a narrow-band signal (Fig 8B) reduces the noise by extracting the patterns present in the chosen frequency band alone, but these wave patterns are typically complicated as they are influenced by two different types of propagating activity: amplitude patterns (Fig 8C), which capture the movement of the overall shape of the wave and travel at the group velocity [51], and phase patterns (Fig 8D), which capture the progression of timing differences between electrodes and travel at the phase velocity.
In a general oscillating system, the phase and amplitude are independent properties which have no a priory reason to affect each other. However, there is some evidence that phase and amplitude patterns can be related in some neural systems: Phase patterns in rabbit sensory cortices are more commonly observed around the formation of new amplitude patterns [52], and spiral waves in mammalian neocortex exhibit consistently reduced amplitudes at their centers [12,44]. Examining both phase and amplitude separately may therefore uncover similar relationships in other experimental protocols and can reveal a more comprehensive understanding of the underlying dynamics of cortical circuits. For example, two simple patterns can be resolved from the complicated activity in Fig 8B: A gradually expanding activation from a point near the center of the recording array, as revealed by the amplitude velocity field in Fig 8C, and a plane wave propagating across the recording area, as revealed by the phase velocity field in Fig 8D.
Whilst direct identification of wave patterns in velocity fields as described in the previous sections allows for patterns’ individual dynamics to be fully characterized, the procedure does not specify the extent to which these patterns contribute to the overall spatiotemporal dynamics of a recording. To address this, we introduce a complementary method for studying wave activity in neural recordings by using velocity field decomposition, which finds low-dimensional spatiotemporal modes that capture the majority of variance in the system. Dimensionality reduction techniques are commonly used for uncovering underlying neural mechanisms of brain function [6]. However, the majority of existing techniques use principal component analysis (PCA) or similar procedures that decompose data into independent spatial and temporal modes, obscuring activity that is not time-space separable such as propagating waves and patterns [53]. Some studies have used decomposition techniques to specifically detect waves by examining phase gradients in complex decompositions of data [26,29]. To identify dominant spatiotemporal patterns in our framework, we again obtain inspiration from the field of turbulence, in which dimensionality reduction is often directly applied to velocity fields, capturing low-dimensional spatiotemporal dynamics [33].
In turbulence, dimensionality reduction can be performed through a variety of different decomposition methods, including Reynolds decomposition, principal component analysis (or proper orthogonal decomposition), and dynamic mode decomposition [54]. These techniques find modes capturing the majority of the energy in the system, which is not well-defined for velocity fields of neural data as it is in fluid flows, but some of these methods nonetheless can be adapted to find low-dimensional representations of the primary spatiotemporal dynamics of a neural recording. We implement a simple singular value decomposition (SVD) to extract the dominant spatiotemporal patterns from a time series of velocity fields in an efficient and parameter-free way.
To reorganize the velocity fields ux,y,t,vx,y,t into standard form for decomposition with variables in columns and observations in rows, we combine spatial dimensions and rearrange indices to define u~t,r,v~t,r, for time t and recording site r. We then use two alternate approaches to combine ũ and ṽ. In the first approach, we concatenate the two matrices across recording sites to define the real matrix w˜re(t,r')=[u˜|v˜]. In the second, we represent the velocity field as a complex number to form the complex matrix w˜co(t,r)=u˜(t,r)+iv˜(t,r). In either case, the singular vector decomposition (SVD) is defined as
w˜=TΣR*,
(9)
where w~ denotes w~re or w~co, T and R are unitary matrices, * denotes the conjugate transpose, and Σ is a rectangular diagonal matrix of positive numbers σi, called the singular values [55]. This operation finds orthogonal linear combinations of recording sites that explain the greatest variance in the velocity fields, and is closely related to PCA: if w~ has been shifted so that each recording site has zero sample mean, then R comprises exactly the principal component loadings, and σi2 are the principal component scores [56]. However, normalizing the velocities at each recording site is counterproductive in this application, as biases in propagation direction are an important component of wave dynamics. The k-th spatial mode, defined by the velocity field in the k-th column of R, explains a proportion of the overall variance given by σk2/∑iσi2, and has a time course given by the k-th row of T.
The vector SVD procedure is closely related to traditional PCA methods, as both techniques reduce the dimensionality of a dataset by extracting patterns that comprise the bulk of the variance and their evolution over time. The differences between these approaches are illustrated in Fig 9, which again shows marmoset LFP data during moving dot-field stimulation. PCA typically decomposes data into orthogonal spatial modes (Fig 9A), which comprise linear combinations of recording sites [6]. Vector SVD instead processes the velocity vector fields to extract spatial modes which are vector fields themselves (Fig 9B), and therefore represent distinct propagation patterns in the underlying data. In both cases, each spatial mode has a corresponding time-course (or temporal mode), describing its evolution across the duration of a recording (Fig 9C and 9D).
Although the dominant PCA modes explain more variance than their SVD counterparts, their temporal components reveal structured interactions between the dominant spatial modes (Fig 9C), generating spatiotemporal activity patterns which are difficult or impossible to determine directly from the PCA modes. In contrast, SVD spatial modes directly reflect these spatiotemporal patterns, and their evolution over time represents the strength of different pattern types in response to stimulus. In Fig 9, stimulus onset generates large, clear changes in spatiotemporal pattern dynamics revealed by SVD (Fig 9D): Sink pattern activity increases dramatically but transiently (shown by the large deflection in mode 3); plane waves (modes 1 and 2) increase in activity more modestly, but change direction soon after stimulus onset (as indicated by the sign change of mode one at 300 ms) and are sustained for a longer period. These results suggest that stimuli can directly affect the dynamics of propagating wave patterns, but that these changes are obscured when using PCA or other decomposition methods that separate space and time.
We find that in both spontaneous and stimulus-evoked LFP recordings, velocity vector fields in phase and amplitude at all frequencies display consistent dynamics: the most dominant modes typically reflect orthogonal directions of plane wave motion, and the next most dominant modes contain complex patterns including sources, sinks, spirals and saddles (Fig 10A). Despite these similarities, the disparities between spatial modes in different recordings or conditions can reveal major differences in the underlying pattern dynamics, including the primary directions of plane wave motion, the center location of complex patterns, and the relative prevalence of different pattern types. As an example, we compare the dominant SVD modes during stimulus-evoked activity in Fig 9B to those during ongoing activity in the same recording (Fig 10A). The four most dominant modes represent the same activity patterns, but they display slightly different dynamics: The primary propagation direction of plane waves changes (shown by the direction of mode 1); the central locus of source, sink and saddle activity changes location (shown by the critical point in nodes 3 and 4); and ongoing activity overall contains more plane wave and less source, sink and saddle activity (as revealed by the percentage of variance explained).
In the SVD method discussed thus far, each class of spatiotemporal pattern may be represented across multiple modes (e.g., modes one and two both reflect plane wave activity), making their overall prevalence more difficult to calculate. To address this issue, we also implement a modified SVD procedure that we call complex singular value decomposition (cSVD), which treats velocity vectors as complex numbers. In this approach, temporal modes have both a real and imaginary component, allowing spatial modes to both scale and rotate over time: The amplitude of the temporal mode gives the relative strength of the pattern, and the argument of the real and imaginary components gives the angle by which all vectors are rotated. This approach effectively combines real SVD modes together (Fig 10): Modes with plane waves travelling in any direction are combined, as are modes containing source, sink and spiral patterns with the same center location, or saddles with the same center location. This allows the overall relative contribution of each type of activity pattern (plane waves, expanding or contracting waves, saddle patterns) to be accurately evaluated, but information about the direction of patterns is removed to the complex time evolution.
In this paper, we have introduced a methodological framework and associated MATLAB toolbox for the classification and analysis of propagating wave patterns in neural recordings, and illustrated these methods using simulated data, LFP recordings from marmoset visual area MT and whole-brain optical imaging data from mouse cortex. The toolbox is freely available under an open source agreement from [https://github.com/BrainDynamicsUSYD/NeuroPattToolbox]. As we have demonstrated, our methods provide a framework for uncovering the spatiotemporal organization principles of these patterns and for examining how they are related to brain function.
We have introduced velocity vector fields to characterize how neural oscillatory signals change across space and time. Based on these vector fields a range of mathematical techniques including order parameters, critical point analysis and winding number calculation are uniquely combined to detect a diverse range of wave patterns and to characterize their key spatiotemporal organization properties. Our methods thus build upon the application of optical flow analysis for detecting wave patterns developed in previous studies [15,57,30,58]. As we have demonstrated, order parameters can be used to detect the presence of large-scale plane waves or synchronous activity, and critical point analysis can detect complex wave patterns, comprising sources, sinks, spirals-in, spirals-out, and saddles. Calculation of the winding number around critical points can then be used to measure the precise size of wave patterns, which may be useful in future studies to examine the spatial scale of neural features and effects across different frequencies [59]. These approaches allow multiple classes of waves to be tracked simultaneously and systematically.
Applying these methods to experimental data, we successfully identified both small-scale wave patterns in LFP recordings from anaesthetized marmoset visual cortex and large-scale patterns in whole-brain optical recordings from awake mice. In both datasets, multiple coexisting patterns were commonly active and all patterns were significantly more prevalent than in noise-driven surrogate data (Fig 5). Furthermore, we showed that visual stimulation can change the direction of plane wave activity and the position of complex waves in marmoset area MT (Fig 6), and that these waves evolve between different types in a structured way beyond what is expected by chance (Fig 7). These results are consistent with previous studies associating visual stimuli and traveling waves [9,10,19], and showing that neural dynamics evolve following preferred pathways [13,15,60]. However, unlike previous work, our methodology allows these patterns and their dynamics to be simultaneously detected and quantified, and places them into a framework of explicit pattern behavior to more precisely study underlying neural dynamics.
In our methods, dominant spatiotemporal activity patterns can be extracted from a recording using novel vector field decomposition methods. These present a promising approach to the task of dimensionality reduction in large-scale neural recordings. Current dimensionality reduction methods typically process data into separable temporal and spatial modes which reproduce the dynamics of a recording [6]. However, these approaches find population structures that are often dominated by single-cell response properties and correlated activity [61], and do not adequately capture activity patterns that are not time-space separable, such as propagating waves [53]. In contrast, vector field decomposition specifically targets propagating waves by directly extracting spatiotemporal pattern modes from data. We showed that stimulus onset in marmoset LFP recordings generated complicated responses in spatial PCA modes that are difficult to interpret, but clear effects on the activity of spatiotemporal pattern modes. We also showed that the dominant spatiotemporal modes are consistent across recordings but change in dynamics depending on cognitive function. In the future, this approach could be useful for examining how sensory stimuli and cognitive tasks affect wave dynamics of population-level responses, and for visualization and exploration of the underlying spatiotemporal activity in large neural data sets. Together, the detailed wave pattern tracking approach, and broad, parameter-free velocity decomposition approach provide a comprehensive analysis of spatiotemporal activity patterns in neural recordings.
There are many ways that our methodology can be extended to explore spatiotemporal neural pattern dynamics beyond what has been presented in this paper. For instance, plane waves and synchronous activity are currently detected as global patterns active across the whole recording area, but it would be advantageous (particularly in large-scale recordings) to identify discrete local regions exhibiting these patterns. This would support accurate simultaneous analysis of brain areas displaying different dynamics, be useful for studying the spread of synchrony or plane wave propagations, and provide consistency with the localized nature of complex patterns as defined by the winding number. However, implementing localized order parameters across regions of different sizes would significantly slow the pattern detection procedure and introduce additional free parameters to the framework. Future work may develop more efficient methods to characterize localized regions displaying synchrony or planar propagations, potentially allowing entire cortical sheets to be fully and dynamically segmented into multiple interacting patterns.
Additionally, our methods can be further extended to explore currently unknown mechanisms of wave pattern interactions in the brain. Firstly, localized patterns with complex dynamics that are active simultaneously may directly interact. Such interactions are prevalent in modelling studies including spiking neural networks [37] and neural firing rate models [62], and they are theorized to be directly involved in distributed dynamic computation [25]. In experimental studies, interactions between sharp-wave ripple patterns in rat hippocampus can result in their reflection or annihilation [63], but more complex interactions have not been examined. Secondly, patterns may interact across oscillations at different frequencies. Currently, the phase of low-frequency oscillations is known to influence the amplitude of high frequency oscillations in the brain [64,65], but it is not clear how this cross-frequency coupling actually influences or is influenced by the underlying patterns in these systems. Finally, wave pattern dynamics may interact across multiple spatial and temporal scales in more complex ways, creating cascades of pattern dynamics comparable to energy cascades in turbulence studies [66]. By effectively detecting and analyzing neural spatiotemporal activity patterns simultaneously across multiple scales, our methods provide a framework for further exploring these key questions in future studies.
In this section we describe how the methodological framework introduced in the Results section is implemented in the NeuroPatt toolbox and outline how user-set parameters affect computations. We also briefly describe the experimental protocols of the data shown in the figures of this manuscript. The NeuroPatt Toolbox was written in MATLAB 2016b, and is freely available from [https://github.com/BrainDynamicsUSYD/NeuroPattToolbox].
NeuroPatt follows the workflow shown in Fig 11 and includes data filtering to extract oscillatory activity; optical flow estimation to quantify the direction and speed of propagations by constructing velocity vector fields; turbulence-inspired classification and tracking of simple waves (synchrony, planar travelling waves) and complex patterns (sources, sinks, spiral waves, saddles); and vector field decompositions to find dominant spatiotemporal dynamics.
NeuroPatt includes two complementary methods to band-pass filter oscillatory neural data to extract amplitude or phase at a chosen frequency prior to detection of spatiotemporal patterns. The first method uses an eighth-order Butterworth filter as implemented in MATLAB’s Signal Processing Toolbox to filter data to a specified frequency range. This filter is applied in both forward and reverse directions to minimize phase distortion. The oscillation amplitude, A, and phase, θ, are then extracted from the analytic signal, zf+iz^f=Aeiθ, where z^f is the Hilbert transform [39] of the filtered data zf. The second method estimates the analytic signal at a specified center frequency using the complex Morlet wavelet transform to filter data and extract phase and amplitude with an optimal trade-off between time and frequency resolution [40], as implemented by MATLAB’s Wavelet Toolbox.
These two procedures give comparable outputs [67], but each has advantages in different situations: The Hilbert transform allows the properties of the filtering to be precisely specified but can be invalid if the underlying frequency is not sufficiently narrow-band; the wavelet transform is usually faster to compute and always results in a valid analytic signal, but gives a less concretely defined frequency range. Both procedures are included in the toolbox, with the wavelet transform as the default option. Users without access to either the Signal Processing Toolbox or the Wavelet Toolbox can detect spatiotemporal patterns in unfiltered data, which are valid but can be contaminated by noise from multiple frequencies as illustrated in Fig 8, or can calculate the analytic signal through other implementations of the Hilbert or wavelet transform for use in the later steps of the toolbox. Any band-pass filtering procedure necessarily involves some degree of temporal smoothing [68], which can inhibit the extraction of precise timing information in later analysis steps. We note that this effect will not change the timing of maxima or minima in time series, as both wavelet and Hilbert filtering techniques do not distort signal phase, but they will smear out activity between these points.
Velocity vector fields are calculated by solving the Euler-Lagrange equations corresponding to the minimization problem given by Eq 4:
ρ′(Ed2)Dx[Dxu+Dyv+Dt]−α∇⋅[ρ′(Es2)∇u]=0,
ρ′(Ed2)Dy[Dxu+Dyv+Dt]−α∇⋅[ρ′(Es2)∇v]=0,
(10)
Where ρ′(x2)=(2x2+β2)−1. Note that for large values of β, ρ'x2 is approximately constant, and the optical flow for the Charbonnier penalty approaches that of the quadratic penalty. For clarity, we let ρd=1αρ′(Ed2) and ρs=ρ′(Es2), and rewrite these equations as
ρdDx[Dxu+Dyv+Dt]−ρs∇2u−∇ρs⋅∇u=0,
ρdDy[Dxu+Dyv+Dt]−ρs∇2v−∇ρs⋅∇v=0,
(11)
where ∇2≐(∂x2, ∂y2) denotes the Laplace operator. These equations can be solved through fixed point iteration for the functions ρd and ρs after linearizing all other terms. In the toolbox, we approximate partial first derivatives with a five-point stencil 112-1,8,0,-8,1 where possible (or with centered or forward differences when close to edges), and approximate the Laplacian with a 2D five-point stencil [69]. If D represents phase data, it will contain temporal and spatial discontinuities where phase wraps from -π to +π, which invalidate linear difference stencils taken near these points. Instead, we approximate partial derivatives with centered or forward differences calculated using circular subtraction, θ1-θ2=modθ1-θ2+π,2π-π. All figures in this report use parameters α=0.1 and β=10 unless otherwise specified.
To create valid velocity vector fields, NeuroPatt has some restrictions on the format and content of input data sets. Firstly, data must be spatially arranged in a 2D square lattice of recording sites. This restriction exists primarily because of the optical flow estimation methods implemented in the toolbox, which assume spatial uniformity to maximize efficiency and accuracy. Alternate optical flow estimation methods exist for 3D data sets sampled volumetrically [70] or from non-uniform surfaces [71], but these implementations require significant modifications to the methodology described here and are much more computationally intensive. Secondly, data must be consistent across multiple recording sites: Because velocity vector fields are computed based on local dynamics, any recording sites with erroneous activity can significantly influence the surrounding velocity vectors. As in previous studies using optical flow for brain recordings [58,72], we recommend that highly noisy data are spatially filtered prior to the application of optical flow methods, and that any invalid or discontinuous recording channels are interpolated over. When amplitude spatiotemporal patterns are being examined, data should also be normalized across recording sites by subtracting the baseline or z-scoring to remove factors that could cause any regional bias, such as uneven electrode impedance or dye intensity. These processes are included as optional pre-processing steps in the toolbox.
Finally, the changes between consecutive time steps in recorded signals must also be sufficiently small, as optical flow cannot be estimated if there are large discontinuities between frames. Such discontinuities can occur if signals are changing on a shorter time scale than the sampling frequency, which may be an issue for fast neural signals such as action potentials and high-frequency oscillations, or for recording techniques with low temporal resolution such as fMRI. There is no strict rule to determine if the sampling frequency is sufficiently high, but as a general guideline we suggest that signals at a single recording site should typically change by less than 10% of their maximum range between consecutive time steps. The toolbox will warn if this condition is not satisfied, potentially leading to non-convergent optical flow estimation or invalid velocity vector fields, or if the change in data is significantly below this threshold, indicating that it can be safely down-sampled for computational efficiency.
The toolbox includes multiple parameters for the identification and tracking of spatiotemporal patterns. Firstly, the user can specify a minimum distance from the edge of the recorded area Ledge (default 2 grid spaces) for critical points to occupy, as velocity fields can be inaccurate and contain spurious critical points close to the boundary (mainly due to the use of forward differences to approximate derivatives at these points). The user can also specify a minimum radius Lradius (default 2 grid spaces) for critical point patterns to occupy, to exclude small-scale and potentially noise-driven local patterns from analysis.
Once both simple and complex patterns are detected in individual velocity fields, these individual observations are combined across time to identify persistent spatiotemporal patterns. NeuroPatt allows the user to specify a minimum duration tdur (default 5 time steps) for all patterns (including global synchrony and plane waves): patterns which persist for less than this amount of time are discarded. To add some error tolerance to the process of linking patterns together over time, a maximum time gap tgap (default 1 time step) can be specified between successive instances of a pattern for it to still be counted as one spatiotemporal structure. Finally, complex patterns can move over time, so critical points of the same type in successive time steps are considered part of the same spatiotemporal pattern only if they are separated by less than a maximum displacement Ldisp (default 0.5 grid spaces).
The pattern detection methods and all relevant parameters can therefore be summarized as follows: Synchrony occurs when Rt>Tsyn at least every tgap+1 time steps for a period of at least tdur. Plane waves occur when φt>Tpw at least every tgap+1 time steps for a period of at least tdur. Nodes, foci and saddles occur when a critical point, at least Ledge away from the edge of the grid with a minimum spatial radius of Lradius, can be linked to another critical point of the same type within tgap+1 time steps and distance Ldisp, and this chain of critical points persists for at least tdur time steps. Nodes and foci with the same stability properties are typically treated as separate patterns. However, they represent the same type of motion (expansion from or contraction to the critical point), so the toolbox can optionally group these critical points together for a more robust characterization of these pattern types.
NeuroPatt contains a few key parameters that must be carefully selected by the user to ensure valid and accurate results. As illustrated in Fig 2, the optical flow smoothness parameter α controls many properties of the computed velocity vector fields and can over-smooth the data and create spurious plane wave activity if too large or generate flow fields dominated by local noise if too small. To assist with the selection of appropriate values of these parameters, the NeuroPatt toolbox can automatically generate simulated datasets with recording size, sampling frequency and oscillation frequency identical to input data, allowing α and β to be optimized based on the user’s data, as shown in Fig 3. Because the pattern dynamics are typically unknown prior to processing they must be guessed for the simulated data, but we find that in most cases the optimal parameter choices do not change significantly with the pattern types or sizes present.
Even within valid velocity fields, the detection of plane waves and synchronized activity is largely dependent upon the thresholds Tpw and Tsyn, which are typically arbitrary parameters set by the user. This is a persistent problem in the detection of such activity patterns: Previous studies have identified plane waves using template matching [24,30,73] or though alignment statistics [17]; and synchronous activity through correlation or coherence measures [74,75]. All these methods rely on largely arbitrary thresholds to explicitly detect patterns. To assist with the choice of these thresholds in our methodology, we implement an optional visual inspection step in the toolbox, allowing users to view sample periods of plane wave and synchronous activity for various thresholds and display distributions of the underlying velocity field statistics across a recording. If such distributions are multi-modal, they can suggest meaningful boundary points for threshold values.
To help to validate results, the NeuroPatt toolbox implements surrogate data generation to test results obtained from real data against results obtained from noise with the same basic dynamics as the input data. To achieve this, we construct time series comprising white noise for each recording site with the same mean and standard deviation of the corresponding site in the original data. We then repeat all processing steps including pre-processing, filtering, optical flow estimation and pattern detection with multiple random surrogate datasets and compare identified pattern statistics and dynamics with those obtained from the true data. If results are comparable between the real and surrogate datasets, it suggests that they may have been introduced through smoothing or other processing steps rather than being real effects in the data. This surrogate data testing process will therefore flag most false positive detections made by the toolbox. To ensure that all findings are robust to changes in parameters, we recommend that users verify that their results are consistent across a range of values for key parameters in the pattern detection process.
Once simple and complex spatiotemporal patterns have been detected, the evolution dynamics between different pattern types can be quantified. For each pair of pattern types (pA, pB), the observed number of transitions from pA to pB, nobspA→pB, can be counted in each trial by searching for instances where pB starts within a short time gap tgap of pA ending. This can be compared to the expected number of transitions if the patterns in the trial occurred at random times,
nexp(pA→pB)=nAnBtgapttrial,
(12)
where nA and nB are the observed number of patterns pA and pB within the trial and ttrial is the total length of the trial in seconds. The fractional change between observed and expected transition counts is then defined as nobs/nexp-1. We used paired t-tests with the Bonferroni correction for multiple comparisons to evaluate whether nobs and nexp were significantly different for each pattern transition across multiple trials of a recording.
We test the pattern detection procedures in NeuroPatt by generating data sets with known pattern properties. To create a simulated wave pattern zsim with wavenumber k and angular frequency w, we use the formula
zsim(x,y,t)=A(x,y,t)ei(ks(x,y,t)−wt),
(13)
where A(x,y,t) is a function giving the spatial amplitude profile of the wave, and s(x,y,t) is a function giving the spatial phase profile. Using different functions for s allows different types of critical point patterns to be generated. All patterns are specified with an initial location (x0,y0), and a constant velocity (vx,vy), given in grid spaces per time step, so the location of a pattern at time t is xt,yt=(x0+vxt,y0+vyt). For source or sink patterns, we use
ssourcesink(x,y,t)=(x−xt)2+(y−yt)2,
(14)
for spiral patterns we use
sspiral(x,y,t)=(x−xt)2+(y−yt)2+1katan2(y−yt,x−xt),
(15)
where atan2 is the multi-valued inverse tangent, and for saddle patterns we use
ssaddle(x,y,t)=|x−xt|−|y−yt|.
(16)
To ensure that all patterns are localized, we define A(x,y,t) as a symmetric two-dimensional Gaussian centered on the critical point location:
A(x,y,t)=A0exp(−((x−xt)2+(y−yt)2)2c2),
(17)
where A0 is the maximum amplitude and c is the Gaussian width parameter.
To generate complex datasets, we add multiple wave patterns and then add normally distributed white noise with mean 0 and standard deviation proportional to the amplitude of each grid point. For Fig 3, we used a 12×12×10 spatiotemporal grid to generate datasets comprising two random critical point patterns, both with parameters w = 2π×0.01 and k = 2π/5. Start positions, velocities, maximum amplitudes and Gaussian width parameters were randomized in each dataset: x0 and y0 were picked uniformly randomly but rejected if patterns were within 2 grid spaces of each other or an edge, vx and vy were picked randomly between −0.1 and +0.1, A0 was picked between 1 and 2 and c was picked between 3 and 5. Pattern detection in simulated data was performed with default parameters.
To demonstrate methods in NeuroPatt, we analyze recordings from the middle temporal area of adult male marmosets (Callithrix jacchus). Details of preparation are given previously [49,76]. Anesthesia and analgesia were maintained by intravenous infusion of sufentanil citrate (6–30 μg kg−1 h−1) and inspired 70:30 mix of N2O and carbogen (5% CO2, 95% O2). Dominance of low frequencies (1–5 Hz) in the EEG recording and absence of EEG or electrocardiogram changes under noxious stimulus (tail pinch) were taken as the chief signs of an adequate level of anesthesia. Drifts towards higher frequencies (5–10 Hz) in the EEG record were counteracted by increasing the rate of venous infusion or the concentration of anesthetic. The typical duration of a recording session was 48–72 h.
Stimuli were presented on a cathode-ray-tube monitor (Sony G500, refreshed at 100 Hz, viewing distance 45 cm, mean luminance 45–55 cd m−2), and comprised either a grey screen held at constant luminance for the duration of the recording (5–25 minutes, ongoing activity), or a pattern alternating every two seconds between a grey screen and a field of drifting circular white dots (Weber contrast 1.0; dot diameter 0.4°; drift velocity 20 deg/s) presented in a large, stationary circular window (30°). For dot fields, different motion directions (90° steps) were presented pseudo-randomly, and the procedure was repeated until 100 repetitions were made for each of the four directions. Data were recorded using multielectrode arrays (10×10 electrodes, 1.5 mm length, electrode spacing 400 μm, Blackrock Microsystems). Recording surface insertion depth was targeted to 1 mm.
To demonstrate pattern dynamics at a larger spatial scale and during waking activity, we also examine optical voltage recordings from awake mice, obtained with permission from Thomas Knöpfel. Details of recording have been described previously [50,77,78]. Briefly, excitatory neurons in mouse layer 2/3 were targeted with the gene encoding VSFS Butterfly 1.2 [79], and mice were implanted with a head post and thinned skull cranial window. Image acquisition was performed with a dual emission wide-field epifluorescence macroscope during anesthesia induced by pentobarbital sodium (40 mg/kg i.p.). The data presented here were taken as anesthesia was wearing off, when mice were responsive to touch and exhibited spontaneous whisker and limb movement between recordings. Image sequences of 60 s duration were acquired at 50 Hz temporal resolution and 320 × 240 pixel spatial resolution, with each pixel corresponding to 33 × 33 μm of a projected cortical area. The voltage imaging signal was calculated as the ratio of mKate2 to mCitrine fluorescence after equalization of heartbeat-related fluorescence modulation. The resulting ratiometric sequences of voltage maps were then spatially down-sampled by a factor of 5 using the MATLAB function imfilter, to smooth over noise and reduce the density of calculated velocity fields.
|
10.1371/journal.ppat.1002724 | The Wor1-like Protein Fgp1 Regulates Pathogenicity, Toxin Synthesis and Reproduction in the Phytopathogenic Fungus Fusarium graminearum | WOR1 is a gene for a conserved fungal regulatory protein controlling the dimorphic switch and pathogenicity determents in Candida albicans and its ortholog in the plant pathogen Fusarium oxysporum, called SGE1, is required for pathogenicity and expression of key plant effector proteins. F. graminearum, an important pathogen of cereals, is not known to employ switching and no effector proteins from F. graminearum have been found to date that are required for infection. In this study, the potential role of the WOR1-like gene in pathogenesis was tested in this toxigenic fungus. Deletion of the WOR1 ortholog (called FGP1) in F. graminearum results in greatly reduced pathogenicity and loss of trichothecene toxin accumulation in infected wheat plants and in vitro. The loss of toxin accumulation alone may be sufficient to explain the loss of pathogenicity to wheat. Under toxin-inducing conditions, expression of genes for trichothecene biosynthesis and many other genes are not detected or detected at lower levels in Δfgp1 strains. FGP1 is also involved in the developmental processes of conidium formation and sexual reproduction and modulates a morphological change that accompanies mycotoxin production in vitro. The Wor1-like proteins in Fusarium species have highly conserved N-terminal regions and remarkably divergent C-termini. Interchanging the N- and C- terminal portions of proteins from F. oxysporum and F. graminearum resulted in partial to complete loss of function. Wor1-like proteins are conserved but have evolved to regulate pathogenicity in a range of fungi, likely by adaptations to the C-terminal portion of the protein.
| Plant pathogenic fungi can have devastating effects on crop yield and quality. In addition, these fungi may generate mycotoxins that pose health risks when the contaminated crops are consumed. The pathogen Fusarium graminearum infects wheat heads and grows through the rachis by synthesizing trichothecene toxins. The mechanisms and environmental cues triggering the production of trichothecene toxins have been studied for many years. Here, we describe a fungal gene, Fgp1, that is absolutely required for pathogenicity and mycotoxin synthesis during infection and in culture. Fgp1 is not required for vegetative growth of F. graminearum but is important for reproductive development and potentially for a putative switch from a vegetative to a pathogenic phase. Deletion of Fgp1 results in reduced expression of the trichothecene biosynthetic gene cluster and genes specifically expressed during toxin synthesis in planta and in vitro. Fgp1 contains a conserved N-terminal domain and a divergent in the C-terminal region. The corresponding C-terminus from a sister species, F. oxysporum, is unable to function fully in F. graminearum when fused to the conserved F. graminearum N-terminal domain, suggesting the gene function is highly species specific.
| Pathogenic fungi have evolved sophisticated ways to infect their hosts, mainly by adapting to the host environment and by producing pathogenicity-related products such as toxic secondary metabolites, effector proteins and/or extracellular enzymes. The expression of genes involved in the adaptation to a host and synthesis of pathogenicity factors are under tight regulation to assure successful infection and survival. The pathogenic fungus Fusarium graminearum is a devastating pathogen of wheat and barley [1], [2] and modulates its pathogenicity largely by regulating a cluster of genes encoding enzymes for the biosynthesis of trichothecene toxins [3], [4], [5]. These mycotoxins are required during wheat infection to breach the rachis node of spikelets which acts as a barrier to systemic infection and maximal head blight symptoms. Toxin production during infection depends on multiple cellular and environmental factors (see recent review [6]) and yet, exactly how the genes for trichothecene biosynthesis are regulated is still largely unknown.
Another member of the Fusarium genus, F. oxysporum [7], contains both pathogenic and non-pathogenic strains. Pathogenic F. oxysporum strains modulate their pathogenicity in part by secreting small secreted proteins which may act both as virulence and as avirulence factors [8]. In the tomato wilt pathogen F. oxysporum f. sp. lycopersici, the nuclear protein Sge1 was demonstrated to be required for parasitic growth and expression of the small-secreted proteins genes SIX1, SIX2, SIX3 and SIX5, during conditions mimicking in planta growth [9]. A Δsge1 mutant was still able to colonize roots, but was unable to reach the normally infected xylem vessels. Recently an orthologous gene in the necrotrophic plant pathogen Botrytis cinerea, REG1, was shown to be required for infection of bean leaves [10], indicating that these genes may have a conserved role in fungal plant pathogenicity.
SGE1 and REG1 are orthologs of WOR1 from Candida albicans and RYP1 in Histoplasma capsulatum [11], [12]. In these human pathogenic fungi, both proteins are involved in the dimorphic switch, a transition correlated with the ability to cause disease [13], [14]. This suggests that the orthologous proteins found in plant pathogens also may be involved in switching from a saprophytic to a parasitic lifestyle. In this study we characterize the ortholog of WOR1 in F. graminearum, and show that it is absolutely required for pathogenicity and regulates the expression of the trichothecene biosynthetic (TRI) genes in planta as well as in vitro. It also apparently mediates a morphological change during toxin production, and, as determined by transcriptome analysis, regulates the expression of other genes, many of which are related to pathogenicity.
In order to study functional conservation of the orthologs from the two Fusarium species, we interchanged the genes and as well as swapped the C-terminal portions of the genes between the species. We found that the gene from the opposite species was not fully functional in the other species, nor were the chimeric genes. Additionally, very little overlap was found among genes regulated by each protein during vegetative growth. The results from these experiments show that the family of Wor1-like proteins regulates pathogenicity in many fungi through transcriptional reprogramming. Additionally, two members in two closely related Fusarium species have evolved different function, presumably in order to adapt to successfully infect their different hosts.
Most fungi have two homologous proteins encoded by the WOR1 gene family [9], which are recognized by a common GTI1/PAC2 domain, named after the WOR1-like genes GTI1 and PAC2 of Schizosaccharomyces pombe [15], [16]. Phylogenetically, the two paralogous genes from each species sort into two different clades [9] with e.g. GTI1, WOR1, RYP1 and SGE1 residing in one clade [9] and e.g. PAC2 from S. pombe and PAC2 from F. oxysporum in the other clade [9]. In F. oxysporum SGE1 is required for pathogenicity whereas PAC2 is not [9]. F. graminearum also has these two WOR1-like homologs. The gene that shows the highest similarity to SGE1 from F. oxysporum and aligns with the WOR1 clade is FGSG_12164 that we have named FGP1 (F. graminearum GTI1/PAC2 1). The predicted FGP1 gene is 1029 bp, intronless and encodes a 342 amino acid protein. The paralogous gene from F. graminearum that shows the highest similarity to PAC2 from F. oxysporum and aligns with the PAC2 clade is FGSG_10796 that we have renamed FGP2 (F. graminearum GTI1/PAC2 2). The predicted FGP2 gene is 1266 bp, intronless and encodes a 421 amino acid protein.
In order to assess the conservation of the Fgp1/Sge1 and Fgp2/Pac2 proteins in more Fusarium species, we obtained the sequences of the respective proteins from two other sequenced Fusarium strains, F. verticillioides and Fusarium solani f. sp. pisi (also known as Nectria haematococca), using the BLAST function [17] on the websites of the Broad Institute and the DOE Joint Genome Institute, respectively. As for F. oxysporum and F. graminearum, two genes were found for F. verticillioides as well as for F. solani. The genes that show the highest similarity to SGE1 and FGP1 and align with the WOR1 clade are FVEG_09150 and Fs_81912 of F. verticillioides and F. solani, respectively. The genes that show the highest similarity to PAC2 and FGP2 and align with the PAC2 clade are FVEG_11476 and Fs_60837 of F. verticillioides and F. solani, respectively.
When we aligned the four protein sequences of the Wor1 clade (Fo Sge1, FVEG_09150, Fg Fgp1 and Fs_81912), we found great divergence between the sequences of the four Fusarium proteins. Conservation is mainly restricted to the N-terminal portion (first ±220 amino acids in Figure 1A) of which the total similarity percentage ranges from 64–90.5% (Figure 1B). The C-terminal portions (last first ±140 amino acids, grey shaded in Figure 1A) of the Fusarium orthologs from the Wor1-clade are highly diverged despite of the high numbers of glutamine residues in all four sequences (5.6–13.6%, Table S1B). Overall sequence similarity is as low as 37% between F. graminearum and F. verticillioides. The sequences of the F. verticillioides ortholog and F. oxysporum Sge1 share the highest similarity; 65% (Figure 1B). On the contrary, when we aligned the four protein sequences of the Pac2 clade (Fo Pac2, FVEG_11476, Fg Fgp2 and Fs_60837), high similarities and conservation among the four Fusarium proteins (80–97%, Table S1A) are observed throughout the N-terminal and C-terminal regions (Figure S1).
The N-terminal portion of all the proteins from the four Fusarium strains in both clades contains the conserved WOPRa box (black line above the sequence, Figure 1 and S1) and WOPRb box (dashed black line above the sequence, Figure 1 and S1) [18]. The WOPRa and WOPRb boxes, previously recognized as the GTI1/PAC2 domain, have been shown to be involved in DNA binding of Wor1 in C. albicans. In the WOPRa box, a conserved threonine residue is present (boxed, Figure 1) that functions as a putative phosphorylation site. Mutation of this site in Gti1, Wor1 and Sge1 impairs the function of the respective proteins [9], [15], [19]. It has been shown previously that the N-terminal portion containing the WOPRa and WOPRb boxes from Fo Sge1 and Fo Pac2 align with the same domains from other proteins of the Wor1 family [18] and the same likely holds for the other Fusarium species because of the high conservation of these domains among the Fusarium proteins. The C-terminal portions of the Wor1-like proteins from Fusarium however have not only greatly diverged from each other but also from other fungal species. For example, the C-termini of the Fusarium proteins do not align at all with the C-terminus of Wor1 or Reg1 due to differences both in sequence and in length (data not shown).
Strains deleted for FGP1 or FGP2 were generated in F. graminearum using constructs produced in plasmid pPK2HPH-GFP [20] containing a cassette for a hygromycin B resistance gene fused at the C-terminus to GFP and regulated by a GPD1 promoter and TRPC terminator. In the respective plasmids, the cassette is flanked by the up- and downstream sequences (+/−2000 bp) of FGP1 or FGP2. Using Agrobacterium tumefaciens mediated transformation, multiple independent transformants were obtained of which five were selected for both FGP1 or FGP2. For each gene, all five proved by PCR and Southern blot to be homologous gene deletions, indicating that no ectopic transformants were obtained (Figure S2).
Wheat heads of the variety “Norm”, point-inoculated with the wild type (WT) strain PH-1 and four Δfgp1 strains were assessed after two weeks for the number of diseased spikelets. Wild type strain PH-1 was able to spread from the inoculated spikelet (arrow) to other spikelets in the head causing blight symptoms; bleached spikelets and deformed anthers (Figure 2B, left panel, lower head). In contrast, the Δfgp1 strains did not spread from the inoculated spikelet (arrows) to other spikelets (Figure 2A). In the inoculated spikelet, the Δfgp1 strains only cause minor disease symptoms; some browning of the palea (Figure 2B, left panel, upper four heads). When the inoculated spikelets were analyzed for trichothecene content, no toxin was detected in the spikelets inoculated with the Δfgp1 strains, in contrast to spikelets inoculated with the wild type, which accumulated significant levels of the toxins deoxynivalenol (DON) and 15-deoxynivalenol (15-ADON) (Figure 2C).
Reintroduction of the wild type gene FGP1 into a Δfgp1 strain resulted in complemented strains that were tested for spread through the wheat head and for toxin production in the inoculated spikelet. In total, four independent complemented strains used regained the ability to spread throughout the wheat head (Figure 2A) and cause disease symptoms similar to wild type; spread was present from the inoculated spikelets (arrows) to other spikelets in the head (Figure 2B, right panel). The four complemented strains also regained the ability to produce toxin similar to wild type (Figure 2C).
In contrast, Δfgp2 strains were not reduced in the ability to cause disease. The Δfgp2 strains were able to cause head blight symptoms almost to the same extent as wild type PH-1 (Figure 2D) and accumulated mycotoxin levels comparable to wild type (Figure 2E). These results affirm previous observations that genes from the WOR1 clade are involved in pathogenicity whereas genes from the PAC2 clade are not and suggest that WOR1 orthologs may regulate secondary metabolite synthesis.
As previously mentioned, no spread of disease symptoms was observed with the Δfgp1 mutant compared to wild type. To determine the reason for this, we investigated whether the symptoms in the Δfgp1 and Δfgp2 mutants are correlated to the growth of the mutant strains within the wheat head, using bright light and fluorescent microscopy. Since no differences in symptoms were observed between wild type PH-1 and Δfgp2, the latter was used as a fluorescent positive control strain. After inoculation, spread of the Δfgp1 and Δfgp2 strains were monitored through the spikelet for two weeks. At three days after inoculation, “fluffy” mycelium growing from of the inoculated spikelet was observed for the Δfgp1 strain to the same extent as wild type (data not shown) and the Δfgp1 strain had infected the palea and a portion of the lemma but had not spread towards the glume on the outer part of the spikelet (Figure 3A). Even after two weeks, the glume was still green and apparently healthy. Presence of the fungus in the palea and lemma is manifested by browning of the tissue and was verified by microscopy (data not shown). In contrast to Δfgp1, the Δfgp2 strain had infected the palea, the lemma and had spread to the glume after three days (Figure 3B). For both Δfgp1 and Δfgp2 strains, GFP is coupled to the constitutively expressed hygromycin B resistance gene HPH, making it possible to follow the growth inside the plant using hyphal fluorescence. Doing so, we identified the rachis node as the flower tissue where growth of the Δfgp1 mutant was halted. No fluorescence from the Δfgp1 mutant was found in the rachis node (Figure 3C, red circle) nor was fluorescence detected beyond the rachis over the complete time period of two weeks. Growth of Δfgp1 was observed in other spikelet parts after three days (Figure 3C, white arrow). In contrast, fluorescence of the Δfgp2 mutant was found in the flower (arrow) as well as in the rachis node and beyond after three days (Figure 3D, red circle). The inability of the Δfgp1 mutant to penetrate the rachis node and the consequential pathogenicity loss might be caused solely by the lack of trichothecene production by the fgp1 mutant. This inference is made because Δtri5 mutants of F. graminearum, lacking the gene for the first enzymatic step in toxin synthesis, are unable to produce trichothecene toxins in planta and also are stalled during infection at the rachis node [21].
To determine whether the inability to cause disease could be correlated to a major growth or developmental deficiency, the growth characteristics of Δfgp1 strains were assessed on different media. Germination rate was assessed by placing freshly produced spores on PDA agar. Spores from WT PH-1 show almost complete (≥95%) germination after 8 hours, while Δfgp1 spores show a slight delay with complete (≥95%) germination observed after 12 hours (data not shown). On PDA and minimal medium, the deletion of FGP1 in wild type results in a slightly reduced radial growth phenotype, which can be attributed to the delayed germination. Complementation of the Δfgp1 strain with the wild type FGP1 gene restored the wild type growth phenotype. No growth differences were observed on complex solid media including V8, carrot or mung bean agar plates (data not shown).
Fgp1 is required for full asexual spore formation as was observed for Sge1 in F. oxysporum [9]. During growth on mung bean agar (MBA), fewer (p≤0.05, Student's t-Test) macroconidia are formed in four independent Δfgp1 deletion strains compared to wild type (Figure 4A). Additionally, spores of the Δfgp1 strain are smaller during growth on mung bean agar (MBA) (p≤0.05, Student's t-Test) or in carboxymethylcellulose (CMC) (p≤0.01, Student's t-Test) (Figure 4B) or may exhibit precocious germination or appear not fully developed (three middle pictures, Figure 5C representative spores are shown). In the complementation strain, the quantity of spores is restored, albeit incompletely, still, the spores produced resemble wild type (Figure 4A, B and C, representative spores of wild type and complemented strain are shown). Using Calcofluor white and Hoechst staining, no defects in cell wall composition or nucleus quality were observed in the Δfgp1 strain (data not shown).
When strains were grown on carrot agar to induce perithecium and ascospore formation, it was observed that perithecia formed by the Δfgp1 strain are comparable to wild type (data not shown). Ascospore formation on the other hand was delayed by one week in the Δfgp1 strain and altogether only a few ascospores emerged from the perithecia. Cirrhi of wild type PH-1 and a FGP1 complementation strain appear one week after the formation of the perithecia and consist of strings of ascospores emerging from perithecia (data not shown). Cirrhi of the FGP1 deletion strains appear two weeks after the formation of the perithecia and contain only a few ascospores at the perithecial apex (data not shown). When the ascospores were harvested and counted, fewer ascospores (p≤0.01, Student's t-Test) were produced by the Δfgp1 strains compared to wild type (Figure 4D). This defect was restored in the complementation strains (Figure 4D). These observations indicate that FGP1 is involved the developmental processes of conidiogenesis and ascospore formation although these processes are not fully abolished in the deletion mutant. As a consequence of this impaired conidiogenesis, the Δfgp1 strain probably displayed delayed germination on certain media. On the other hand, Fgp1 has no impact on normal vegetative growth.
Since the Fgp2 homolog Pac2 is involved in the sexual cycle of Schizosaccharomyces pombe, we also tested whether perithecium and ascospore formation was altered in the Δfgp2 strain. No major differences with respect to perithecium and cirrhi formation were noted when compared to wild type (data not shown).
Previous studies have shown that F. graminearum produces trichothecene toxins when grown on medium containing polyamine compounds [22] and that this medium induces the expression of genes involved in trichothecene production [5]. Since trichothecene toxins did not accumulate in wheat spikelets inoculated with Δfgp1, we investigated whether deletion strains also are impaired in production of toxin in vitro. In order to do this, the wild type, four independent Δfgp1 strains and two independent complemented strains were inoculated into putrescine containing medium and after one week assessed for the presence of trichothecene toxins. Toxin was found in wild type cultures but no toxin was detected in the cultures of the Δfgp1 strains (Figure 5A). Toxin was also found in the cultures of the two complemented strains albeit to slightly lower levels than wild type (Figure 5A).
The lack of toxin production in the Δfgp1 strains could be due to the inability to express the genes from the TRI cluster. In order to test for this, a northern blot experiment was performed. Both wild type PH-1 and a Δfgp1 strain were grown in minimal medium containing the polyamine putrescine or in control minimal medium with NaNO3 as the sole nitrogen source instead of putrescine. Samples for RNA isolation were taken 8, 16, 24, 32, 40 or 48 hrs after inoculation and the RNA was used for gel electrophoresis and capillary blotting. Northern blots were hybridized with TRI14, a gene from the trichothecene biosynthetic cluster [23] or with a constitutively expressed actin gene used as the loading control. TRI14 is expressed to high levels during growth in polyamine medium [24] and is therefore a good marker gene for the expression of the TRI cluster. No TRI14 transcript was observed in the Δfgp1 mutant strain grown in putrescine medium at any time point in contrast to wild type PH-1 in which expression of TRI14 was detected at 32, 40 and 48 h (Figure 5B). RNAs corresponding to TRI14 were not detected in cultures grown in minimal medium for either wild type or Δfgp1 (data not shown). Actin transcripts were detected in all samples of wild type and Δfgp1, indicating equal loading.
Analysis of the cultures filtrates of the different samples revealed that trichothecene toxins were present in the putrescine medium containing wild type PH-1 at 32 h and accumulated to higher levels at 40 and 48 h (Figure 5B). No toxin was detected for the Δfgp1 strain grown in putrescine medium at any time point nor in control medium, which is consistent with the previous toxin analysis experiment and the lack of TRI14 gene expression. The Δfgp1 strain exhibits growth equal to wild type in putrescine medium (data not shown), suggesting that it likely can take up putrescine and use it as a nitrogen source but this does not lead to TRI gene expression. In minimal medium lacking any nitrogen source, growth of both wild type and the Δfgp1 strain was noticeably reduced and no toxin was detected (data not shown). This suggests that TRI gene expression is not triggered by the absence of a nitrogen source but rather that putrescine is a specific inducer. Clearly, Fgp1 is required for putrescine induced trichothecene production.
The full HPLC spectra of the wild type and Δfgp1 samples grown for 40 hours in putrescine medium were, besides used for trichothecene toxin identification, also inspected for other differences in metabolite profiles. Doing so, no significant peaks other than the peaks corresponding to DON and 15-ADON are missing in the spectrum from the Δfgp1 strain compared to the wild type (Figure S3). This suggests that Δfgp1 can still produce a majority of the other metabolites present in the spectrum and that it does not appear to be impaired in production of metabolites in general.
To examine in more detail why TRI gene expression was not observed in the Δfgp1 strain, wild type PH-1, Δfgp1 and complemented strain were grown in putrescine medium and examined microscopically. In all strains, a distinct morphological change in hyphae was observed when cultures were grown in putrescine medium for 40 h compared to control medium. In control medium, hyphae display a uniform thickness over their entire length and grow in long branches in wild type, Δfgp1, and complemented strain (Figure 6, upper panel). In putrescine medium, the wild type and complemented strains display hyphae that produce bulbous sub-apical structures, which are to some extent also observed in the Δfgp1 strain (Figure 6, center panel). In wild type and complemented strain, bulbous structures up to 21 µm in diameter are observed among numerous other, but fewer and smaller bulbous structures are observed in the Δfgp1 strain (Figure 6, lower panel). Another independent Δfgp1 and complemented strain displayed similar growth morphology as the Δfgp1 and complemented strain given in Figure 6.
To investigate whether this morphological change in the putrescine medium parallels the expression of TRI genes and trichothecene accumulation we also examined the cultures at the different time points. After 16 and 24 h, some hyphae begin to swell and form bulbous structures in both wild type and Δfgp1 strains (see arrows, Figure S4). These structures become more abundant at 32 h in the wild type but not in the Δfgp1 strain (see arrows, Figure S4). At 40 and 48 h, large bulbous structures appear in the wild type which are less apparent in the Δfgp1 strain (see arrows, Figure S4). This suggests that the formation of these bulbous structures might be involved in the production of trichothecene toxins and that somehow the Δfgp1 strain is less capable to form the same structures as in wild type.
To examine whether the bulbous structures observed in wild type are formed as a consequence of toxin production or whether the structures may accommodate toxin production, a Δtri6 mutant was grown in putrescine and control minimal medium. TRI6 encodes a transcription factor that is absolutely required for the expression of the genes in the TRI cluster and for trichothecene synthesis [25]. The Δtri6 strain lacks detectable trichothecene accumulation in the putrescine medium (data not shown), but the hyphae show the same degree of hyphal swelling and bulbous structure formation as the wild type (data not shown). We conclude from this that the bulbous cells do not form in response to trichothecene synthesis because they appear to develop before transcription of the TRI genes, and the toxins themselves, can be detected. It remains unclear what the exact role this hyphal swelling plays in toxin synthesis and how Fgp1 might be involved in the facilitation of this process.
To investigate the underlying genetic basis for the differences in toxin accumulation and hyphal morphology seen in putrescine medium, RNA was extracted from wild type PH-1 (three independent samples) and Δfgp1 strains (three independent samples) after growing 40 h in putrescine and labeled for microarray experiments. Additionally, in order to find genes regulated by Fgp1 during in planta growth, RNA was extracted from wheat heads inoculated for 72 hours with wild type (three independent samples, three heads per sample with ten inoculated spikelets per head) or Δfgp1 (three independent samples, three heads per sample with ten inoculated spikelets per head) and labeled for microarray experiments. The number of genes showing a >2-fold difference in average expression between wild type PH-1 and Δfgp1 strains during growth on putrescine medium and during wheat infection were determined (Figure S5). Fgp1 appears to regulate gene expression positively as well as negatively during growth in putrescine medium and infection of wheat with hundreds of genes differentially expressed. In total, 654 genes show a >2.0-fold (P<0.05) lower expression in Δfgp1 compared to wild type when grown in putrescine and 536 genes show higher expression in Δfgp1 (Table S2). Additionally, 634 genes were ≥2-fold lower expressed (P<0.05) in the Δfgp1 strain compared to wild type during wheat infection and 39 genes were found >2 fold higher expressed (Table S3).
Some overlap is found between genes positively or negatively regulated by Fgp1 in putrescine medium and those regulated during wheat infection (91 and 3 genes, respectively (Figure S5)). Recently, a gene set considered to be expressed “in planta only” was described containing genes from F. graminearum exclusively expressed during plant infection but not in axenic cultures [26]. Out of the 369 genes from this gene set, 243 were found expressed in the experiments described here, either in wild type or the Δfgp1 strain grown in putrescine medium (45 genes) or in infected wheat (199 genes) (Table S4). Overlap was found between these 243 “in planta only” genes and genes lower or higher expressed in the Δfgp1 strain in putrescine medium or during wheat head infection (Figure S5). Altogether, Fgp1 regulates a remarkably large proportion of genes considered “in planta only”; 99 out of 244 or 41% were positively regulated and another nine (3.7%) were negatively regulated.
Fgp1 positively regulates thirteen “in planta only” genes in putrescine as well as in wheat heads. Among these thirteen are six genes from the trichothecene biosynthetic cluster (FGSG_03534, FGSG_03536, FGSG_03537, FGSG_03538, FGSG_03540 and FGSG_03543) as well as four other genes involved in trichothecene toxin synthesis or co-expressed with TRI5 [24] (FGSG_00007, FGSG_01819, FGSG_07562 and FGSG_10397). Among the remaining three are FGSG_08079 a gene involved in butenolide synthesis, FGSG_08309 encoding an ABC transporter with orthologs in other fungi and FGSG_02120 for which there is no annotated function or ortholog found in any other fungi with a sequenced genome. Fgp1 negatively regulates one “in planta only” gene expressed in putrescine medium and during infection: FGSG_11033.
TRI6 (FGSG_03536) encodes a transcription factor that regulates the expression of other TRI genes and genes involved in the isoprenoid biosynthesis pathway. Since TRI6 and an ancillary transcription factor, TRI10, are both extremely down-regulated in the Δfgp1 strain, genes positively regulated by Tri6 and Tri10 might be expected to be expressed at lower levels in the Δfgp1 strain. Indeed, 56 and 62 genes expressed at lower levels in the Δfgp1 strain in putrescine and wheat head, respectively (Table S5), were also found to be expressed at lower levels in the Δtri6/tri10 strains 96 h after wheat inoculation. Among the common regulated genes are the ones involved in the isoprenoid biosynthesis pathway. This suggests that Fgp1 may assume a portion of its regulatory control by acting upstream of the Tri6 and Tri10 transcription factors.
As co-regulated clusters of genes often are associated with secondary metabolite biosynthesis, expression of putative gene clusters from F. graminearum identified previously [27] was examined using the expression values obtained from the microarray experiment. Unlike in wild type, expression of all genes involved in trichothecene biosynthesis was absent or was greatly reduced in the Δfgp1 strain during growth in putrescine medium as well as during wheat infection (Figure 7A). This loss of expression was confirmed by the absence of TRI14 transcript in the northern blot experiment (Figure 5) and lack of the trichothecene toxins produced by the Δfgp1 strain. In addition to the twelve genes from the TRI cluster, 19 of the 38 genes that showed co-expression with TRI5 in agmatine medium [24] are also regulated by Fgp1, either during growth in putrescine medium or during wheat infection (Table S6). Fgp1 negatively regulates two of the 19 genes: FGSG_01832 and FGSG_03132. Five of the 17 genes that are positively regulated by Fgp1 are also regulated by Tri6 (Table S6).
Genes involved in butenolide synthesis, also reside in a cluster [28]. Other than FGSG_08077, FGSG_08078 and FGSG_08079, expression of genes from the butenolide cluster was not detected in putrescine medium for both wild type and Δfgp1. However, during wheat infection expression of all genes from the cluster were detected on the microarray chip in wild type and in the Δfgp1 strain, albeit to a significantly lesser extent for the mutant (Figure 7B).
A cluster of genes involved in the production of the yellow or purple pigment aurofusarin [29] as well as genes comprising the non-ribosomal protein synthase 8 (NPS8) [24] gene cluster also appear to be regulated by Fgp1. In the Δfgp1 strain, the aurofusarin cluster genes are up-regulated when compared to wild type suggesting a role for Fgp1 in the negative regulatory control of this cluster (Figure 7C). The genes of the NPS8 cluster are highly down-regulated or not detected in putrescine medium in the Δfgp1 strain when compared to wild type (Figure 7D). For both gene clusters no gene expression was detected during wheat infection.
Fgp1 also regulates other genes that encode polyketide synthases (PKS) [30] and non-ribosomal protein synthases (NPS) [31] that may be involved in secondary metabolite synthesis. During wheat infection NPS9 (FGSG_10990), NPS12 (FGSG_11294), NPS14 (FGSG_11395), PKS7 (FGSG_13295) and PKS15 (FGSG_04590) are expressed at lower levels in the Δfgp1 mutant compared to wild type. During growth in putrescine medium NPS11 (FGSG_03245) is expressed at lower levels in the Δfgp1 mutant. Higher expression was observed for NPS18 (FGSG_13783) in the Δfgp1 mutant during growth in putrescine medium. Transcriptome experiments therefore show that Fgp1 regulates gene clusters and genes predicted to encode secondary metabolite synthesis proteins and genes previously found solely expressed during plant infection.
To determine whether Fgp1 and Sge1 share common regulatory targets, a comparative transcriptional study was conducted during growth in culture. Both wild type strains Fol4287 and PH-1 as well as deletion strains Δsge1 and Δfgp1 from F. oxysporum and F. graminearum, respectively, were grown in complete medium (CM) for 48 h. RNA was extracted, reverse-transcribed into cDNA, labeled and hybridized to microarray chips. In F. graminearum, 119 genes show a ≥2 fold lower expression in the Δfgp1 strain on CM and 83 genes a higher expression compared to WT (Table S7). For F. oxysporum, 394 genes show a ≥2 fold lower expression in the Δsge1 strain on CM and 819 a higher expression compared to WT (Table S7). When the gene sets regulated by either Fgp1 or Sge1 were compared, only a few orthologous genes were found to be regulated by both Sge1 and Fgp1 in the two Fusarium species (Figure S6). A set of 16 down-regulated and eight up-regulated orthologous genes were found in both deletion strains (Table S8). Remarkably, eleven genes up-regulated in Δsge1 were found down-regulated in Δfgp1 and one gene up-regulated in Δfgp1 was found down-regulated in Δsge1 (data not shown). Altogether, the results from this comparative transcriptomic experiment suggest that Fgp1 and Sge1 regulate largely non-overlapping sets of genes.
Because of the effect of the Δsge1 mutation on sporulation in F. oxysporum it was noteworthy that there was differential regulation between wild type and the Δsge1 mutant for several genes known to effect sporulation in Fusarium and other fungi. REN1 [32], an ortholog of MEDUSA [33] from Aspergillus nidulans, ABA1, an ortholog of ABACUS [34] from A. nidulans and FOXG_01756, an orthologs of FlbC [35] from A. nidulans, were among the genes expressed at lower levels in the Δsge1 mutant. The lower expression of these genes could account for the fewer number of spores produced by the Δsge1 mutant [9]. F. graminearum does not produce conidia in CM, so, in order to study these genes, we grew wild type and Δfgp1 in liquid carboxymethylcellulose (CMC) medium or mung bean agar (MBA), two media that induce sporulation. Quantitative PCR experiments were conducted using cDNA obtained from F. graminearum wild type and Δfgp1 and from F. oxysporum wild type and Δsge1 strains during growth in conidia inducing medium. We used the constitutive expressed gene FRP1 [36] as a reference in both F. oxysporum and F. graminearum. The expression levels of FRP1 proved not to be influenced by deletion of SGE1 or FGP1 (results not shown). The results showed that in F. oxysporum all three genes (REN1, ABA1 and FOXG_01756) were expressed at lower levels in the Δsge1 strain, confirming the microarray results, but in the Δfgp1 strain of F. graminearum, only the expression of ABA1 was significantly reduced during growth on MBA and CMC compared to wild type (Figure S7); the expression of REN1 and FGSG_ 07052 (FLBC ortholog) were not significantly lower; both genes even show a slightly higher expression in the Δfgp1 strain compared to wild type when grown on CMC. However, the expression of FGSG_ 07052 was lower in the Δfgp1 strain compared to wild type when grown on CM (Figure S7). These results suggest that Sge1 regulates a number of sporulation genes and that Fgp1 only regulates Aba1 during conidia formation. This expression difference may be attributable to the difference in spores that are formed under these conditions: microconidia in F. oxysporum versus macroconidia in F. graminearum.
To study the functional conservation of FGP1 and SGE1, the two genes were introduced in the opposite species in order to test whether they can take over each other's function. Doing so, SGE1 was introduced into the Δfgp1 F. graminearum strain and FGP1 into the Δsge1 F. oxysporum strain. Two independent transformants that carry the SGE1 complement in the Δfgp1 strain in F. graminearum were obtained and were used to inoculate wheat spikelets. In contrast to the transformants harboring the FGP1 complement, both SGE1 complements were unable to restore disease causing ability on wheat heads (Figure 8A, bars below the FGP1 box and SGE1 box, respectively). This suggests that SGE1 has diverged from FGP1 in such way that it is unable to restore its function.
This observation led to the hypothesis that the functional specificity of the protein may be present in the highly diverged C-terminal portion of the protein. To test this, combinations of the two genes from F. oxysporum and F. graminearum were made and transformed into a Δfgp1 strain of F. graminearum and into a Δsge1 strain of F. oxysporum. A combination consisting of the N-terminal portion of Fgp1 (amino acids (aa) 1–219) and the C-terminal of Sge1 (aa 220–330), expressed from the native FGP1 promoter was used to complement the Δfgp1 strain. The strains obtained showed less spread (p≤0.01, Student's t-Test) and disease in inoculated wheat spikelets after two weeks compared to the full length FGP1 complements but more disease than complements with full length SGE1 (Figure 8A, bars below the FGP1/SGE1 box). This suggests that the highly diverged C-terminus of Fgp1 is required for full function in F. graminearum. Reintroduction of the complete wild type gene FGP1 of F. graminearum (Figure 8B, bars below the FGP1 box) used to complement the Δfgp1 strain into a Δsge1 strain of F. oxysporum failed to restore pathogenicity towards tomato in contrast to complementation with full length SGE1 (Figure 8B, bars below the SGE1/FGP1 box). The combination of the N-terminal domain of SGE1 (encoding aa 1–218) and the C-terminal domain of FGP1 (encoding aa 219–342) expressed from the native SGE1 promoter also failed to complement the mutant phenotypes in the Δsge1 strain. This suggests that the diverged C-terminal portions of the genes are critical for their function in the species of origin.
In this study, the Wor1-like protein Fgp1 from F. graminearum was demonstrated to greatly influence plant infection and trichothecene mycotoxin production. The Fgp1 ortholog Wor1 in C. albicans also regulates pathogenicity by controlling dimorphic switching essential for mammalian infection. By analogy we hypothesize that Fgp1 and other orthologs like Sge1 (F. oxysporum) and Reg1 (Botrytis cinerea) in phytopathogenic fungi may also act as master regulators of plant infection. Deletion of the WOR1 orthologs in these fungi does not cause obvious vegetative growth defects yet the deletions appear to lock the mutants in a nonpathogenic state. Since the diseases caused by these fungi are vastly different in their mechanisms of infection, pathogenesis and tissue- and host specificity, the orthologous Wor1 regulatory proteins must also have evolved to regulate these unique aspects of their disease causing ability.
This study is the first to demonstrate that a Wor-1 like protein has the ability to control the expression of genes for mycotoxin biosynthesis and well as other gene clusters for synthesis of fungal secondary metabolites such as aurofusarin and butenolide. Fgp1 positively regulates the TRI cluster but negatively the AUR cluster responsible for aurofusarin production. This phenotype is also seen in a heterochromatin protein encoding gene deletion mutant of F. graminearum, Δhep1 [37]. The similarity in phenotype between Δfgp1 and Δhep1 could suggest that Fgp1 regulates chromatin modification too. Trichothecene toxins produced by F. graminearum are central to its pathogenicity to wheat [21]. These toxins normally are synthesized during pathogenic growth and are induced within specific plant parts [3]. As trichothecene biosynthesis is strongly under the control of Fgp1, especially through the transcription factor genes TRI6 and TRI10, the protein must have evolved to regulate the genes for mycotoxin synthesis that are peculiar to this disease. A morphological change that accompanies trichothecene accumulation and gene expression is noticeably altered in the Δfgp1 mutant, suggesting that Fgp1 may be involved in a morphological change required for pathogenicity. A similar morphological change occurs during infection of wheat; in planta, F. graminearum forms thickened hyphae and coralloid structures that resemble the bulbous hyphae that are observed in putrescine medium [38]. This morphological phenomenon has been associated with toxin biosynthesis gene expression in planta in various studies [38], [39], [40], however, the necessity of the morphological change for toxin accumulation remains to be determined.
Using microarray analysis many putative downstream targets of Fgp1 were identified. Among these are several pathogenicity related genes [26], including ones previously described as infection specific, such as genes from the TRI cluster and others involved in toxin production [24]. We also found many putative downstream targets of Fgp1 and of its F. oxysporum counterpart Sge1 when grown on CM. Both proteins show a high degree of specificity towards their putative targets as very little overlap was found between orthologous genes in the two species regulated by either Fgp1 or Sge1. Additionally, of the hundreds of genes that are regulated by Δfgp1 in putrescine medium and during wheat infection, only eight are also regulated by Fgp1 during growth in CM (Table S9), suggesting that Fgp1 regulates specific sets of genes during toxin induction conditions (putrescine or wheat head) or during growth in rich, toxin non-inducing medium (CM).
A few target genes of Sge1 and Fgp1 identified in this study and confirmed by quantative PCR analysis, are involved in sporulation and some may also play a role during pathogenicity. For example, Sge1 regulates the expression level of REN1, a conserved gene required for adherence, biofilm formation and virulence in A. fumigatus [41] and absolutely required for microconidia formation but not for pathogenicity in F. oxysporum [32]. The lower expression level of REN1 in Δsge1 might explain, at least partly, the lower number of microconidia produced in CM. F. graminearum does not produce microconidia under any condition tested and perhaps as a consequence, no difference in REN1 expression is observed between wild type and Δfgp1 strains during macroconidia formation. Sge1 also regulates the expression level of ABA1, another gene involved in conidium formation. In F. oxysporum f. sp. melonis, a pathogen of melon, Aba1 regulates production of both micro- and macroconidia and is required for full virulence. The Δaba1 mutant of this strain shows delayed pathogenicity towards melon probably due to fewer spores produced inside the xylem vessels (personal communication Dr. Tsutomu Arie, Tokyo University of Agriculture and Technology, Japan). In addition to sporulation, the Aba1 ortholog in Penicillium marneffei is involved in the dimorphic switch [42] and in A. fumigatus in autolysis and cell death [43]. The expression levels of ABA1 of F. graminearum are also regulated by Fgp1 during conidia formation but not during growth in putrescine medium or infection of wheat heads. The role of ABA1 during infection is therefore still elusive. Another gene known to control conidia formation in fungi that is regulated by Sge1 is FlbC. In A. nidulans, FlbC regulates the developmental processes of conidia formation and sexual fruiting and is required for normal vegetative growth [35]. To the contrary, we found that Fgp1 does not regulate the expression of FLBC of F. graminearum during conidia formation. Instead we found that expression of FLBC is regulated by Fgp1 during growth on CM and putrescine medium. Interestingly, FlbC in F. graminearum is absolutely required for wheat infection but not for toxin production [44]. How FlbC regulates virulence independent of toxin production and whether and how FLBC is regulated by Fgp1 during wheat infection is not yet known. FLBC is an interesting candidate gene to investigate further regarding its role in pathogenicity and its putative function downstream of both Fgp1 and Sge1.
A transcription factor not involved in sporulation but nevertheless regulated both by Fgp1 and Sge1 during growth on CM is DAL81. Dal81 is a general activator of nitrogen metabolic genes in yeast, including those for γ-aminobutyrate (GABA) [45]. Its exact role in Fusarium species is not yet known but this transcription factor (FGSG_02068) is required for toxin production in F. graminearum but, paradoxically, apparently not for virulence [44].
Whether Fgp1 and Sge1 are transcription factors and bind DNA directly is not known. Their conserved N-terminal regions contain putative DNA binding domains, previously called the GTI1/PAC2 domain [15], [16] but recently renamed the WOPR box (Wor1, Pac2 and Ryp1) [18]. This motif consists of two globular peptide domains: WOPRa and WOPRb. Via these two domains Wor1 is able to bind a 14-bp DNA sequence and thereby activates its target genes and itself via a positive feedback loop [18]. Of the fourteen base pairs, the ones located in positions 6 through 14 (TTAAAGTTT) are absolutely required for binding. By scanning the upstream regions of both FGP1 and SGE1, variations to the WOR1 motif were found 557 upstream of the ATG of FGP1; TTAAAGTTC and 644 bp upstream of the ATG of SGE1: TTAACGCTT. Whether these domains in the promoters of FGP1 and SGE1 are DNA binding sites required for FGP1 and SGE1 expression is unknown. However, a search conducted for these patterns in upstream regions in both the genomes did not unveil any enrichment for this pattern in the genes found up- or down-regulated by either Sge1 or Fgp1 (data not shown). Likewise, a specifically conserved DNA pattern was not found upstream of genes regulated by Fgp1 or Sge1 under different conditions using search engines like RSAT [46].
The evolution of the Wor1-like proteins involved a duplication of the ancestral Wor1 like gene sequence prior to the divergence of the yeast -like and filamentous ascomycetous fungi. The duplication resulted in paralogous genes (FGP1/WOR1 and FGP2/PAC2) that apparently have evolved quite different regulatory functions. While FGP2/PAC2 orthologs have been shown to be dispensable for pathogenicity, several studies now have established a role for FGP1/WOR1 in pathogenicity for a surprisingly diverse array of fungal pathogens of both plants and animals [9], [10], [11], [12]. The function of FGP1/WOR1 however, clearly extends beyond its role in pathogenicity since transcriptome studies demonstrate its regulatory control over many other functions such as reproduction and secondary metabolism. FGP1/WOR1 and FGP2/PAC2 orthologs also are found in strictly non-pathogenic fungi like Podospora anserina (data not shown); however, interestingly, in the non-pathogenic fungus Neurospora crassa only the FGP2/PAC2 ortholog is present.
Unlike the N-termini, the divergent glutamine-rich C-termini of FGP1/WOR1 genes in Fusarium likely allow specificity of regulatory control that has evolved independently in each species. Indeed the sets of genes regulated by orthologs FGP1 and SGE1 in F. graminearum and F. oxysporum, respectively, show little overlap and swapping experiments indicate the function of the C-terminal domains is largely specific to the species of origin.
How WOR1/FGP1 genes are regulated themselves is still elusive although the mitogen activated protein kinase (MAPK), as well as the protein kinase A (PKA) pathway could be involved. In B. cinerea, REG1 expression levels are regulated by two mitogen activated kinases: BcSAK1 and Bmp3 [10]. For SGE1, higher expression levels (±5-fold) are observed during in planta growth compared to growth in axenic culture [9], which might indicate that SGE1 is regulated through expression levels too, but for FGP1 no significant differences in expression levels were observed during the conditions tested (data not shown). In F. oxysporum and F. graminearum, both mitogen activated kinases are required for pathogenicity [47], [48] of which the Δgpmk1 mutant in F. graminearum lacks the ability to form bulbous infection hyphae in planta [38] which might indicate that this strain is also defective in production of the bulbous hyphea in putrescine medium. In C. albicans, Wor1 phosphorylation and subsequent activation is believed to be performed by Tpk2, a subunit of the PKA [19]. The conserved phosphorylation site of the Wor1-like proteins resembles a PKA site, making it likely to be phosphorylated by PKA. But whether Sge1 or Fgp1 are also phosphorylated and which kinase may be responsible for that is still unknown. However, a PKA mutant in F. oxysporum (ΔfocpkA) is also impaired in root penetration and virulence [49].
In C. albicans, Wor1 negatively regulates Efg1 transcription levels in opaque cells directly and indirectly via Czf1 and is itself regulated by Wor2 [18], [50]. The conserved EFG1 ortholog in Fusarium species is called STUA and has been studied in F. graminearum (FgSTUA) [51] and F. oxysporum (FoSTUA) [52]. Czf1 and Wor2 have no conserved ortholog in Fusarium species as sequences homologous to these genes cannot be located using low stringency BLAST searches of the Fusarium genomes. The expression level of FGP1 seems to be negatively regulated in F. graminearum by StuA. During growth of the ΔfgstuA mutant in CMC and in a two-stage toxin induction medium levels of FGP1 transcripts are ±10 and ±100-fold higher compared to wild type. During growth of the ΔfgstuA mutant on wheat head, on the other hand, no significant difference in FGP1 levels were observed [51]. In contrast, FgSTUA or FoSTUA expression levels are not significantly different from wild type in Δfgp1 and Δsge1 mutants, respectively ([9] and data not shown). The ΔfostuA mutant is still able to infect its host but the ΔfgstuA mutant is impaired in pathogenicity and toxin production [51]. In both F. graminearum and F. oxysporum, StuA is involved in conidia formation [51], [52]. Overall, the ΔfgstuA demonstrates a more severe phenotype than the Δfgp1 mutant with greatly reduced vegetative growth and spores almost entirely absent [51]. These observations suggest that regulation of STUA and FGP1 in Fusarium species occurs differently than their orthologs WOR1 and EFG1 in C. albicans.
In-depth phosphorylation experiments and expression studies of FGP1 and SGE1 with different mutant strains will be needed to identify other putative upstream activation factors. Additional work also will be required to fully understand the divergent roles of the N- and C- terminal domains on protein function and target specificity. Lastly, further investigations into the genome-wide impact of Fgp1 and Sge1 regulation on cell homeostasis, spore development and host infection will be needed to grasp a better understanding of this very interesting protein family.
The fungal isolates used in this study are the sequenced strains of F. graminearum PH-1 and F. oxysporum f.sp. lycopersici (Fol) strain 4287. Also used were the Fol Δsge1 strain SGE1KO4 and complementation strain SGE1com79 reported earlier [9]. All fungal strains were kept at −80° and revitalized on potato dextrose broth plus agar (PDB and Bacto agar, Difco). F. graminearum strains were grown for five days in carboxymethylcellulose (CMC) medium and F. oxysporum in rich complete medium (CM) for macroconidia and microconidia production, respectively. Agrobacterium tumefaciens EHA105 [53] used for Agrobacterium mediated transformations was grown in LB containing 20 µg/ml rifampicin and at 28°C.
The wheat varieties “Norm” and “Bobwhite” and the wilt susceptible tomato variety “Bonny Best” (Reimer Seeds, North Carolina USA) were used for plant infection studies. Wheat pathogenicity assays were performed with cultivar “Norm” using point inoculation [54]. Pathogenicity was scored two weeks after inoculation by counting the number of infected spikelets. Wheat infection used for microarray studies were performed with cultivar “Bobwhite” using point inoculations of 10 spikelets in the head. 72 hour after inoculation, anthers were cut off and the infected spikelets were detached from rachis, collected and frozen until RNA processing [25]. Tomato infections were performed with two-week-old seedlings sown in vermiculite and given 20-20-20 NPK fertilizers after one week. Seedlings were inoculated using the root dip method and disease was scored as described previously [9].
In order to generate deletion constructs of FGP1 and FGP2, PCR was used to amplify the up- and down- stream sequences of each gene using primers 2 & 3 and 6 & 7 (FGP1) and primers 9 & 10 and 12 & 13 (FGP2) (Table S10). PCR products were ligated into plasmid pPK2hphgfp [20] after both product and plasmid were digested using the appropriate restriction enzymes. The upstream flanks of FGP1 and FGP1 were cut with KpnI and PacI and the downstream flanks with HindIII and XbaI. For FGP1, first the upstream flank was ligated into the plasmid and then the downstream flank. For FGP2, first the downstream flank was ligated into the plasmid and then the upstream flank.
A FGP1 complementation construct was generated using PCR and primers 14 & 15 (Table S10), which contain KpnI and EcoRI restriction sites, and ligated in pGEMT-easy (Promega) and sequenced. A correct product was ligated into plasmid pRW1p [55], which was cut using the same enzymes: KpnI and EcoRI. The complementation constructs for chimeric FGP1 and SGE1 genes were prepared using the complementation constructs for both SGE1 and FGP1 in plasmid pRW1p. The combination of the N-terminal SGE1 with the C-terminal FGP1 was amplified using primers 16 & 17 and 18 & 19 (Table S10). The combination of the N-terminal FGP1 part and the C-terminal SGE1 part was amplified using primers 20 & 21 and 22 & 23 (Table S10). PCR products were ligated in pGEMT-easy and sequenced. Correct N-terminal SGE1 and C-terminal FGP1 products were cut from pGEMT-easy using enzymes BglII and XbaI and XbaI and PvuII, respectively. Correct N-terminal FGP1 and C-terminal SGE1 products were cut from pGEMT-easy using enzymes AdhI and XbaI and XbaI and BglII, respectively. Plasmid pRW1pSGE1 [9] was cut using enzymes BglII and PmeI. Using a three-point ligation strategy, the N-terminal SGE1 and C-terminal FGP1 products were ligated into plasmid pRW1pSGE1 cut using enzymes BglII and PmeI and the N-terminal FGP1 and C-terminal SGE1 products were ligated into plasmid pRW1pFGP1 cut using enzymes BglII and AdhI.
For transformations of F. graminearum, a neomycin resistance cassette was ligated into the different pRW1p plasmids. This neomycin cassette was previously cut from pSM334 [56] using the flanking XbaI site and ligated into plasmid pAG1 [57] cut with XbaI. Subsequently, the neomycin cassette was cut from pAG1-Neo with BamHI and ligated into pRW1p cut with the same enzyme.
Agrobacterium mediated transformation used for F. oxysporum was performed as described previously [58]. Agrobacterium mediated transformation used for F. graminearum was performed as described previously [29] with the following alterations: A different A. tumefaciens strain was used (EHA105), filters containing resistant colonies were not transferred to fresh selection plates but instead an agar plug containing a drug resistant colony was placed in liquid CMC medium and after two days of growth, spores were filtered through one layer of sterile miracloth and plated onto PDA plates containing cefoxitin (300 µg/ml) and hygromycin or geneticin (150 µg/ml). Single spore colonies were subsequently transferred to a fresh PDA plate and mycelial plugs were stored at −80°C. Deletion mutants were tested by PCR and Southern analysis. Transformants made to complement the different deletion strains were checked by PCR for presence of the inserted construct (data not shown) and strains containing an insertion of the gene of interest were used.
DNA was extracted using the CTAB protocol [59] and 5–10 µg was used for restriction and loaded for gel electrophoresis. Transfer of DNA to HyBond N+ (GE Health Care) was performed using standard alkaline procedures according to the manufacturer's protocol. Probes for FGP1 and FGP2 were amplified by PCR using primers 24 & 25 and 26 & 27, respectively (Table S10). Probe hybridization and detection was performed using an AlkPhos kit and CDP-Star chemiluminescent solution (GE Health Care) according to the manufacturer's protocol.
For quantification of microconidia, three independent experiments were performed, each with two replicates Microconidia were harvested after five days of growth in 100 ml CMC medium and 1 ml of 2*104 spores/ml were inoculated into 25 ml of fresh CMC medium. Alternatively 50 µl of a 1*106/ml spore suspension were inoculated into 5 ml of CMC in a 24-deep well plate (1*104 spores per well) and incubated for 5 days. the third method used was to count spores produced on mung bean agar (MBA), using 2 µl of a 1*106/ml spore suspension (2*103 spores) to inoculate a mung bean agar plate. After one week of incubation, two ml of water was added to the plate and spread over the mycelium to collect spores. One ml of the spore suspension was pipetted into a tube and spores within a volume of 10 µl were counted using a haemocytometer.
To assess the spore length, 40–75 spores produced in either CMC or on MBA were placed under a Nikon Eclipse 90i microscope and their length was measured. Perithecium formation was analyzed using the modified carrot agar method as described previously [60] in three replicas. To count the amount of ascospores, two ml of water was spread over the perithecia to collect ascospores. One ml of the ascospore suspension was pipetted into a tube and spores within a volume of 10 µl were counted using a haemacytometer.
For in vitro toxin analysis, conidia (1*104 sp/ml) of each strain were inoculated in six wells containing 2 ml of putrescine medium (30 g/l sucrose, 1 g/l KH2PO4, 0.5 g/l MgSO4, 0.5 g/l KCl, 0.8 g/l putrescine, 2 ml/l FeSO4*H2O solution (5 mg/ml) and 200 µl trace elements (50 g/l citrate, 50 g/l ZnSO4*7H2O, 2.5 g/l CuSO4*5H2O, 0.5 g/l H3BO3, 0.5 g/l NaMoO4*2H2O, 0.5 g/l MnSO4*H2O) and grown for 1 week in the dark at 25°C. For microscopy and the time series experiment, flasks containing 25 ml of putrescine medium or minimal medium (30 g/l sucrose, 1 g/l KH2PO4, 0.5 g/l MgSO4, 0.5 g/l KCl, 2 g/l NaNO3, 2 ml/l FeSO4*H2O solution (5 mg/ml) and 200 µl trace elements (50 g/l citrate, 50 g/l ZnSO4*7H2O, 2.5 g/l CuSO4*5H2O, 0.5 g/l H3BO3, 0.5 g/l NaMoO4*2H2O, 0.5 g/l MnSO4*H2O) were inoculated with 2000 spores/ml and grown in the dark at 25°C with shaking 150 rpm. For each time point, the filtrate was collected by passing it through one layer of miracloth and 250 µl of culture filtrate was placed in a glass vial and lyophilized. In planta trichothecene analysis was performed by placing the inoculated spikelet in a glass vial and measuring its weight. Determination of DON, 3ADON and 15ADON concentration per unit mass in the vials was performed as described earlier [54].
RNA was extracted using Trizol (Invitrogen) according to manufacturer's protocol with an alternative precipitation step using ½ volume of isopropanol and a ½ volume of salt solution (0.8 M Sodium Citrate, 1.2 M NaCl). RNA was extracted from mycelium growing in complete medium for 48 hours in the dark at 25°C with shaking 150 rpm, from CMC grown cultures in a 24 well plate in 12 h light cycle at 25°C with shaking 150 rpm, from MBA grown cultures in 12 h light cycle at 25°C and from putrescine and minimal medium grown cultures for each time point as described above. Mycelium from cultures was harvested by filtration over one or two layers of miracloth, washed with water and frozen in liquid nitrogen. Mycelium was then lyophilized and ground in a mortar and pestle prior to Trizol extraction. RNA was also isolated from inoculated wheat spikelets, which were harvested, frozen in liquid nitrogen and ground in a mortar and pestle prior to Trizol extraction.
For northern blotting 15 µg RNA in loading buffer (0.5× 3-(N-morpholino) propanesulfonic acid buffer (MOPS), 1 M deionized glyoxal, 50% DMSO) was loaded onto a 1× MOPS 1% agarose gel. The RNA was subsequently blotted onto Hybond-N+ (GE Health Care) using a capillary blotting protocol provided by the manufacturer with 20× SSC as transfer buffer. The TRI14 probe was amplified using primers 28 & 29 (Table S10) followed by BamHI digestion and gel purification of the 897 bp fragment containing the portion of TRI14 downstream of the preditcted intron. The ACTIN probe was amplified using primers 30 & 31 (Table S10). Probe hybridization and detection was performed using the AlkPhos kit and CDP-star chemiluminescent solution (GE Health Care) according to the manufacturer's protocol.
RNA cleanup was done using the RNeasy Mini Kit (Qiagen) prior to reverse transcriptase or microarray labeling. RNA labeling reactions were performed according to the standard Affymetrix protocols. The putrescine and minimal medium samples were hybridized to the Affymetrix F. graminearum GeneChips [61] and the complete medium and wheat infection samples were hybridized to an updated Affymetrix nine fungal plant pathogen GeneChip (www.plexdb.org). Hybridizations were performed at the BioMedical Genomics Center of the University of Minnesota.
RNA (2 µg) was treated with DNase (Invitrogen) and used for RT-PCR with SuperScript III Reverse Transcriptase (Invitrogen) according to manufacturer's protocol. The cDNA obtained by different methods was used as template for Quantitative PCR (qPCR), which was performed in two replicates with DyNamo™ SYBR® Green qPCR (Finnzymes) using a DNA-Engine Peltier thermal cycler (BioRad) equipped with a Chromo4™ real-time PCR detector and MJ Opticon Monitor™ analysis software. To quantify mRNA levels of genes of interest the ΔCt method was used. Primers used for constitutively expressed FRP1 genes (primers 32–35) and the respective sporulation genes (primers 36–47) are listed in Table S10.
Fungal infection in spikelets was monitored using an Olympus SZX16 Research Stereo Microscope and perithecia and cirrhi formation was observed using an Olympus SZX12 Research Stereo Microscope. Hyphal morphology of fungal strains growing in putrescine and control minimal medium was observed using a Nikon Eclipse 90i microscope.
CEL files were imported in Refiner 5.3 software (Expressionist) and RMA preprocessing was applied. Signal values (p-value 0.04) obtained in the Analyst software (Expressionist) were normalized to the median. Fold-expression filters were applied as described in the results. The probe sets and the corresponding probe descriptions from the F. graminearum GeneChips were converted from the annotation of the FG1 assembly to the FG3 assembly using the Fusarium graminearum database of MIPS (http://mips.helmholtz-muenchen.de/genre/proj/FGDB/) in order to compare the experiments [62]. Probe sets on the nine fungal plant pathogen genome array GeneChips was designed based on the FG3 assembly for F. graminearum and the FO2 assembly for F. oxysporum (Fusarium Comparative Sequencing Project, Broad Institute of Harvard and MIT (http://www.broadinstitute.org/)). For the experiments with both F. graminearum and F. oxysporum grown in CM, a F. oxysporum ortholog was queried by BLAST [17] for every F. graminearum gene showing altered expression. In order to establish whether a gene was conserved in both F. oxysporum and F. graminearum, we searched for the respective orthologs using BLAST and a hit was considered a full ortholog at a bit score of >200. Data and CEL files for microarray experiments are available at www.plexdb.org [63] under accession numbers NF2 (wheat infection), NF3 (growth on complete medium) and FG18 (growth on putrescine medium).
|
10.1371/journal.pcbi.1006195 | Limits on reliable information flows through stochastic populations | Biological systems can share and collectively process information to yield emergent effects, despite inherent noise in communication. While man-made systems often employ intricate structural solutions to overcome noise, the structure of many biological systems is more amorphous. It is not well understood how communication noise may affect the computational repertoire of such groups. To approach this question we consider the basic collective task of rumor spreading, in which information from few knowledgeable sources must reliably flow into the rest of the population. We study the effect of communication noise on the ability of groups that lack stable structures to efficiently solve this task. We present an impossibility result which strongly restricts reliable rumor spreading in such groups. Namely, we prove that, in the presence of even moderate levels of noise that affect all facets of the communication, no scheme can significantly outperform the trivial one in which agents have to wait until directly interacting with the sources—a process which requires linear time in the population size. Our results imply that in order to achieve efficient rumor spread a system must exhibit either some degree of structural stability or, alternatively, some facet of the communication which is immune to noise. We then corroborate this claim by providing new analyses of experimental data regarding recruitment in Cataglyphis niger desert ants. Finally, in light of our theoretical results, we discuss strategies to overcome noise in other biological systems.
| Biological systems must function despite inherent noise in their communication. Systems that enjoy structural stability, such as biological neural networks, could potentially overcome noise using simple redundancy-based procedures. However, when individuals have little control over who they interact with, it is unclear what conditions would prevent runaway error accumulation. This paper takes a general stance to investigate this problem, concentrating on the basic information-dissemination task of rumor spreading. Drawing on a theoretical model, we prove that fast rumor spreading can only be achieved if some part of the communication setting is either stable or reliable. We then provide empirical support for this claim by conducting new analyses of data from experiments on recruitment in desert ants.
| Systems composed of tiny mobile components must function under conditions of unreliability. In particular, any sharing of information is inevitably subject to communication noise. The effects of communication noise in distributed living systems appears to be highly variable. While some systems disseminate information efficiently and reliably despite communication noise [1–5], others generally refrain from acquiring social information, consequently losing all its potential benefits [6–8]. It is not well understood which characteristics of a distributed system are crucial in facilitating noise reduction strategies and, conversely, in which systems such strategies are bound to fail. Progress in this direction may be valuable towards better understanding the constraints that govern the evolution of cooperative biological systems.
Computation under noise has been extensively studied in the computer science community. These studies suggest that different forms of error correction (e.g., redundancy) are highly useful in maintaining reliability despite noise [9–12]. All these, however, require the ability to transfer significant amount of information over stable communication channels. Similar redundancy methods may seem biologically plausible in systems that enjoy stable structures, such as brain tissues.
The impact of noise in stochastic systems with ephemeral connectivity patterns is far less understood. To study these, we focus on rumor spreading—a fundamental information dissemination task that is a prerequisite to almost any distributed system [13–16]. The literature on rumor spreading is quite vast and encompasses different disciplines over the last decades [17, 18]. For a succinct overview as for theoretical computer science, see Section Related works in computer science in the Supplementary Information.
A successful and efficient rumor spreading process is one in which a large group manages to quickly learn information initially held by one or a few informed individuals. Fast information flow to the whole group dictates that messages be relayed between individuals. Similar to the game of Chinese Whispers, this may potentially result in runaway buildup of noise and loss of any initial information [19]. It currently remains unclear what are the precise conditions that enable fast rumor spreading. On the one hand, recent works indicate that in some models of random noisy interactions, a collective coordinated process can in fact achieve fast information spreading [20, 21]. These models, however, are based on push operations that inherently include a certain reliable component (see more details in Section Separation between PUSH and PULL). On the other hand, other works consider computation through noisy operations, and show that several distributed tasks require significant running time [22]. The tasks considered in these works (including the problem of learning the input bits of all processors, or computing the parity of all the inputs) were motivated by computer applications, and may be less relevant for biological contexts. Moreover, they appear to be more demanding than basic tasks, such as rumor spreading, and hence it is unclear how to relate bounds on the former problems to the latter ones.
In this paper we take a general stance to identify limitations under which reliable and fast rumor spreading cannot be achieved. Modeling a well-mixed population, we consider a passive communication scheme in which information flow occurs as one agent observes the cues displayed by another. If these interactions are perfectly reliable, the population could achieve extremely fast rumor spreading [16]. In contrast, here we focus on the situation in which messages are noisy. Informally, our main theoretical result states that fast rumor spreading through large populations can only be achieved if either
the system exhibits some degree of structural stability, or
some facet of the pairwise communication is immune to noise.
In fact, our lower bounds hold even when individuals are granted unlimited computational power and even when the system can take advantage of complete synchronization. In light of these theoretical results, we then turn to discuss several examples of information sharing in distributed biological systems. We provide new analyses of the efficiency of information dissemination during recruitment by desert ants. These suggest that this system lacks reliability in all its communication components, and its deficient performances qualitatively validate our predictions. Finally, we revisit existing rumor spreading solutions in large biological systems and discuss different strategies for confronting noise.
An intuitive description of the model follows. For more precise definitions, see, Section The models in the Supplementary Information.
Consider a population of n agents. Thought of as computing entities, assume that each agent has a discrete internal state, and can execute randomized algorithms—by internally flipping coins. In addition, each agent has an opinion, which we assume for simplicity to be binary, i.e., either 0 or 1. A small number, s, of agents play the role of sources. Source agents are aware of their role and share the same opinion, referred to as the correct opinion. The goal of all agents is to have their opinion coincide with the correct opinion.
To achieve this goal, each agent continuously displays one of several messages taken from some finite alphabet Σ. Agents interact according to a random pattern, termed as the parallel-PULL model: In each round t ∈ N +, each agent u observes the message currently displayed by another agent v, chosen independently and uniformly at random from all agents. Importantly, communication is noisy, hence the message observed by u may differ from that displayed by v. More precisely, for any m, m′ ∈ Σ, let Pm,m′ be the probability that, any time some agent u observes an agent v holding some message m ∈ Σ, u actually receives message m′. The probabilities Pm,m′ define the entries of the noise-matrix P [21], which does not depend on time.
The noise is characterized by a noise parameter δ > 0. Our model encapsulates a large family of noise distributions, making our bounds highly general. Specifically, the noise distribution can take any form, as long as it satisfies the following criterion.
Definition 1 (δ-uniform noise) We say that the noise is δ-uniform if Pm,m′ ≥ δ for any m, m′ ∈ Σ.
When messages are noiseless, it is easy to see that the number of rounds that are required to guarantee that all agents hold the correct opinion with high probability is O ( log n ) [16]. In what follows, we aim to show that when the δ-uniform noise criterion is satisfied, the number of rounds required until even one non-source agent can be moderately certain about the value of the correct opinion is very large. Specifically, thinking of δ and s as constants independent of the population size n, this number of rounds is at least Ω(n).
To prove the lower bound, we will bestow the agents with capabilities that far surpass those that are reasonable for biological entities. These include:
We show that even given this extra computational power, fast convergence cannot be achieved. All the more so, fast convergence is impossible under more realistic assumptions.
The purpose of this work is to identify limitations under which efficient rumor spreading would be impossible. Our main result is theoretical and, informally, states that when all components of communication are noisy fast rumor spreading through large populations is not feasible. In other words, our results imply that fast rumor spreading can only be achieved if the system either exhibits some degree of structural stability or that some facet of its communication is immune to noise. These results in hand, a next concern is how far our highly theoretical analysis can go in explaining actual biological systems.
Theoretical results with a high degree of generality may hold relevance to a wider range of biological systems. Lower bound and impossibility results follow this approach. Indeed, impossibility results from physics and information theory have previously been used to further the understanding of several biological systems [23, 24]. The results we present here are, similarly, in the form of lower bounds but, this time, they are derived from the realm of distributed computation. As such, our theorems are general enough to constrain the performances of a vast class of computational systems regardless of their particulars or the specific computational algorithms which they apply. This generality stretches over to biology and can provide us with fundamental lessons regarding the limitations faced by distributed biological systems [24–26].
While the generality of our lower bound results makes them relevant to a large number of biological systems it also constitutes a weakness. Namely, the assumptions on which such theorems are based are not tailored to describe a particular system. This implies that comparisons between the model assumptions and the actual details of a specific system will not be perfect. Nevertheless, we show how our theoretical results can shed light on some non-trivial behaviors in a specific biological system whose characteristics are close enough to the underlying theoretical assumptions (see Section Recruitment in desert ants). Particularly, we empirically show that when desert ants communicate information regarding a new food source they are subject to limitations which are similar to those assumed by our model. We then demonstrate a non-trivial slowdown in the speed at which information spreads through the system as a function of group size. Despite the non-perfect matching between the theoretical assumption and the biological system, this non-trivial result stands in direct accordance with our theoretical lower bounds.
Distributed computing provides an effective means of studying biological groups [27–30]. However, to the best of our knowledge, there are no examples in which algorithmic lower bounds, one of distributed computing most powerful tools, have been applied to a particular living system. This work uses lower bounds to provide insights into non-trivial dynamics observed during ant recruitment behavior.
In all the statements that follow we consider the parallel-PULL model satisfying the δ-uniform noise criterion, with cs/n < δ ≤ 1/|Σ| for some sufficiently large constant c, where the upper bound follows from the criterion given in Definition 1. Hence, the previous lower bound on δ implies a restriction on the alphabet size, specifically, |Σ| < n/(cs).
Theorem 1.1 Any rumor spreading protocol cannot converge in less than Ω ( n δ s 2 ( 1 - δ | Σ | ) 2 ) rounds.
Observe that the lower bound we present loses relevance when s is of order greater than n, as our proof technique becomes uninformative in presence of a large number of sources (see Remark 2 in the Supplementary Information). Recall also that we assume that a source is aware that it is a source, but if it wishes to identify itself as such to agents that observe it, it must encode this information in a message, which is, in turn, subject to noise. We also consider the case in which an agent can reliably identify a source when it observes one (that is, this information is not noisy). For this case, the following lower bound, which is weaker than the previous one but still polynomial, apply (see also the S1 Text, Detectable sources):
Corollary 1.1 Assume that sources are reliably detectable. There is no rumor spreading protocol that converges in less than Ω ( ( n δ s 2 ( 1 - δ | Σ | ) 2 ) 1 / 3 ) rounds.
Our results suggest that, in contrast to systems that enjoy stable connectivity, structureless systems are highly sensitive to communication noise (see Fig 1). More concretely, the two crucial assumptions that make our lower bounds applicable are: 1) stochastic interactions, and 2) δ-uniform noise (Fig 1, right hand panel). When agents can stabilize their interactions the first assumption is violated. In such cases, agents can overcome noise by employing simple error-correction techniques, e.g., using redundant messaging or waiting for acknowledgment before proceeding to the next action. As demonstrated in Fig 1, (left hand panel), when the noise is not uniform, it might be possible to overcome it with simple techniques based on using default neutral messages, and employing exceptional distinguishable signals only when necessary.
Our theoretical results assert that efficient rumor spreading in large groups could not be achieved without some degree of communication reliability. An example of a biological system whose communication reliability appears to be deficient in all of its components is recruitment in Cataglyphis niger desert ants. In this species, when a forager locates an oversized food item, she returns to the nest to recruit other ants to help in its retrieval [31, 32].
In our experimental setup, summarized in Fig 2, recruitment occurs in the small area of the nest’s entrance chamber (Fig 2a). We find that within this confined area, the interactions between ants follow a near uniformly random meeting pattern [33]. In other words, ants seem to have no control over which of their nest mates they will meet next (Fig 2b). This random meeting pattern approximates the first main assumption of our model. Another of the model’s assumptions is that ants interact in parallel. This implies that the interaction rate per ant be constant and independent of group size. Indeed, the empirical rate of interaction during the recruitment process was measured to be 0.82 ± 0.07 (mean ± sem, N = 44) interactions per minute per ant and induces a small increase with group size: 0.62 ± 0.13 for two ants (N = 8) and 1 ± 0.2 for a group sizes of 9-10 (N = 5).
It has been shown that recruitment in Cataglyphis niger ants relies on rudimentary alerting interactions [34, 35] which are subject to high levels of noise [32]. Moreover, the information an ant passes in an interaction can be attributed solely to her speed before the interaction [32]. Binning ant speeds into three discrete messages and measuring the responses of stationary ants to these messages, we can estimate the probabilities of one message to be mistakenly perceived as another one (see Estimating δ in the Methods). We find that this communication is extremely noisy which complies with the uniform-noise assumption with a δ of approximately 0.3 (Fig 2c). While artificially dividing the continuous speed signals into a large number of discrete messages (thus creating a larger alphabet) would inevitably decrease δ, this is not supported by our empirical data (see Section Methods).
Finally, the interaction scheme, as exhibited by the ants, can be viewed somewhere in-between the noisy-push and the noisy-pull models. Moving ants tend to initiate more interaction [32] and this may resemble, at first glance, a noisy-push interaction scheme. However, the ants’ interactions actually share characteristics with noisy-pull communication. Mainly, ants cannot reliably distinguish an ant that attempts to transmit information from any other non-communicating individual [32]. The fact that a receiver ant cannot be certain that a message was indeed communicated to her coincides with the lack of reliability in information transmission in line with our theoretical assumptions (see more details on this point in the Section Separation between PUSH and PULL).
Given the coincidence between the communication patterns in this ant system and the requirements of our lower bound we expect long delays before any uninformed ant can be relatively certain that a recruitment process is occurring. We therefore measured the time it takes an ant, that has been at the food source, to recruit the help of two nest-mates for different total group size. One might have expected this time to be independent of the group size or even to decrease as two ants constitute a smaller fraction of larger groups. To the contrary, we find that the time until the second ant is recruited increases with group size (p < 0.05 Kolmogorov-Smirnov test over N = 24 experiments, see Fig 2d).
Our theoretical results set a lower bound on the minimal time it takes uninformed ants to be recruited. Note that our lower bounds actually correspond to the time until any individual can be sure with more than 2/3 probability of the rumor. In the context of the ant recruitment experiment this means that if an ant goes out of the nest only if she is sure with some probability that there is a reason to exit, then the lower bounds correspond to the time until the first, and similarly the second (see Fig 2d), ants exit the nest.
Our lower bound is linear in the group size (Theorem 1.1). Note that this does not imply that the ants’ biological algorithm matches the lower bound and must be linear as well. Rather, our theoretical results qualitatively predict that as group size grows, recruitment times must eventually grow as well. This stands in agreement with Fig 2d. Thus, in this system, inherently noisy interactions on the microscopic level have direct implications on group level performance.
Here, we provide the intuition for our main theoretical result, Theorem 1.1. For a formal proof please refer to the S1 Text, The lower bounds. The proof can be broken into three parts and, below, we refer to each of them separately.
Several of the assumptions discussed earlier for the parallel-PULL model were made for the sake of simplicity of presentation. In fact, our results can be shown to hold under more general conditions, that include: 1) different rate for sampling a source, and 2) a more relaxed noise criterion. In addition, our theorems were stated with respect to the parallel-PULL model. In this model, at every round, each agent samples a single agent u.a.r. In fact, for any integer k, our analysis can be applied to the model in which, at every round, each agent observes k agents chosen u.a.r. In this case, the lower bound would simply reduce by a factor of k. Our analysis can also apply to a sequential variant, in which in each time step, two agents u and v are chosen u.a.r from the population and u observes v. In this case, our lower bounds would multiply by a factor of n, yielding, for example, a lower bound of Ω(n2) in the case where δ and s are constants. Observe that the latter increase is not surprising as each round in the parallel-PULL model consists of n observations, while the sequential model consists of only one observation in each time step. See more details in the Supplementary Information.
Our lower bounds on the parallel-PULL model (where agents observe other agents) should be contrasted with known results in the parallel-PUSH model (this is the push equivalent to parallel-PULL model, where in each round each agent may or may not actively push a message to another agent chosen u.a.r.). Although never proved, and although their combination is known to achieve more power than each of them separately [16], researchers often view the parallel-PULL and parallel-PUSH models as very similar on complete communication topologies. Our lower bound result, however, undermines this belief, proving that in the context of noisy communication, there is an exponential separation between the two models. Indeed, when the noise level is constant for instance, convergence (and in fact, a much stronger convergence than we consider here) can be achieved in the parallel-PUSH using only logarithmic number of rounds [20, 21], by a simple strategy composed of two stages. The first stage consists of providing all agents with a guess about the source’s opinion, in such a way that ensures a non-negligible bias toward the correct guess. The second stage then boosts this bias by progressively amplifying it. A crucial aspect in the first stage is that agents remain silent until a certain point in time that they start sending out messages. This prevents agents from starting to spread information before they have sufficiently reliable knowledge and allows for a balanced control of the rumor spread. More specifically, marking an edge corresponding to a message received for the first time by an agent, the set of marked edges forms a spanning tree of low depth, rooted at the source. The depth of such tree can be interpreted as the deterioration of the message’s reliability. On the other hand, as shown here, in the parallel-PULL model, even with the synchronization assumption, rumor spreading cannot be achieved in less than a linear number of rounds.
Perhaps the main reason why these two models are often considered similar is that with an extra bit in the message, a PUSH protocol can be approximated in the PULL model, by letting this bit indicate whether the agent in the PUSH model was aiming to push its message. However, for such a strategy to work, this extra bit has to be reliable. Yet, in the noisy PULL model, no bit is safe from noise, and hence, as we show, such an approximation cannot work. In this sense, the extra power that the noisy PUSH model gains over the noisy PULL model, is that the very fact that one node attempts to communicate with another is reliable. This, seemingly minor, difference carries significant consequences.
Communication in man-made computer networks is often based on reliable signals which are typically transferred over highly defined structures. These allow for ultra-fast and highly reliable calculations. Biological networks are very different from this and often lack reliable messaging, well defined connectivity patterns or both. Our theoretical results seem to suggest that, under such circumstances, efficient spread of information would not be possible. Nevertheless, many biological groups disseminate and share information, and, often, do so reliably. Next, we discuss information sharing in biological systems within the general framework of our lower-bounds.
The correctness of the lower bounds relies on two major assumptions: 1) stochastic interactions, and 2) uniform noise. Communication during desert ant recruitment complies with both these assumptions (see Fig 2b and 2c) and indeed the speed at which messages travel through the group (see Fig 2d) is low. Below, we discuss several biological examples where efficient rumor spreading is achieved. We expect that, in these examples, at least one of the assumptions mentioned above should break adding some degree of reliability to the overall communication. The group can then utilize this reliability and follow one of the strategies mentioned in Section Theoretical results, in order to yield reliable collective performance. We begin by discussing examples that violate the first assumption, namely, that of stochastic interactions, and then discuss examples that violate the second assumption, namely, uniform noise.
All experimental results presented in this manuscript are re-analyses of data obtained in Cataglyphis niger recruitment experiments [32]. In short, ants in the entrance chamber of an artificial nest were given access to a tethered food item just outside the nest’s entrance (Fig 2a). The inability of the ants to retrieve the food induced a recruitment process [32].
The reaction of the ants to this manipulation was filmed and the locations, speeds and interactions of all participating ants were extracted from the resulting videos.
To estimate the noise parameter δ we used interactions between ants moving at three different speed ranges (measured in cm/sec), namely, ‘a’: 1-10, ‘b’: 10-20, and ‘c’: over 20 and “receiver” ants. Only interactions in which the receiver ant was initially stationary were used as to ensure that the state of these ants before the interaction is as similar as possible. The message alphabet is then assumed to be Σ = {a, b, c}. The response of a stationary ant v to the interaction was quantified in terms of her speed after the interaction.
An alphabet of three messages was used since the average responses of v to any two messages were significantly different (all p-values smaller than 0.01) justifying the fact that these are not artificial divisions of a continuous speed signal into a large number of overlapping messages. On the other hand, dividing the bins further (say, each bin divided into 2 equal bins) yielded statistically indistinguishable responses from the receiver (all p-values larger than 0.11). Therefore, our current data best supports a three letter alphabet.
Assuming equal priors to all messages in Σ, and given specific speed of the receiver ant, v, the probability that it was the result of a specific message i ∈ Σ was calculated as pi(v) = p(v ∣ i)/∑k∈Σ p(v ∣ k), where p(v ∣ j) is the probability of responding in speed v after “observing” j. The probability δ(i, j) that message i was perceived as message j was then estimated as the weighted sum over the entire probability distribution measured as a response to j: δ(i, j) = ∑v p(v ∣ j) ⋅ pi(v). The parameter δ can then be calculated using δ = min{δ(i, j) ∣ i, j ∈ Σ}.
|
10.1371/journal.pbio.1001682 | Biotic and Human Vulnerability to Projected Changes in Ocean Biogeochemistry over the 21st Century | Ongoing greenhouse gas emissions can modify climate processes and induce shifts in ocean temperature, pH, oxygen concentration, and productivity, which in turn could alter biological and social systems. Here, we provide a synoptic global assessment of the simultaneous changes in future ocean biogeochemical variables over marine biota and their broader implications for people. We analyzed modern Earth System Models forced by greenhouse gas concentration pathways until 2100 and showed that the entire world's ocean surface will be simultaneously impacted by varying intensities of ocean warming, acidification, oxygen depletion, or shortfalls in productivity. In contrast, only a small fraction of the world's ocean surface, mostly in polar regions, will experience increased oxygenation and productivity, while almost nowhere will there be ocean cooling or pH elevation. We compiled the global distribution of 32 marine habitats and biodiversity hotspots and found that they would all experience simultaneous exposure to changes in multiple biogeochemical variables. This superposition highlights the high risk for synergistic ecosystem responses, the suite of physiological adaptations needed to cope with future climate change, and the potential for reorganization of global biodiversity patterns. If co-occurring biogeochemical changes influence the delivery of ocean goods and services, then they could also have a considerable effect on human welfare. Approximately 470 to 870 million of the poorest people in the world rely heavily on the ocean for food, jobs, and revenues and live in countries that will be most affected by simultaneous changes in ocean biogeochemistry. These results highlight the high risk of degradation of marine ecosystems and associated human hardship expected in a future following current trends in anthropogenic greenhouse gas emissions.
| Climate change caused by human activity could damage biological and social systems. Here we gathered climate, biological, and socioeconomic data to describe some of the events by which ocean biogeochemical changes triggered by ongoing greenhouse gas emissions could cascade through marine habitats and organisms, eventually influencing humans. Our results suggest that the entire world's ocean surface will be simultaneously impacted by varying intensities of ocean warming, acidification, oxygen depletion, or shortfalls in productivity. Only a very small fraction of the oceans, mostly in polar regions, will face the opposing effects of increases in oxygen or productivity, and almost nowhere will there be cooling or pH increase. The biological responses to such biogeochemical changes could be considerable since marine habitats and hotspots for several marine taxa will be simultaneously exposed to biogeochemical changes known to be deleterious. The social ramifications are also likely to be massive and challenging as some 470 to 870 million people – who can least afford dramatic changes to their livelihoods – live in areas where ocean goods and services could be compromised by substantial changes in ocean biogeochemistry. These results underline the need for urgent mitigation of greenhouse gas emissions if degradation of marine ecosystems and associated human hardship are to be prevented.
| As CO2 and other greenhouse gas emissions continue to rise, ocean biogeochemistry is being altered in ways that could potentially impact nature and mankind. Atmospheric CO2 concentrations have already risen to ∼400 ppm from ∼280 ppm in pre-industrial times and could rise to between 550 and 900 ppm by 2100, depending upon the emission scenario [1]–[7]. In the marine realm, the surplus of CO2 has been associated with ocean warming from the greenhouse effect [1] and acidification caused by the fact that approximately 25% of the annually emitted CO2 enters the ocean, where it reacts with water to produce carbonic acid, thereby reducing pH [6]–[8]. Ocean warming and other climatic changes can trigger additional responses in connection to ocean circulation and stratification, which in turn reduce oxygen concentration [9],[10] and primary productivity [11] (additional responses may include sea-level rise and extreme weather events, which we do not analyze here but that certainly will add to the stress likely to be exerted by greenhouse gas emissions [10]). Several analyses predict that, by the year 2100, depending on the emission scenario, surface ocean temperature could increase by 2 to 3°C [9], pH decline by over 0.2 units [6],[7], oxygen concentration decrease by 2% to 4% [9], and ocean productivity by 2% to 20% [11], from current values. The magnitude of these changes would be unprecedented in the Earth's history during the last 20 million years [12],[13].
Species are adapted to their environment, and therefore shifts in environmental parameters can induce considerable change in species fitness and trigger additional responses in community composition, functioning, and overall biodiversity [2],[3],[9]–[11],[14]–[16]. Ocean warming, acidification, oxygen depletion, and reduction in primary production have all been highlighted as potentially having negative biological consequences [2],[3],[9]–[11],[14]–[16]. Changes in temperature, for instance, can affect metabolism, reproduction, and survival [10],[17], which is already evident in multiple shallow and deep-sea ecosystems [2],[18]. Parameters related to food supply, such as primary productivity and sinking organic-carbon flux, and dissolved oxygen can influence metabolism, body size, reproduction, and thus control, in part, the biomass that can be sustained in any given area of the ocean [19]. Moreover, depending on the magnitude of shifts in biogeochemical parameters and/or their proximity to physiological thresholds, these changes can make entire areas essentially unsuitable for metazoans (except for some meiofaunal organisms, as well as viruses, prokaryotes, and certain protists [20]–[23]). There is already evidence that oxygen minimum zones have increased in vertical extent over recent decades, with important consequences for ecosystems and coastal communities [24]. Likewise, pH can influence rates of calcification and several other physiological processes [10],[15],[25],[26]. Co-occurring changes in biogeochemical parameters could also accelerate biological responses, either additively or synergistically [10],[27]–[30]. Warming, for instance, can increase metabolism but, if combined with a reduction in dissolved oxygen and food availability, it could also lead to considerable reductions in body size [31], survival, and synergistic responses of ecosystems [32] and cause range expansions or contractions [2],[10],[17]. Studies on marine invertebrates have also revealed that embryos that survive exposure to warming may later die as larvae if exposed to acidification [33]. This is not to say that all species will be impacted negatively. Some species may expand to new areas or thrive in areas where they were once rare. It is certain, however, that biogeochemical changes in the ocean, especially their co-occurrence, have considerable potential to reorganize patterns in biodiversity, body size, and abundance (Table 1). Additionally, the number of species within ecosystems, variations in life histories, and susceptibility to climate change among species suggest that ecosystem responses to ocean biogeochemistry change are likely to be varied and highly idiosyncratic (Table 1).
Socioeconomic systems can also be sensitive to ocean biogeochemical changes, depending upon the exposure of ocean goods and services to environmental change, human dependence on affected services, and social adaptive capacity [34]–[38]. Examples of the goods and services likely to be impacted by ocean climate change are diverse. Ocean warming and acidification, for instance, are causing a new set of conditions that are very close to the tolerance thresholds of corals, making them vulnerable to massive bleaching and mortality when long-term trends related to climate change are “added” to natural variability. The decay of coral reefs could potentially impair their ability to deliver goods and services such as fisheries, tourism, coastal protection, and in some cases aesthetic and spiritual values [35],[37], which have been grossly valued at over US$375 billion annually [39]. Likewise, future changes in ocean temperature are expected to cause a redistribution in the global diversity of cetaceans [40], which in turn could impact local economies that rely on tourism or the fishing of these species. A similar example is the effect of ocean climate change on the world's fisheries, where a combination of warming, oxygen depletion, and reduction in primary productivity can induce changes in body size [31], abundance, and distribution of exploited species [41],[42]. These would add to the ongoing decline of fisheries yields, which are considerable sources of food, revenues, and jobs [36],[42],[43]. Shifts in the distribution and abundance of species could also bring new opportunities for local communities, although adaptability (e.g., flexibility and responsiveness) will be needed to realize any potential benefits [38]. However, the vulnerability of societies to the changes in ocean goods and services ultimately depends on the balance among exposure to environmental change, human dependency on impacted goods and services, and social adaptability [34],[35],[44],[45]. In that context, we are aware of two relevant studies analyzing social vulnerability to ocean climate change over large spatial scales: one for fisheries [34] and the other for coral reefs [35]. Given the limited availability of ocean climate projections at the time, the former study used projected mean surface air temperature to 2050 as the underlying indicator of exposure to climate change, while the latter study focused on five countries of the Western Indian Ocean and used thermal stress on coral reefs as a proxy of climate change. As far as we are aware, more detailed studies connecting the exposure to several and co-occurring stressors of climate change with a variety of ocean goods and services at the global scale are lacking.
As indicated above, we have a relatively good understanding of the potential changes in ocean biogeochemical parameters expected under different greenhouse gas scenarios [7],[9],[11],[46], and conceptually we know some of the mechanisms through which ecological and social systems may be impacted by such changes. However, we lack a synthetic global quantification of the simultaneous projection of biogeochemical changes on the ocean and how they may pertain to marine biota and people worldwide. To address this gap, we compiled all available data generated by Earth Systems Models as part of the Coupled Model Intercomparison Project Phase 5 (CMIP5) to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change [47] to assess the extent of co-occurrence of changes in temperature, pH, oxygen, and primary productivity. We complemented the analysis by assembling global distribution maps of 32 marine habitats and biodiversity hotspots to assess the potential vulnerability of biological systems to co-occurring biogeochemical changes in the ocean. Finally, we used available data on human dependency on ocean goods and services and social adaptability to quantify the vulnerability of coastal people to projected ocean biogeochemical change. We would like to emphasize that our results primarily concern the vulnerability of biological and social systems resulting from their exposure to projected anthropogenic ocean biogeochemical change, while cautioning that, although biotic and social responses will certainly occur, the type and magnitude of such responses will be difficult to predict.
Important contributions have been made to understand future projections in the ocean biogeochemical parameters analyzed here [9],. We repeated the collection of such projections for the purposes of identifying patterns of co-occurrence in biogeochemical variables and to quantify sources of error due to model accuracy and precision. Earth Systems Models in the CMIP5 improve upon earlier models and runs by incorporating better knowledge of the climate, improved computational capability, and CO2 pathways that use more detailed and up-to-date data and integrate multiple forcing agents of climate change [48]. Additionally, as noted below, multimodel averages were always more accurate than individual models, further justifying the assembly of biogeochemical projections based on all available Earth System Models. Although these data on ocean biogeochemical parameters represent an important component of our study, our main goals are to identify how their patterns of co-occurrence may pertain to marine biota and thereby social systems worldwide that rely on marine biodiversity goods and services.
The reliability of climate change projections is primarily determined by the skilfulness with which climate models are able to predict the climate [49],[50]. Climate model realism has improved over recent years owing to increased computing power, better scientific understanding of Earth System processes, and the ability to integrate atmosphere, ocean, land, and sea-ice components of the climate system [49]–[51]. However, our theoretical understanding of the climate system is still incomplete and a myriad of unresolved differences exist among models (e.g., spatial and temporal resolution, numerical solution techniques, process parameterizations, and complexity of atmospheric convection, carbon cycle coupling, ocean mixing, unresolved attributes of the biosphere, etc. [49]). As a result, one of the major motivations in climate research has been to quantify the agreement among models as well as between models and actual climate observations. To address these standing concerns, we measured two proxies for model precision and accuracy. Accuracy was defined as the proximity of the model projections to actual data and precision as the standard deviation among the projections of all models. Of course, the availability of actual observations is restricted to recent times and so we assume that a model that accurately simulates present climate will produce better projections of future climates [51].
We found that the average of all models was always closer to actual observations than any model was individually (Tables S1, S2). Thus, errors in precision were often larger than those in accuracy (Figure 1). That is, there were often large differences among the suite of model predictions, but their multimodel average was often closer to actual observations (Tables S1, S2, Figure 1). We also found that the accuracy of the multimodel average varied by parameter and ocean domain. Specifically, there was a stronger predictability of temperature, oxygen, and pH at the ocean surface and a lower predictability of phytoplankton carbon concentration and of all parameters at the seafloor (Figure 1; complete results and details of accuracy and precision plus Taylor diagrams are presented in Tables S1, S2). This low predictability may emerge from the limited availability of actual observations [this may be the case for “phytoplankton carbon concentration” and “particulate organic carbon flux,” which are modeled from other parameters, and there is often a significant disagreement among available products of such parameters (Page 5 in Table S2)] [11] and the inevitable complexity of deep-water processes, which may remain poorly modeled by Earth System Models. With these considerations in mind, we used results based on the upper layer of the ocean and the multimodel average, unless otherwise indicated.
It is worth noticing that discrepancies between Earth System Models outputs and present-day climate observation are partly due to the fact that these models simulate their own internal climate variability (i.e., complex, nonlinear interactions among different components such as atmosphere, ocean, ice, physics, biogeochemistry, etc.) rather than those observed in reality. Thus, a perfect match between any individual model output and observations is unlikely for all places and times. However, these offsetting errors between a given global model and current-day observations have been found to be ameliorated by averaging the output of multiple models (e.g., this study, [52]). This property of multimodel averaging is likely to be just as useful in future climate projections, which highlight the key reason for using the broad range of available models in future predictions of the climate, including those models with moderate capacity to predict current observations [50],[52].
In this study, we analyzed ocean biogeochemical projections under two alternative pathways in which CO2 concentrations could increase to 550 and 900 ppm by 2100 (as reference, atmospheric CO2 concentrations are now at ∼400 ppm from 280 ppm in pre-industrial times; see Figure S1 [1],[48],[53]). These two scenarios are based on Representative Concentration Pathways 4.5 (RCP45) and 8.5 (RCP85) and represent alternative mitigation efforts between a concerted rapid CO2 mitigation and a “business-as-usual” scenario, respectively [48]; there is a more aggressive mitigation scenario called RCP26, which we do not use because it was not consistently used among models and some consider it realistically unattainable (see Figure S1).
Projections of biogeochemical parameters under RCP45 and RCP85 were variable in magnitude among analyzed Earth System Models (semitransparent lines in Figure 2E–H) but followed remarkably similar trends overall (solid lines in Figure 2E–H, Table S1). By 2100, global averages for the upper layer of the ocean could experience a temperature increase of 1.2 to 2.6°C (Figure 2E, Table S3), a dissolved oxygen concentration reduction of 0.11 to 0.24 ml l−1 (i.e., a ∼2% to 4% reduction of current values, Figure 2F, Table S3), a pH decline of 0.15 to 0.31 (Figure 2G, Table S3), and a diminished phytoplankton concentration of 0.001 to 0.003 mg C l−1 (i.e., a ∼4% to 10% reduction of current values, Figure 1H, Table S3) according to RCP45 and RCP85, respectively. In contrast, the world's seafloor was projected to experience smaller changes in temperature and pH (i.e., warming of 0.20 to 0.31°C and acidification of 0.03 to 0.04 pH units) but larger reductions in particulate carbon flux (i.e., food supply) reaching the seafloor (i.e., particulate carbon flux will decline 0.18 to 0.36 mg C m−2 y−1 or 6% to 13% reduction of current values, Table S3); reductions in dissolved oxygen will be similar to those observed at the sea surface (i.e., oxygen will decline by 0.11 to 0.14 ml l−1 compared to current values, Table S3); all values are according to RCP45 and RCP85, respectively.
By 2100, projected changes in temperature, dissolved oxygen, pH, and primary food supply vary significantly among regions (Figure 2A–D). For the ocean surface, the smallest projected changes for pH are in the tropics, for temperature and productivity in temperate regions, and for oxygen in the Southern Ocean (Antarctica). At the seafloor, all variables analyzed experienced the largest changes along continental margins, with decreasing oxygen being common over larger areas of the world's seafloor, particularly at the poles (Figure S2). In general, however, with the exception of the Antarctic and small areas in the South Pacific and North Atlantic, most of the world's oceans will be simultaneously exposed to change in all parameters (Figures 3–4, Figure S2). With the exception of productivity and all parameters at the seafloor, current errors in accuracy and precision of the Earth System Models are of insufficient magnitude to offset projected changes; that is, projected changes in temperature, oxygen, and pH in the upper ocean layer were larger than their errors in accuracy and precision, meaning that trends in these three parameters are robust and are unlikely to be reversed by current sources of model errors (Figure 1, Table S2).
To identify patterns of co-occurrence in biogeochemical changes, we differentiate changes in biogeochemistry that are negative (i.e., warming, acidification, oxygen depletion, and primary food reduction) from those that are positive (i.e., cooling, basification, oxygenation, and productivity increase). Note that the terms “negative” and “positive” are used to indicate the direction of biogeochemical changes, not their potential effects upon biodiversity or social systems. The resulting values were then scaled from 0 to 1 (i.e., 0 meaning no change and 1 the upper 97.5% most extreme absolute change predicted in the world). The scaled-scores for each biogeochemical parameter were added to generate a composite global map of “negative” and “positive” changes in ocean biogeochemistry (Figure 3). The composite global scores were differentiated between “positive” and “negative” changes to avoid neutralization of biogeochemical changes (e.g., a cell with a warming score of −1 and a productivity increase score of 1 will yield a composite global score of 0, which would be confounded with no change). Additionally, separation of the global composite scores into positive and negative changes allows a better appreciation of the preponderance of the directions of biogeochemical change in the world's oceans.
The results of this analysis indicate that the entire ocean surface will be impacted by warming, acidification, or reductions in oxygen and productivity (Figure 4A,C)—over 99% by the largest negative change in at least one full parameter (Figure 4A). In contrast, only oxygen and productivity will experience positive changes at the surface over a small fraction (Figure 4D) of the polar regions (Figure 3C,E); almost no place in the world's ocean surface will face cooling or pH increase (Figure 3B,D). Co-occurring negative changes will also occur extensively over the world's ocean seafloor (Figure 4C,D), although the magnitude of such changes will be smaller: only about 20%–27% of the ocean's seafloor will be exposed to the largest negative change projected in more than one biogeochemical parameter (Figure 4A). Patterns of co-occurrence in biogeochemical parameters were very similar between the RCP45 and RCP85 (Figure 4).
By overlaying the global distribution of marine habitats and hotspots of biodiversity for individual taxa with the projected changes in temperature, oxygen, pH, and primary food supply, we found that, to varying degrees, all projected biogeochemical changes will occur simultaneously within all habitats and biodiversity hotspots (Figure 5; Table S4 provides detailed statistics for the change in each parameter at each marine habitat and biodiversity hotspot and sources of error owing to accuracy and precision in the Earth System Models). Among marine habitats, the smallest absolute changes in biogeochemical parameters are expected to occur in deep-sea habitats (e.g., soft- and hard-bottom benthos, seamounts, and vents; Figure 5A, Table S4), whereas the largest changes will likely occur in shallow-water habitats like coral and rocky reefs, seagrass beds, and shallow soft-bottom benthos (Figure 5A, Table S4). Like the biota, biodiversity hotspots (i.e., areas with high numbers of species of a particular taxon [54]) will also be differentially stressed by ocean biogeochemistry change (Figure 5B). Among biodiversity hotspots analyzed in this study, the smallest cumulative exposure to future biogeochemical change is projected to occur in hotspots of mangrove and coral reef species, whereas the largest exposure will occur in hotspots of euphausiid (i.e., krill; a crucial component of food webs at mid and high latitudes), cetacean, squid, and pinniped species (Figure 5B, Table S4).
For the purpose of assessing the co-occurrence of biogeochemical change, we considered all absolute changes in an additive manner; however, this is not to say that biological responses will follow an additive or linear response to such changes. Realistically, empirical data are unavailable for a sufficient number of species to predict the biological responses of an entire ecosystem to the exposure of biogeochemical change in the ocean (i.e., given variations in physiological adaptations, tolerance thresholds, nonlinear responses, ecological interactions, and resulting cascade effects, etc.). Even broad generalizations could be prone to limitations. For instance, it is often argued that diverse ecosystems can be resilient to climate change as redundancy in species functions could allow the buffering of species lost by climate change. This, in itself, implies a change in community structure [55], although empirical evaluation of this idea has suggested that, perhaps due to strong niche specialization, diverse ecosystems may actually exhibit reduced functional redundancy and be particularly prone to disturbances [56].
Despite our inability to predict the type and magnitude of biological responses to ocean biogeochemistry change, existing knowledge suggests that ocean biogeochemical changes could exert a major selective pressure upon species and have the capability to reorganize patterns of body size, abundance, distribution, species richness, and ecosystem functioning (Table 1). Biological and ecological responses are likely to be magnified, especially if in interaction with other stressors [10],[28], as there will be a need for multiple physiological adaptations. The expected biological response is further highlighted by the biological changes already observed in certain monitored ecosystems in response to recent environmental change. Coral reefs, in which massive bleaching and growth reduction have been linked to relatively minor contemporary warming and acidification [3],[57], provide an excellent example of this. Even deep-sea ecosystems, for which the magnitude of biogeochemical shifts will be smaller (dotted lines in Figure 2E–G), may undergo substantial biological responses, mainly because the deep ocean is much more stable, and thus its faunas are likely adapted to narrower ranges of environmental variation than those in shallow marine habitats [14],[58]. We reemphasize that a standing challenge is to determine the preponderance of taxa from different marine habitats and ecosystems that will be sensitive to ocean biogeochemistry change.
Here we quantified the relative vulnerability of coastal people to ocean biogeochemistry change in the traditional sense of exposure to environmental change, dependency of potentially impacted ocean goods and services, and social adaptability [34],[35],[44],[45].
We determined the level of exposure of each Exclusive Economic Zone in the world to the cumulative negative ocean biogeochemistry change, on a scale ranging between 0 (i.e., no ocean biogeochemistry change) and 4 (i.e., maximum observed biogeochemistry change in all four analyzed parameters) (data from Figure 3B; we analyzed only negative changes given their overwhelming coverage globally, and because those changes are likely to have the largest impacts on the supply of ocean goods and services). For the purpose of classification, cumulative negative biogeochemical changes were divided into three equal bins to classify countries with low, medium, and high exposure to ocean biogeochemistry change.
To quantify levels of dependency, we used three different metrics of peoples' dependence on the ocean: jobs, revenues, and food. Job dependency was measured as the fraction of the countries' work force employed by marine fishing, the marine tourism industry, mariculture, and marine mammal watching. Revenue dependency was measured as the fraction of a country's Gross Domestic Product (GDP) generated by revenues from marine tourism, fishing, mariculture, and marine mammal watching. Food dependency was the fraction of animal protein consumption supplied by seafood. All three dependencies were added and divided in three equal bins to indicate countries of low, medium, and high dependency.
Societal adaptability to environmental change was quantified as per capita GDP, assuming that richer countries will have more alternatives, higher capacity, and adaptability. For the purpose of classification, we defined low-, medium-, and high-income countries depending on whether annual per capita GDP was smaller than US$4,000, between US$4,000 and US$12,000 and larger than US$12,000, respectively (sources of data are presented in Table S6).
For each country, we estimated the number of coastal people (i.e., living within 50 km of the coast) within each category of exposure, dependency, and adaptability (we provide global summaries in the main text and detailed country results in the Supporting Information section). We found that approximately 1.4 billion people live in the coastal areas of countries whose Exclusive Economic Zones will experience medium to high ocean biogeochemistry change by 2100 under the RCP45. Of those, ∼690 million live in countries with a medium to high ocean dependence, and of these ∼470 million live in low-income countries (Figure 6A). The situation will be more dramatic under the RCP85, according to which 2.02 billion coastal people will live in countries with medium to high ocean biogeochemistry change; of those, 1.12 billion live in countries of medium to high ocean dependence; and of these, ∼870 million live in low-income countries (Figure 6B; detailed statistics of the change in each biochemical parameter at each Exclusive Economic Zone and sources of error owing to accuracy and precision in the Earth System Models are shown in Table S5). These results highlight the considerable challenges for human adaptability likely to emerge from ocean biogeochemistry change. Not only does a considerable fraction of the world's human population constantly use resources that will be impacted by ocean climate change, but such people are also located in developing countries with low capacity for adaptation to climate change. This limited socioeconomic capacity could also hamper the ability to benefit from “positive” ecosystem changes, if such new opportunities require costly adaptability [38].
Although a mechanistic model of how ocean biogeochemical changes alter biological and social systems will be difficult to develop, existing knowledge suggests that the responses to the exposure of expected ocean biogeochemical change could be considerable. First, the array of interrelated parameters affected by increasing CO2 emissions provide a much more worrisome picture than consideration of single stressors alone, as most of the world's oceans will be influenced by changes in multiple biogeochemical parameters, and thus adaptation will require multiple physiological adjustments from marine species. Additionally, there is the potential for synergistic responses to co-occurring stressors, and indirect ecological releases and trophic cascades. Secondly, human dependence on marine goods and services is also substantial in countries that will experience considerable ocean biogeochemistry change, particularly among low-income countries. This highlights the looming vulnerability to climate change in developing/low-income countries, and an unfortunate disparity between those who benefit economically from the processes creating climate change and those who will have to pay most of the environmental and social costs. The kind of biogeochemical stressors identified here will be further compounded by sea level rise, which has already been identified as a major potential socioeconomic consequence from climate change. These results provide a refined and synoptic numerical projection of change in key biogeochemical parameters upon marine biota and human societies, and indicate that if global CO2 emissions are not reduced, substantial degradation of marine ecosystems and associated human hardships are very likely to occur.
Our analysis builds on recent ocean physical and biochemical projections developed as part of the Coupled Model Intercomparison Project Phase 5 to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change [47]. As of July 2012, there were 31 Earth System Models from 18 centers in nine countries that modeled at least one of the ocean parameters analyzed here (Table S1). For analysis, all parameters were interpolated into a common 1° by 1° grid (assessment of multiple interpolation methods is provided in the Supplement S1). In total, over 27,000 years of data from the different models and variables were processed. Given the number and size of the files, we used several tools to optimize data processing, which are made available in Supplement S2. To quantify the robustness of Earth System Models, we compared projections among models (to measure model precision) and with actual data (to measure model accuracy) (data sources are indicated in Table S6). The multimodel average projections in the different biogeochemical parameters, in response to the analyzed CO2 scenarios, were overlapped with the distribution of different marine habitats and biodiversity hotspots to calculate how much individual and combined change will occur upon each habitat and hotspot (additional details are provide in Figure 5 and Table S4). Finally, for each Exclusive Economic Zone in the world, we calculated the projected cumulative change in all biogeochemical parameters analyzed here (Figure 3B), and quantified human vulnerability to this change by using country-level data on current social resilience (in terms of wealth and assuming that richer countries will have more alternatives, higher capacity, and adaptability) and human dependence for ocean goods and services arising from food, jobs, and revenue (Results for individual countries are shown in Table S5 and data sources in Table S6.)
|
10.1371/journal.pntd.0000855 | Paromomycin for the Treatment of Visceral Leishmaniasis in Sudan: A Randomized, Open-Label, Dose-Finding Study | A recent study has shown that treatment of visceral leishmaniasis (VL) with the standard dose of 15 mg/kg/day of paromomycin sulphate (PM) for 21 days was not efficacious in patients in Sudan. We therefore decided to test the efficacy of paramomycin for a longer treatment duration (15 mg/kg/day for 28 days) and at the higher dose of 20 mg/kg/day for 21 days.
This randomized, open-label, dose-finding, phase II study assessed the two above high-dose PM treatment regimens. Patients with clinical features and positive bone-marrow aspirates for VL were enrolled. All patients received their assigned courses of PM intramuscularly and adverse events were monitored. Parasite clearance in bone-marrow aspirates was tested by microscopy at end of treatment (EOT, primary efficacy endpoint), 3 months (in patients who were not clinically well) and 6 months after EOT (secondary efficacy endpoint). Pharmacokinetic data were obtained from a subset of patients weighing over 30 kg.
42 patients (21 per group) aged between 4 and 60 years were enrolled. At EOT, 85% of patients (95% confidence interval [CI]: 63.7% to 97.0%) in the 20 mg/kg/day group and 90% of patients (95% CI: 69.6% to 98.8%) in the 15 mg/kg/day group had parasite clearance. Six months after treatment, efficacy was 80.0% (95% CI: 56.3% to 94.3%) and 81.0% (95% CI: 58.1% to 94.6%) in the 20 mg/kg/day and 15 mg/kg/day groups, respectively. There were no serious adverse events. Pharmacokinetic profiles suggested a difference between the two doses, although numbers of patients recruited were too few to make it significant (n = 3 and n = 6 in the 20 mg/kg/day and 15 mg/kg/day groups, respectively).
Data suggest that both high dose regimens were more efficacious than the standard 15 mg/kg/day PM for 21 days and could be further evaluated in phase III studies in East Africa.
ClinicalTrials.gov NCT00255567
| Visceral leishmaniasis (VL) is a parasitic disease transmitted through the bite of sandflies. The WHO estimates 500,000 new cases of VL each year, with more than 90% of cases occurring in Southeast Asia, East Africa, and South America. If left untreated, VL can be fatal. We had previously conducted a large multi-center study in Sudan, East Africa, to assess the efficacy of paromomycin (PM) alone or in combination with sodium stibogluconate. Clinical studies in India have shown that 15 mg/kg/day PM for 21 days was an effective cure. However, the same treatment regimen was not efficacious in two study sites in Sudan. Here, our aim was to assess two high-dose regimens of PM in Sudan: 15 mg/kg/day for 28 days and 20 mg/kg/day for 21 days. The results suggest that, at these total doses, PM is more efficacious than when given daily at 15 mg/kg for 21 days, and that high doses are required to treat VL in Sudan. Efficacy of 20 mg/kg/day PM for 21 days is currently being evaluated in a prospective, comparative phase III trial in East Africa.
| According to the WHO estimates, visceral leishmaniasis (VL) is a parasitic disease that affects more than 500,000 people globally each year [1], and has a fatality rate of up to 100% if left untreated [2]. 90% of cases occur in five countries: India, Bangladesh, Nepal, Sudan, and Brazil [1], with the affected communities mostly located in remote regions of these endemic areas without ready access to treatment.
Although drugs (mainly antimonials such as sodium stibogluconate [SSG]) currently exist to treat this parasitic infection, their use has been limited because of high cost, toxicity, or development of parasite resistance [3]–[5]. A multi-center phase III study in India showed that PM is a very efficacious, affordable, and safe treatment [6], and is now registered for VL treatment in India.
In an effort to identify an effective treatment for VL in East Africa, we had previously initiated a multi-center phase III study in Sudan, Ethiopia, and Kenya comparing the efficacy of PM alone at the dose shown to be efficacious in India (15 mg/kg/day for 21 days) against SSG alone (20 mg/kg/day for 30 days) and against a combination treatment of SSG and PM (same dose of individual treatments but for 17 days). PM monotherapy did not show adequate efficacy, particularly in Sudan where parasite clearance was below 50% in patients at 6 months after end of treatment (EOT), and the study had to be prematurely stopped [7].
In the current study, we sought to find an efficacious dose of PM for the treatment of VL in Sudan and to explore possible reasons for the failure of the drug at the previous dose studied of 15 mg/kg/day for 21 days. In our previous study using this dose, conducted in 5 sites in Ethiopia, Kenya and Sudan, we found an overall end of treatment cure of 67.4% and 6-month post-treatment cure of 63.8% [7]. Cure at both sites in Sudan was below 50% [7]. The cure rate in this study of SSG was 92.2% at 6 months post-treatment [7].
Therefore a total dose increase of 33% was attempted through two possible regimens- an increased dose of 20 mg/kg for 21 days or a prolonged course of 15 mg/kg for 28 days. The former regimen has been evaluated in some clinical trials in India [8], [9]. There was no previous clinical experience with the 15 mg/kg dosage given for 28 days. The rationale was that the longer course of treatment would provide additional time for the patient's general condition to improve, and for their immunological response to develop, and that this might translate into a better clinical response without increasing the daily dosage.
The main objective was to assess the efficacy of two dosing regimens of PM monotherapy for the treatment of VL: 20 mg/kg/day for 21 days and 15 mg/kg/day for 28 days. Secondary objectives were to assess the safety of PM and compare the pharmacokinetic (PK) profiles of the two groups in a subset of patients.
Patients with clinical symptoms and signs suggestive of VL and confirmed by visualization of parasites in bone-marrow aspirates were eligible for enrollment according to the National VL guidelines for Sudan for treatment and control. To be included in the study, patients had to: be between 4 and 60 years of age; be able to comply with the protocol (Protocols S1, S2 and S3); and provide written informed consent signed by themselves or by parents or legal guardians.
Patients were excluded from the study if they: had negative bone-marrow smears; were clinically contraindicated to having a bone-marrow aspirate; received any anti-leishmania drug in the past 6 months; had severe protein or caloric malnutrition (Kwashiorkor or marasmus); had previous hypersensitivity reaction to aminoglycosides; suffered from a concomitant severe infection, ie tuberculosis, HIV, or any other serious underlying disease (cardiac, renal, hepatic); suffered from other conditions associated with splenomegaly such as schistosomiasis; had previous history of cardiac arrhythmia or an abnormal electrocardiogram (ECG); were pregnant or lactating; or had pre-existing clinical hearing loss. If tuberculosis or schistosomiasis were suspected, these were screened through laboratory testing. Additionally, patients with the following laboratory values were excluded: hemoglobin less than 5 g/dL; white blood cell less than 103/mm3; platelets less than 40,000/mm3; liver function test values more than three times the normal range; and serum creatinine values outside the normal range for age and gender.
This was a two-arm, randomized, open-label, dose-finding study done at a single site in Sudan (Kassab Hospital, Ministry of Health, Gedaref State). This site participated in the previous study conducted on PM [7]. Eligible patients were randomly assigned to 20 mg/kg/day PM for 21 days (n = 21) or 15 mg/kg/day PM for 28 days (n = 21), and started a treatment regimen upon allocation to their treatment. Restricted-block randomization was done for the two groups. Randomization was done using sequentially numbered sealed envelopes that were prepared according to a centrally generated randomization list. Treatment was administered via daily intramuscular injection, and patients remained in the hospital for the duration of treatment. Patients were followed up at 3 and 6 months after treatment as outpatients. Parasitological assessments (bone-marrow aspirates only) were done at baseline, end of treatment (EOT), 3 months (only on patients who were not clinically well) and 6 months after treatment. Safety and clinical laboratory assessments were done at baseline, day 7 and day 14 of drug administration, EOT, and at 3 and 6 months follow-ups. These included a clinical assessment, (clinical symptoms, vital signs, weight, spleen and liver size), ECG, HIV testing (at baseline only), hemoglobin, white cell count, platelets, urea, creatinine, liver function tests (bilirubin, aspartate aminotransferase, alanine aminotransferase, alkaline phosphatase), urinalysis and audiometry. Audiometry was performed using a standardized procedure by site investigators who were trained by a qualified audiometrist and recorded as hearing levels in dB at 0.25, 0.5, 1, 2, 4 and 8 kHz frequencies [10]. All reported abnormal audiometric readings were reviewed by the audiometrist. An audiometric shift was defined in patients for whom there was one of the following: an increase in hearing level between baseline and EOT of ≥25 dB at ≥1 threshold frequency; an increase in hearing level between baseline and EOT of ≥20 dB at ≥2 adjacent threshold frequencies. Disabling hearing impairment was determined as an average of at least 41 dB across 0.5, 1, 2 and 4 kHz frequencies in adults (ages 15 years and above) and at least 31 db across 0.5, 1, 2 and 4 kHz frequencies in children (less than 15 years of age) [10].
Parasitology slides were prepared from bone-marrow aspirates, read, and reported according to a standardized, approved WHO method [11], [12]. Standardised parasitology readings were done from freshly prepared bone-marrow aspirates taken directly from the patients to the laboratory. Slide fields were examined and counted for parasites under oil emersion 100× magnification for 30 minutes (timed) before being declared negative (absence of parasites on microscopy slide). All parasitology was performed by a trained laboratory technician.
For the PK analysis, the first six consenting patients weighing 30 kg or more were selected from each treatment group and had additional venous blood and urine samples on day 1 and day 14 in the 20 mg/kg/day group, and on day 1 and day 26 in the 15 mg/kg/day group. The timing for blood sampling was 0 (before treatment) and at 0.25, 0.5, 1, 2, 4, 6, 8, 12, and 24 hours after administration of the drug, and at 0–2, 2–4, 4–6, 6–8, 8–12, and 12–24 hours after administration of the drug for urine sampling.
The trial was done in accordance with the Declaration of Helsinki (2002 version) for the conduct of research on human subjects and followed the International Committee for Harmonization guidelines for the conduct of clinical trials. All trial site personnel received relevant training in Good Clinical Practices.
The Ethics Committee of the Institute of Endemic Diseases, University of Khartoum, and the Directorate of Health Research, Federal Ministry of Health, Sudan approved the study protocol (July 8, 2005), which was submitted as a protocol amendment (Protocol S3) to our previous study [7].
All participants or their parents or legal guardians gave their written informed consent before entry into the trial. Children were included in this study because they represent more than 50% of VL cases in this endemic area, and were included in the PK sampling if they met the weight criteria (>30 kg). This study was registered at ClinicalTrials.gov (registration number NCT00255567).
The study medication was 1 g/2mL paromomycin sulphate (Gland Pharma, India). Doses in the study groups were 20 mg/kg/day paromomycin sulphate (equivalent to 15 mg/kg/day of paromomycin base) and 15 mg/kg/day paromomycin sulphate (equivalent to 11 mg/kg/day paromomycin base). The rescue medication was AmBisome (a liposomal formulation of amphotericin B, Gilead, USA), which was reconstituted according to the manufacturer's instructions for a dosage of 3 mg/kg/day for 10 days.
104 patients with suspected VL were screened for entry into this study. Of these, 42 patients were enrolled in the study (21 per group; figure 1). Demographics and baseline characteristics were similar in the two groups (table 1). One patient in the 20 mg/kg/day PM group was considered lost by the 6-month follow-up. The first patient was recruited in October 2005 and the last patient followed-up in October 2006.
Data were available for all patients at EOT (figure 1). 18 patients in the 20 mg/kg group and 19 in the 15 mg/kg group had parasite clearance at EOT, indicating an efficacy of 85.7% (95% CI: 63.7% to 97.0%) and 90.5% (95% CI: 69.6% to 98.0%), respectively (table 2).
At 3-months follow-up, two patients had relapsed in the 20 mg/kg/day for 21 days regimen and three in the 15 mg/kg/day for 28 days regimen; however, there were no additional relapses at 6 months. At 6-months follow-up, the complete-case analysis efficacy in both groups was similar (80.0% in the 20 mg/kg/day group versus 81.0% in the 15 mg/kg/day group) (table 2). All treatment failures were given rescue medication. An exception was one patient in the 20 mg/kg/day PM group who was parasite positive at EOT but clinically responded. This patient was lost to follow-up at 6 months, leading to a lower efficacy estimate in the worst-case analysis (table 2). There was one slow responder (ie, parasite-positive patient at EOT, but clinically well and ultimately recovered) in the group treated with 15 mg/kg/day PM for 28 days.
PM was well tolerated in this study. 48 AEs were reported in total; 20 in the 20 mg/kg for 21 days group and 28 in the 15 mg/kg for 28 days group (table 3), and none was regarded as serious. This gives an AE rate of 0.05 per person-day on treatment in both groups. All AEs, except diarrhea and malaria, were judged to be related to the treatment. The most frequent AE was injection site pain (n = 33). Audiometric shifts were seen in five patients at EOT (n = 3 in the 15 mg/kg group and n = 2 in the 20 mg/kg group), but completely resolved by 6 months follow-up. Disabling hearing impairment, detected at EOT, which improved but persisted at 6 months (ie still met the criteria for audiogram shift), occurred in one patient in the 20 mg/kg group.
Although six patients from each group should have taken part in the PK study, only data from three patients in the 20 mg/kg/day PM group and six in the 15 mg/kg/day PM group were obtained. Only one patient was a child (age of 12 years and weight of 39 kg in the 15 mg/kg group). The others were aged between 17 and 28 years.
Mean plasma PM concentrations at the earlier time points were similar between the two treatment groups (figure 2). Nevertheless, the peak mean plasma PM concentration on day 1 was slightly higher in the 20 mg/kg/day group compared with that in the 15 mg/kg/day group (7.8±4.9 µg/mL versus 5.6±4.2 µg/mL). Six hours after administration, PM was not detected in the plasma of patients receiving 15 mg/kg/day PM but was seen at concentrations slightly lower than peak in the plasma of patients receiving 20 mg/kg/day PM (figure 2).
To date, the standard treatment for VL in East Africa still consists of antimonials. This study is part of the first large-scale multi-centre clinical trial to assess the efficacy of PM for the treatment of VL for the East African region. The initial study [7] showed poor efficacy results when 15 mg/kg/day PM was administered for 21 days to VL patients. This finding is in contrast to an earlier phase III study in India [6]. The results of this study show that increasing the total dose of PM from 15 mg/kg/day for 21 days to 15 mg/kg/day for 28 days or 20 mg/kg/day for 21 days improves efficacy in VL patients in Sudan. However, it should be cautioned that the results found in this study apply to one site only and might not apply to the whole East African region.
Although efficacy is normally assessed as parasite clearance at 6 months in trials for VL, in this study we chose to use parasite clearance at EOT as the primary endpoint because a chance of loss to follow-up of just a few patients would significantly affect the result. In addition to the small sample size, another potential limitation is the use of bone-marrow aspiration for diagnosis and test of cure. However, spleen aspiration remains contraindicated in rural hospitals in Sudan, making bone marrow the best viable alternative.
At 6 months after treatment, efficacy was 80.0% (95% CI: 56.3% to 94.3%) and 81.0% (95% CI: 58.1% to 94.6%) in the 20 mg/kg/day and 15 mg/kg/day groups, respectively, compared with less than 50% (in Sudan) at 6 months observed in the previous study [7]. This result shows that efficacy improved to levels closer to those obtained in trials in India (∼95%) [6].
Serious safety issues that would limit the evaluation of PM at high doses were not identified in this study. Otoxicity, which has been seen as a transient side-effect of PM in other studies [6], was also identified as a potential issue in this study because one patient had audiometric shift at 6 months. This shift occurred at high frequencies, as expected with aminoglycosides [6]. We suggest that this adverse event needs to be monitored closely in subsequent studies.
PK analyses showed that peak plasma PM concentration occurred 1–2 hours after administration and suggest that, at the high daily dose of 20 mg/kg, elevated plasma PM concentrations may be maintained for a longer period of time (up to 8 hours). Unpublished data (Mahmoud Mudawi, personal communication) of PM administration (15 mg/kg) to healthy Sudanese volunteers showed peak PM plasma concentrations similar to those in American volunteers who received a similar dose [14]. Sudanese VL patients had a much lower plasma concentration (30–40%) than that of healthy Sudanese (19.5±7.6µg/mL; n = 6) and American volunteers. Therefore, Sudanese VL patients may have different PK characteristics from both Sudanese and American healthy volunteers, and Indian VL patients. However, PK data were very limited and derived from only a small subset of patients. A PK study with more patients is currently underway as part of the larger phase III study.
Even though interpretation of our results is limited because of the small sample size, we identified what seems to be a more efficacious dose of PM than the one previously used in Sudan [7]. A meeting of the principal investigators was held to discuss the PM efficacy and PK dose-finding results. The group chose to use in the large multi-center phase III study, a dose of 20 mg/kg/day for 21 days for a comparison with the previously used doses of SSG and SSG and PM in combination. Our initial study [7] showed that efficacy of PM can vary greatly between geographical regions, and in addition to this study, suggests that different doses may be required to obtain similar levels of efficacy. If confirmed, these results emphasize the importance of considering regional differences in the treatment of VL and show that drugs of proven efficacy in Asian patients might not have the same efficacy in African patients.
|
10.1371/journal.pntd.0002876 | Sarcocystis nesbitti Causes Acute, Relapsing Febrile Myositis with a High Attack Rate: Description of a Large Outbreak of Muscular Sarcocystosis in Pangkor Island, Malaysia, 2012 | From the 17th to 19th January 2012, a group of 92 college students and teachers attended a retreat in a hotel located on Pangkor Island, off the west coast of Peninsular Malaysia. Following the onset of symptoms in many participants who presented to our institute, an investigation was undertaken which ultimately identified Sarcocystis nesbitti as the cause of this outbreak.
All retreat participants were identified, and clinical and epidemiological information was obtained via clinical review and self-reported answers to a structured questionnaire. Laboratory, imaging and muscle biopsy results were evaluated and possible sources of exposure, in particular water supply, were investigated. At an average of 9–11 days upon return from the retreat, 89 (97%) of the participants became ill. A vast majority of 94% had fever with 57% of these persons experiencing relapsing fever. Myalgia was present in 91% of patients. Facial swelling from myositis of jaw muscles occurred in 9 (10%) patients. The median duration of symptoms was 17 days (IQR 7 to 30 days; range 3 to 112). Out of 4 muscle biopsies, sarcocysts were identified in 3. S. nesbitti was identified by PCR in 3 of the 4 biopsies including one biopsy without observed sarcocyst. Non-Malaysians had a median duration of symptoms longer than that of Malaysians (27.5 days vs. 14 days, p = 0.001) and were more likely to experience moderate or severe myalgia compared to mild myalgia (83.3% vs. 40.0%, p = 0.002).
The similarity of the symptoms and clustered time of onset suggests that all affected persons had muscular sarcocystosis. This is the largest human outbreak of sarcocystosis ever reported, with the specific Sarcocystis species identified. The largely non-specific clinical features of this illness suggest that S. nesbitti may be an under diagnosed infection in the tropics.
| Sarcocystis species are protozoan organisms that have been associated with disease in animals but less frequently so in humans. Following a retreat on Pangkor Island off Peninsular Malaysia, a number of persons presented to our hospital with prolonged fever and muscle pain that was initially difficult to attribute to a known infectious cause. Investigations, including muscle biopsies and PCR, showed that this outbreak was most likely due to Sarcocystis nesbitti infection. The most common clinical features were fever and myalgia that was relapsing-remitting in more than half the patients. Some patients had visible swelling of muscle groups, including of the face, with magnetic resonance imaging also demonstrating inflammation in these muscles. Herein, we present the clinical and investigation findings in 89 symptomatic persons in the largest reported outbreak of human muscular sarcocystosis to date. Our findings provide insights and suggestions for the most appropriate forms of investigation, treatment and possible source of infection.
| Sarcocystis spp. are intracellular protozoan parasites which may involve humans either as definitive or intermediate hosts. Humans are definitive hosts for Sarcocystis hominis and Sarcocystis suihominis, acquired by consuming undercooked sarcocyst-containing beef or pork, respectively. Humans can also become accidental intermediate hosts for other Sarcocystis species by consuming food or water contaminated with fecal sporocysts from an infected definitive host. In such cases, hematogenous dissemination can occur with invasion of muscle leading to sarcocysts [1].
As a disease, sarcocystosis is noted in a variety of animals [2] but symptomatic human disease appears to be less common, with fewer than 150 cases reported in the literature [1], [3], [4]. The prevalence of incidental sarcocysts in humans is also difficult to establish. A previous report showed a series of autopsy tongue muscle collected in the University of Malaya Medical Centre (UMMC), Malaysia, to be positive for sarcocysts in 21% of cases [5], yet, of the more than 1,500 limb muscle biopsies received in the past 20 years in the same centre for routine diagnosis of various symptomatic muscle diseases, none have yielded any sarcocyst-positive tissue (Wong KT, unpublished data).
The largest clustered outbreak of symptomatic muscular sarcocystosis previously reported was in 6 American military servicemen involved in a Malaysian jungle mission [3]. There were also recent reports involving a total of 100 foreign persons with sporadic acute muscular Sarcocystis-like illness after returning from Tioman Island, off the east coast of Peninsular Malaysia, between 2011 and 2012 [4]. The association of S. nesbitti infection with symptomatic human sarcocystosis has also been recognized recently [6], [7]. Herein, we report symptomatic muscular sarcocystosis affecting 89 of 92 persons following a retreat in January 2012 in Pangkor Island, off the west coast of Peninsular Malaysia. This report adds to previous molecular work [6], [7] and limited clinical studies by providing a comprehensive overview of the clinical features and time course of the illness. Furthermore, the role and results of blood and imaging investigations, and muscle biopsy, in influencing diagnosis and a consideration of management options is presented.
A case was defined as a person who attended a specified retreat at a hotel on Pangkor Island, Malaysia, from the 17th to 19th January 2012, and developed relevant clinical symptoms (fever, headache, myalgia and/or arthralgia) within 28 days upon return. Cases were subsequently defined as ‘definite sarcocystosis’ if there was histological demonstration or nucleotide sequences of Sarcocystis spp. from muscle tissue. The remaining cases were defined as ‘probable sarcocystosis’. All persons who attended the retreat, whether or not they had medical reviews at UMMC, submitted self-reported responses to a structured questionnaire aimed at ensuring that all cases were identified. To further elucidate the clinical features, participants were asked about the duration of symptoms, episodes of relapse and to describe the severity of myalgia. Myalgia was defined as ‘severe’, if “excruciating” pain was experienced, ‘moderate’, if daily activities were affected and analgesics required, and ‘mild’ if daily activities were not affected and analgesics not required. Fever referred to a subjective sensation of fever as reported by the patient. If there was any conflicting information between the initial medical review and questionnaire responses, the medical review was taken as more accurate. Participants were also questioned regarding activities and food or water exposure to ascertain potential exposure risks during the retreat. They were also asked if any family members or contacts who did not attend the retreat reported similar symptoms.
To investigate possible sources of infection, water samples were examined as previously described [8]. Ten litre samples were collected from different places along the “gravity-feed” water supply system (up-, middle, down stream of the water source, and from water tanks in the hotel) approximately 3 months after the outbreak.
Investigations included full blood counts (FBC), renal function tests (RFT), liver function tests (LFT), serum creatine kinase (CK) levels, chest x-rays, blood cultures, and blood films for malarial parasites. Results were considered ‘early’ if obtained before 12th February 2012 (less than 4 weeks after the start of the retreat) or ‘late’ if obtained after.
Serological testing was done for chikungunya and dengue viruses, Legionella, Mycoplasma, and Leptospira. Testing was performed using an immunofluorescence assay (in-house) for detection of chikungunya IgM and IgG; anti-dengue IgM and IgG capture ELISA (Standard Diagnostics, Inc, Korea) for detection of dengue IgM and IgG; immunofluorescence assays for detection of Legionella IgG (MarDx Diagnostics, Inc, Ireland) and Legionella IgM (Vircell S.L., Spain); SERODIA-MYCO II (Fujirebi Inc., Japan) for detection of Mycoplasma total antibodies, and microscopic agglutination test (in-house) for detection of Leptospira total antibodies. Sarcocystis serology was done at the CDC by an immunoblot assay using S.neurona merozoite-derived antigens (personal communication, CDC, Atlanta, USA) in 10 patients.
Magnetic resonance imaging (MRI) of skeletal muscles was performed in 8 patients who underwent whole-body coronal T2 weighted, T1 weighted and T2 weighted short inversion time inversion recovery (STIR) scans using the 1.5 T SignaHDx MR Systems (GE Healthcare, USA).
Muscle biopsies from affected sites in 4 patients with myalgia and MRI abnormalities were fixed in buffered 10% formalin and routinely processed. Hematoxylin and eosin stained tissue sections were examined by light microscopy. Polymerase chain reaction (PCR) for Sarcocystis spp. was performed on all 4 biopsies (6,7).
Statistical analysis was performed using chi-square testing and Fisher's exact test to compare the categorical variables in relation to patients' nationality and sex. The Mann-Whitney U-test was utilized to compare the median duration of symptoms between nationalities and sexes while Spearman's rank correlation and independent samples t-test were used to analyze the relationship between age and severity of myalgia, and duration of symptoms, respectively.
IBM SPSS Statistics version 21.0 (Armonk, NY: IBM Corp) was used for statistical analysis. P values<0.05 were considered significant.
This investigation was undertaken in response to the presentation of acutely ill patients to our institution with the intent of determining the infectious agent responsible for the outbreak. Patients and their guardians were informed throughout their management that investigations were undertaken to ensure that critical illness did not develop in any person, to identify the causative organism, and, if possible, to determine the mode of acquisition of infection and henceforth, to prevent further infection. Ethics or IRB approval was not requested for this outbreak investigation. All patients who underwent muscle biopsy, the only invasive test, gave written informed consent. Written consent was obtained from the patient in Figure 1 for publication of the photograph.
From 17th to 19th January 2012, a group of 92 college students and teachers attended a retreat at a hotel located on Pangkor Island, Malaysia. There were 55 males and 37 females, with a median age of 35 years (IQR 27 to 44; range, 5 to 64). Seventy-one were Malaysians and 21 non-Malaysians. A total of 58 cases were medically assessed at our institute, and 9 were admitted to hospital.
From a review of available medical records and questionnaire responses, it was determined that 89 persons demonstrated or reported features consistent with the case definition. Of these, 53 were male and 36 female with a median age of 34 years (IQR 27 to 43.5). Sixty-nine (77.5%) symptomatic persons were Malaysians and 20 (22.5%) were non-Malaysians: 9 from China, 3 from Nepal, 2 each from Korea and Indonesia and 1 person each from India, Netherlands, Myanmar and Iran, respectively. Malaysians were older with a median age of 38 years (IQR 27.50 to 45.00) compared to 29·5 years (IQR 24.25 to 33.75) for non-Malaysians (p = 0.012).
The onset of symptoms could be accurately ascertained for 82 patients and occurred between day 1 and day 26 post-retreat. Symptoms began between day 9 and 11 for 58 (70.7%) of these patients. The most common symptoms and their frequencies are listed in Table 1. Retreat participants did not report any similar illness in colleagues or family members who did not attend the retreat.
Overall, the total duration of symptoms lasted from 3 days to nearly 4 months with a median duration of 17 days (IQR 7 to 30). Twenty-seven patients (30.3%) experienced symptoms for 1 month or longer, myalgia being the most prolonged symptom. There was no association between age and duration of symptoms, and no significant difference in median age between the sexes. For non-Malaysians, the median duration of symptoms was 27.5 days (IQR, 17.3 to 42.0) which was significantly longer than the 14 days (IQR, 7.0 to 29.5) experienced by Malaysians (p = 0.001).
In the “early” phase of illness, myalgia was particularly marked in the neck muscles. Later it was experienced most commonly in the lower limbs followed by the back and the upper limbs. Among the 73 persons who graded their myalgia for severity, severe myalgia was reported by 10 persons (13·7%), moderate myalgia by 27 (37·0%), and mild myalgia by 36 (49.3%). There was no significant difference in the median age of those with mild or moderate/severe pain and no significant difference between the sexes. Sixty-two (89.9%) Malaysians and 19 (95.0%) non-Malaysians reported myalgia as a symptom but this was not significantly different. However, 15 out of 18 (83.3%) non-Malaysians reported myalgia as moderate/severe compared to 22 out of 55 (40.0%) Malaysians (p = 0.002). Thus, non-Malaysians were more likely than Malaysians to experience moderate/severe myalgia.
Relapsing fever was reported in 48 of the 84 (57.1%) patients with fever, and 15 (31·3%) of the relapsing fever cases had 3 or more cycles of fever. Each symptomatic episode lasted a median of 5 days (IQR, 3 to 7; range, 1 to 21) and each remission 4 days (IQR 3 to 7; range, 1 to 30). There was no particular pattern regarding the duration of the first or subsequent cycles. Of the 9 patients admitted to hospital, 8 recorded temperatures greater than or equal to 39°C.
Between 34 and 38 days post-retreat, 9 patients developed visible facial muscle swelling with resolution in all patients within 7 days (Figure 1). One of these patients also reported swelling of the thenar eminence of one hand. Four patients presented with calf muscle swelling and another patient had isolated swelling of the interossei muscles of one hand. The higher proportion of non-Malaysians with muscle swelling (5 cases; 25.0%) compared to non-Malaysians (9 cases; 13.0%) was not statistically significant.
Cough was reported as non-productive, transient, and seen only in the first 2 weeks of illness.
Blood cultures from 21 patients were negative for bacteria or fungal growth. Initial chikungunya IgM serology using immunofluorescent staining (IFAT) was positive in 53% (43/81) of patients. However, chikungunya IgG serology up to two weeks after onset of illness confirmed seroconversion in only one patient.
Ten paired patient sera were serologically tested for sarcocystosis including 3 ‘definite sarcocystosis’ cases. There were 3 positive results in both acute and convalescent sera, 2 seroconverted positives, 1 equivocal, and 4 cases that were negative for both acute and convalescent samples. In five patients who had muscle swelling, 3 had positive results. One of the definite cases of sarcocystosis had negative serology.
None of the other serology tests for dengue, Legionella, Mycoplasma or Leptospira were consistent with recent infection.
FBC, LFTs, and CK abnormalities varied according to the time-point of illness. Those parameters that were most commonly abnormal ‘early’ or ‘late’ in the course of the infection are shown in Table 2. Only those symptomatic patients who were reviewed in the hospital had blood tests performed in the early period.
Eight patients who had facial swelling and/or myalgia 4–5 weeks after onset of illness underwent MRI examination. Three muscle groups, muscles of mastication, calf, and superficial back muscles demonstrated focal and heterogeneous high signal intensities on STIR consistent with inflammatory edema and myositis. Non-affected muscles demonstrated low signals on STIR. All 8 patients had edema/myositis in the muscles of mastication (superficial temporalis, deep temporalis and masseter muscles) (Figure 2). These were bilateral in 5 and unilateral in 3 patients. Four of these 8 patients had exhibited or reported facial swelling. Four patients had MRI changes in the back muscles and 2 in the calf muscles (Figure 3).
Of the patients who underwent MRI, 2 had both eosinophilia and raised CK, 3 had eosinophilia alone and 3 had only raised CK.
Four biopsies were taken from the temporalis (1 case), tibialis posterior (1 case), and gastrocnemius (2 cases) muscles (Table 3). In 3 biopsies (temporalis, tibialis posterior, gastrocnemius), one or more sarcocysts were detected within muscle fibres by light microscopy (Figure 4). A whole sarcocyst was also obtained in a muscle tissue culture preparation from the temporalis muscle. In all 3 biopsies with sarcocysts, there were no inflammatory cells immediately surrounding the infected muscle fibre, but mild to moderate inflammation and focal necrosis was noted in other parts of the muscle. Inflammatory cells were mixed and eosinophils were generally not prominent.
We attempted to determine the 18S rDNA sequences from the sarcocysts and/or muscle tissue of 3 biopsies (temporalis, tibialis posterior and gastrocnemius muscles) by polymerase chain reaction (PCR) and direct sequencing [6]. S. nesbitti sequences (accession numbers HF544323 and 544324) were confirmed in 2 patients (temporalis and gastronemius muscles). The 18S rDNA gene shared 100% identity with S. nesbitti found in the muscle of Macaca fascicularis from Yunnan Province, China [9], [10], [11]. The presence of S. nesbitti (accession number JX661499.1) was subsequently confirmed by nested polymerase chain reaction and sequencing from the gastrocnemius muscle of the fourth patient (7).
Thus in our series, 4 cases were classified as definite: Three Malaysians and one non-Malaysian. The clinical features and laboratory findings of these 4 cases are shown in Table 3.
As the etiological agent was not determined until late in the course of illness, only 3 patients received medical treatment aside from basic analgesia. Two patients received oral corticosteroids with 1 patient reporting resolution of symptoms soon after commencement of treatment but the other did not report any change.
A third patient received targeted therapy after diagnosis of biopsy confirmed sarcocystosis 10 weeks into illness. With pre-existing renal impairment, clindamycin 600 mg qid po and fansidar (sulfadoxine 500 mg/pyrimethamine 25 mg) 2 tablets/weekly po were prescribed. After 3 days of treatment, the patient reported that pain in the proximal arm muscles and thighs had reduced significantly and proximal arm strength had improved. There was a decrease in CK from 782 to 653 but this had been falling prior to institution of medications. Fansidar was taken for 2 weeks with an interrupted course of clindamycin for 6 weeks. At this point there was no recurrence of myalgia. The patient was subsequently lost to follow-up.
Water and food contamination were investigated as possible sources of Sarcocystis infection. The hotel is located adjacent to a forested area and about 500 m from the beach. The water supply to the hotel came from two sources: chlorinated (treated) water from the mainland and untreated “gravity-feed” water piped down from forested hillsides on the island. Treated water was meant for drinking and preparing food. Untreated water, intended for bathing and washing, was apparently filtered and stored in large outdoor tanks. There was a temporary breakdown in the treated water supply just prior to the retreat. There were reports of the water appearing “cloudy” at the beginning of the retreat and it is possible that untreated water was inadvertently used to prepare food and drinks during the retreat. All 89 symptomatic persons who attended the retreat drank water/beverages prepared in the hotel. As for the 3 persons who did not fall sick, one person drank only Chinese tea prepared with boiled water, while the other two reported drinking ‘mostly’ bottled water. All 92 participants ate meals prepared by the hotel.
Sarcocystis spp.were not detected by PCR in any water samples.
This outbreak has previously been described in relation to the molecular identification of S. nesbitti as a cause of human muscular sarcocytosis [6], [7]. This report adds to the previous work by providing a comprehensive description of the clinical features of the illness in a large group of affected persons. The focal point of exposure allows for the potential incubation period to be estimated and the temporal changes in both the clinical features and blood investigation results are clearly demonstrated. The important role of MRI is also shown with, for the first time, selected images of the most evident changes. Other aspects of the illness, including treatment options and a possible modification of illness in Malaysians, are also discussed. This report incorporates the clinical, investigative and management aspects of this illness, adding substantially to the previously reported work.
Our findings strongly suggest that muscular sarcocystosis was responsible for the outbreak of acute relapsing febrile myositis in our cohort. Four patients had definite sarcocystosis with sarcocysts or nucleotide sequences demonstrated in the muscles that were shown to be involved by MRI. The other 85 patients had probable muscular sarcocystosis as suggested by the clustered time of onset and recovery from symptoms, overall similarity of symptoms to definite cases, and absence of similar illnesses among patient contacts who did not attend the retreat.
Although sarcocysts have previously been noted as incidental findings in tongue muscles in an autopsy series [5], we believe that their presence in our cases represents a pathological process. Sarcocysts have never been detected in more than 1500 limb muscle biopsies examined at the University of Malaya Medical Centre (Wong KT, unpublished data), thus finding sarcocysts in inflamed skeletal muscles from sites other than the tongue is unlikely to be incidental. In addition, with more than 120 Sarcocystis spp. reported in animals [12] it seems unlikely that 3 cases from our cohort could co-incidentally have the same species identified by PCR, with 3 cases also demonstrating sarcocysts of very similar morphology. Although a few previous human studies have shown that febrile myalgia is associated with biopsy-proven, muscular sarcocystosis [3], [4], [13]–[15], this is the first time that S. nesbitti has been identified as a cause of symptomatic human muscular sarcocytosis.
Our results suggest that S. nesbitti has a very high 97% attack rate. With an identifiable common period of exposure lasting 3 days, the incubation period was determined as most likely to be between 9 and 13 days, but could be up to 28 days. The most common symptoms, fever (94%) and myalgia (91%), were non-specific making initial diagnosis difficult. However, the majority (57.1%) of cases had relapsing fever. Nine patients exhibited facial swelling and an additional 4 had MRI changes consistent with myositis involving the muscles of mastication. Some of the clinical features in the present outbreak appear similar to those described previously. An outbreak involving 6 American military personnel who were believed to be infected in Malaysia, showed that nearly half of them had prolonged fever, myositis and raised liver enzymes [3]. However, there were also reports of bronchospasm, rashes, lymphadenopathy and marked eosinophilia, features not found in our patients. One patient had bitemporal muscle tenderness but facial swelling was not reported. Facial swelling as a striking clinical manifestation has also been reported more recently in a Dutch traveler to Tioman Island, Malaysia [15].
Earlier reports of human muscular sarcocystosis often referred to the disease as an ‘eosinophilic myositis’ [3], [14]. However, we think it important not to exclude this diagnosis on the basis of the eosinophil count since this could be normal early in disease. Normal eosinophil counts were also observed in travellers returning from Tioman Island with possible Sarcocystis-like myositis [4]. In fact, in our cohort, lymphocytosis appeared to be as common as eosinophilia throughout the illness.
The apparent clinical differences between our cases and previous reports could be due to a number of reasons. Firstly, it is possible that a pathogenic Sarcocystis spp. different from S. nesbitti was involved as this is the first time species identification has been successfully performed in human disease. Secondly, the time points at which patients were reviewed appears to affect results as seen in the variations in eosinophil counts and serum CK levels within individual subjects over time. Finally, as the current outbreak occurred within a single community, it is likely that we may have included some patients with milder symptoms who would otherwise not seek medical attention,
Whole body MRI in STIR sequence was helpful to guide the choice of suitable muscle biopsy sites, resulting in a definitive diagnosis in all 4 patients biopsied. This is the highest number of biopsy-proven cases from a single outbreak of symptomatic sarcocystosis, and MRI had an essential role to play. Despite the widespread myalgia, the abnormal MRI changes were confined to 3 distinctive groups comprising of the muscles of mastication, the calves and the back. As discussed, the involvement of the muscles of mastication represents an interesting disease manifestation. In contrast to the utility of MRI, serology performed in a limited number of patients did not appear to be sufficiently sensitive, and this warrants further investigation before diagnostic use in humans can be recommended.
Since clinical manifestations and laboratory investigations may be non-specific, isolated cases of muscular sarcocystosis could be easily missed, and therefore this condition is also likely to be under diagnosed. Nonetheless, prolonged relapsing fever and severe myalgia that lasts several weeks, facial and other muscular swellings, raised LFT, CK, eosinophilia and lymphocytosis could suggest the diagnosis. MRI in STIR sequence, by demonstrating muscles affected by myositis, is a potentially useful tool to guide muscle biopsy to confirm the diagnosis. It cannot be overemphasized that muscle biopsy is most important to confirm the diagnosis and should be done whenever possible. In fact, the definitive diagnosis was only made after muscle biopsy. As our study shows, even in the absence of sarcocysts, the PCR of muscle tissue may be still be useful for diagnosis if clinical suspicion is high.
Interestingly, our study suggests that the only significant risk factor for prolonged disease and moderate/severe myalgia is the patient's nationality. Non-Malaysians were more likely to suffer more prolonged disease and to complain of moderate/severe myalgia than Malaysians. Although perception of severity of pain may be subjective and possibly influenced by cultural values, duration of illness is probably less subjective, suggesting that the observed statistical significance, at least for duration of illness, may indeed be true. The histological demonstration of sarcocysts in 21% of Malaysians along with other reports of incidental findings of sarcocysts in Malaysians suggest that exposure to this organism may not necessarily result in symptomatic infection [5], [16]. However, we speculate that asymptomatic infection could lead to partial immunity which could help explain the shorter duration of symptoms and less severe myalgia experienced by Malaysians in the cohort. It may also explain why all previous reported cases of symptomatic muscular sarcocystosis from Malaysia involved non-Malaysians [3], [4], [13], [15].
There are no conclusive recommendations for the treatment of sarcocystosis [1] although previous reports raise the possibility of clinical improvement with steroids or albendazole but recovery was not rapid [3]. In animals, medications including clindamycin, pyrimethamine and sulphonamides have been suggested as treatment options [17], [18]. There was a previous report of symptomatic improvement and resolution of blood test abnormalities in human sarcocystosis with the use of co-trimoxazole [14]. We elected to treat one patient with prolonged symptoms, as we would a case of toxoplasmosis based on the premise that both organisms are intracellular protozoa. We also elected to utilize medications that have been effective in animals but co-trimoxazole was avoided due to renal insufficiency. Hence the patient received clindamycin and fansidar. It is difficult to attribute the patient's subsequent improvement solely to the medications prescribed but we believe such treatment is worthy of consideration in future cases.
S. nesbitti was first described in muscle tissues of Macaca mulatta monkeys by Mandour in 1969 [19]. An ultrastructural study of human muscular sarcocystosis in Malaysia has previously shown a similarity to sarcocysts found in Macaca fascicularis (Malaysian long tailed macaque) [20]. Phylogenetic analysis suggests that snakes could be a definitive host [10], [19]. More recently, S. nesbitti sequences have been detected in the feces of snakes from disparate parts of Malaysia, confirming that snakes may be the definite hosts [7], [11]. As such, S. nesbitti is likely to be endemic in the country. We have no evidence for any recent change in ecology of snakes or human/snake interaction to account for the recent cases although further investigations are needed. Reports of 100 foreign persons with an acute muscular Sarcocystis-like illness after returning from Tioman Island, Malaysia, between 2011 and 2012 [4] could either suggest a reemerging infection or an increasing recognition of an endemic disease. It is possible that S. nesbitti could cause a clinically more apparent disease compared to other Sarcocystis spp. that could potentially cause human sarcocystosis. Moreover, it is not known if the severity of clinical disease is dependent on the dose of sporocysts consumed or if a low dose results in largely asymptomatic infection.
Infection in this outbreak is likely due to consumption of water contaminated by snake feces. Almost all the water supply in Malaysia are sourced from streams or rivers, so contamination by snake or other reptile feces may be common. We believe that drinking gravity-feed water contaminated by snake feces or ingestion of uncooked food washed therein, is most likely to be responsible for this outbreak. Water treatment by chlorination may not be able to kill sporocysts [21], thus contaminated piped water from the mainland could still be responsible, albeit less likely as the contaminants are likely to be diluted by a significantly larger water supply. Unsafe water supply, especially in the islands and other remote places, poses potentially serious public health risks that will require urgent investigation and intervention. Travellers should be advised to only drink boiled or bottled water and to avoid uncooked food [21]. There should also be heightened awareness for this infection in patients who have recently lived in or travelled to Southeast Asia.
There were some limitations to our study. Firstly, since it was commenced as an outbreak investigation and the diagnosis was not known until well into the course of the illness, investigations and their timing could not be more uniformly undertaken for all the patients. Secondly, in determining the severity of myalgia the adopted definitions may not be mutually exclusive, and other factors, including cultural, may also have influenced the perception of pain. Finally, due to ethical and other considerations, it was neither possible nor appropriate to perform further muscle biopsies.
|
10.1371/journal.ppat.1002159 | Effects of Interferon-α/β on HBV Replication Determined by Viral Load | Interferons α and β (IFN-α/β) are type I interferons produced by the host to control microbial infections. However, the use of IFN-α to treat hepatitis B virus (HBV) patients generated sustained response to only a minority of patients. By using HBV transgenic mice as a model and by using hydrodynamic injection to introduce HBV DNA into the mouse liver, we studied the effect of IFN-α/β on HBV in vivo. Interestingly, our results indicated that IFN-α/β could have opposite effects on HBV: they suppressed HBV replication when viral load was high and enhanced HBV replication when viral load was low. IFN-α/β apparently suppressed HBV replication via transcriptional and post-transcriptional regulations. In contrast, IFN-α/β enhanced viral replication by inducing the transcription factor HNF3γ and activating STAT3, which together stimulated HBV gene expression and replication. Further studies revealed an important role of IFN-α/β in stimulating viral growth and prolonging viremia when viral load is low. This use of an innate immune response to enhance its replication and persistence may represent a novel strategy that HBV uses to enhance its growth and spread in the early stage of viral infection when the viral level is low.
| Hepatitis B virus (HBV) is a major human pathogen that can cause severe liver diseases including hepatitis, liver cirrhosis and hepatocellular carcinoma. Approximately 350 million people worldwide are chronic carriers of this virus. Type I interferons (IFNs), which include IFN-α and IFN-β, are produced by the host to control microbial infections. However, the use of IFN-α to treat HBV patients has generated inconsistent results. By using mice as an animal model, we have investigated the effect of type I IFNs on HBV replication in vivo. Our results indicate that IFN-α/β can suppress HBV replication when viral load is high and enhance HBV replication when viral load is low. These effects of IFN-α/β on HBV are due in part to their abilities to regulate HBV gene expression. Our further studies reveal an important role of IFN-α/β in stimulating viral growth when viral load is low. This use of an innate immune response to enhance its replication may represent a novel mechanism that HBV uses to enhance its growth and spread in the early stage of infection when the viral level is still low.
| Interferon-α (IFN-α) and interferon- β (IFN- β) are type I interferons, which are produced by the host in response to viral infections to inhibit viral replication [1]. After binding to its receptor, IFN-α/β activates the Janus kinase (JAK) and its downstream signal transducer and activator of transcription (STAT) and induces the expression of more than 300 IFN-stimulated genes (ISGs) and many antiviral proteins [2], [3]. IFN-α has been used to treat viral infections including hepatitis B virus (HBV), which chronically infects approximately 350 million people in the world. Unfortunately, IFN- α generates sustained virological response in only a minority of patients [4]. Little is known why the majority of HBV patients do not respond to the IFN-α therapy.
HBV is a small DNA virus that infects liver. Its genome is only 3.2 Kb in size and consists of four genes: the C gene codes for the viral core protein that forms the viral capsid and a related protein termed precore protein, which is the precursor of the secreted e antigen (HBeAg); the S gene codes for the viral envelope proteins, also known as surface antigens (HBsAg); the P gene codes for the viral DNA polymerase; and the X gene codes for a regulatory protein. To understand why IFN-α generates different responses in HBV patients, we studied the effect of IFN-α/β on HBV replication using mice as a model. Interestingly, we found that interferons could suppress HBV replication when viral load is high and enhance HBV replication when viral load is low. The suppression of HBV replication by IFN-α/β apparently involves both transcriptional and post-transcriptional regulations whereas the enhancement of HBV replication by IFN-α/β is mediated by transcription factors HNF3γ and STAT3. This use of type I interferons induced by its infection to enhance its replication thus represents a novel strategy that HBV may use to stimulate its growth and spread in the early stage of viral infection when the viral level is still low.
We have previously produced four HBV transgenic mouse lines that carry either the wild type HBV genome (Tg05 and Tg08 mouse lines) or the mutated HBV genome that is incapable of expressing only the HBV X protein (HBx) (Tg31 and Tg38 lines) [5], which is a regulatory protein. These mouse lines contain replicating HBV DNA in the liver and produce mature viral particles in the blood (Figure S1A and S1B). To examine the possible effects of IFN-α/β on HBV in vivo, HBV transgenic mice were injected intravenously with the IFN-α/β inducer poly(I∶C), or with saline, which served as the control. Mice were sacrificed 12 hours or 24 hours after injection for the studies. The effect of poly(I∶C) on HBV DNA replication in the liver was analyzed by Southern blot. The level of HBV DNA replicative intermediates (RI) in the Tg05 mouse liver at 12 hours and 24 hours after poly(I∶C) injection was reduced by 54% and 80%, respectively (Figure 1A). When the HBV RNA was analyzed by Northern-blot, a slight reduction of the level by poly(I∶C) was also observed, particularly with the HBV C gene transcripts. When the HBV core protein was analyzed by Western-blot, its reduction by poly(I∶C) was inapparent at 12 hours, likely due to the stability of this protein. However, its reduction was apparent at 24 hours after injection (Figure 1A). For the Tg38 mouse line, the HBV DNA in the liver was also reduced in a time-dependent manner, although by a lesser degree (36% and 61% reduction at 12 hours and 24 hours, respectively, after injection). The reduction of the HBV RNA levels was not obvious. However, the core protein reduction was apparent at the 24-hour time point (Figure 1A). In contrast to Tg05 and Tg38 mouse lines, poly(I∶C) increased the HBV DNA, RNA and core protein levels in the liver of Tg31 and Tg08 mice (Figure 1A). These results indicated that poly(I∶C) could have different effects on HBV, depending on the mouse lines.
To determine whether the effect of poly(I∶C) on HBV was mediated by interferons, we first tested whether poly(I∶C) could indeed induce the interferon response in all four mouse lines. Total mouse liver RNA was isolated and analyzed for the expression of 2′-5′ oligoadenylate synthetase (2′,5′-OAS), a gene activated by IFN-α/β, by semi-quantitative reverse transcription PCR (RT-PCR). The expression of 2′,5′-OAS was indeed induced in the liver of all four mouse lines, indicating the induction of interferon response by poly(I∶C) (Figure S2A). We next injected HBV transgenic mice with antibodies directed against IFN-α/β one day prior to the injection of poly(I∶C). The administration of anti-IFN-α/β antibodies, but not the control antibody, abolished the induction of 2′,5′-OAS, indicating the ability of these anti-IFN-α/β antibodies to inhibit the activities of IFN-α/β in the liver (Figure S2B). Finally, we analyzed the effect of anti-IFN-α/β antibodies on HBV replication. The administration of anti-IFN-α/β antibodies, but not the control antibody, prevented poly(I∶C) from reducing the levels of HBV RI DNA, RNA and core protein in Tg05 and Tg38 mice and from increasing their levels in Tg08 and Tg31 mice (Figure 1B). If anti-IFN-α and anti-IFN-β antibodies were administered separately, the latter was found to be more efficient than the former in blocking the effect of poly(I∶C) (Figure S2C). This result might be due to the preferential induction of IFN-β by poly(I∶C) [6], or the difference in activities of these two cytokines [7]. We had also tested directly the role of IFN-α/β in the regulation of HBV replication by injecting Tg05 and Tg31 mice with IFN- α and IFN-β. Our results indicated that IFN- α had only a slight effect on HBV in these two mouse lines whereas IFN-β significantly suppressed HBV replication in Tg05 mice and enhanced HBV replication in Tg31 mice (Figure S2D).
In the studies shown in Figure 1, HBV transgenic mice were only treated with poly(I∶C) for up to 24 hours. To determine whether the effects of poly(I∶C) on HBV could persist, we analyzed its effects on HBV for one week. As shown in Figure S2E, the effects of poly(I∶C) on HBV could persist for one week, the endpoint of the analysis.
Our results thus indicated that the effect of IFN-α/β on HBV could vary depending on the mouse lines. This effect of IFN-α/β on HBV is independent of the HBx protein, as Tg05 and Tg08 mice carried the wild-type HBV genome and yet responded in opposite ways to IFN-α/β. Similarly, Tg31 and Tg38 carried the X-null HBV genome and also responded differently to IFN-α/β. However, there appeared to be a viral load-dependent effect, as IFN-α/β suppressed HBV replication in Tg05 and Tg38 mice, which produced higher levels of HBV, whereas they enhanced HBV replication in Tg08 and Tg31 mice, which produced lower levels of HBV (Figure S1A and S1B).
To examine whether the effect of IFN-α/β on HBV is indeed dependent on viral load, we performed the hydrodynamic injection, which is a rapid and convenient method for gene delivery into the mouse liver [8]. In this study, different amounts of the 1.3mer, over-length HBV DNA genome were injected via the tail vein into mice. Increasing the amount of HBV DNA in the injection led to an increasing level of HBV RI DNA, HBV RNA and the core protein in the liver until the amount of HBV DNA reached 24 µg (Figure 2A). Further increase of the HBV DNA amount to 32 µg for injection did not increase, but rather, decreased HBV RI DNA, HBV RNA and core protein levels in the mouse liver, perhaps due to the reduction in DNA delivery efficiency (Figure 2A). There was a positive correlation between the levels of HBV RI DNA in the liver and HBV titers in the sera (Figure S3). To test whether the effect of poly(I∶C) on HBV replication is dependent on viral load, mice were injected with ploy(I∶C) three days after the hydrodynamic injection of HBV DNA and sacrificed 24 hours later for HBV replication studies. Poly(I∶C) increased HBV DNA, RNA and core protein levels when the HBV DNA used for the injection was 4, 8 or 14 µg. However, poly(I∶C) decreased HBV DNA, RNA and core protein levels when the amount of HBV DNA used for the injection was 20, 24 or 32 µg (Figure 2A).
To determine whether the effect of poly(I∶C) on HBV in this hydrodynamic injection study was also mediated by IFN-α/β, we also injected mice with either the control antibody or the anti-IFN-α/β antibodies. The control IgG had no effect on the increase of HBV DNA, RNA and core protein levels induced by poly(I∶C) when mice were injected with 8 µg HBV DNA. However, this increase was abolished by anti-IFN-α/β antibodies. Similarly, although the control IgG had no effect on the decrease of HBV DNA, RNA and core protein levels by poly(I∶C) when mice were injected with 20 µg HBV DNA, anti-IFN-α/β antibodies diminished the suppressing effect of poly(I∶C) on HBV (Figure 2B).
The results indicate that the effect of IFN-α/β on HBV is dependent on viral load. They enhance HBV replication when viral load in the serum is low and suppress HBV replication when viral load is high. The viral DNA level in the serum that separates these two opposite responses appears to be in the vicinity of 107 copies/ml (Figure S3), as the HBV replication was stimulated by interferons when the viral DNA level in the serum was lower than 107 copies/ml whereas it was suppressed when the viral DNA level was higher than 107 (Figure 2A and Figure S3).
Previous studies indicated that IFN-α/β suppressed HBV replication in transgenic mice that produced high levels of HBV by inhibiting the assembly of the viral capsid or by accelerating their degradation [9]. Our poly(I∶C) injection results indicated that IFN-α/β could also affect HBV RNA transcription or stability, as the reduction of HBV RNA levels in the liver of mice that produced high levels of HBV was also apparent after the injection of poly(I∶C) (e.g., Figure 2A and 2B). However, in mice that produced a low level of HBV, IFN-α/β significantly increased the levels of both HBV C gene and S gene RNA transcripts, suggesting that IFN-α/β might enhance HBV replication in these mice by enhancing the transcription of HBV genes, possibly by acting on the two HBV enhancers, which have global effects on HBV gene expressions [10]. To test this possibility, HBV DNA fragments containing enhancer I and its overlapping X promoter (ENI/Xp) or enhancer II and its overlapping C promoter (ENII/Cp) were linked to the firefly luciferase reporter (Figure 3A). These DNA constructs were delivered together with the 1.3mer HBV genomic DNA into the mouse liver by hydrodynamic injection. Poly(I∶C) decreased the expression of the firefly luciferase approximately three-fold when the ENI/Xp reporter construct was co-injected with 0 or 20 µg 1.3mer HBV genomic DNA. However, it increased the expression level of the luciferase three to four-fold when the reporter construct was co-injected with 8 µg 1.3mer HBV genome (Figure 3B). In contrast, poly(I∶C) reduced the expression level of the luciferase reporter from the ENII/Cp construct, regardless of whether this reporter construct was co-injected with 0, 8 or 20 µg HBV genomic DNA (Figure 3C). These results indicated that poly(I∶C) most likely activated HBV gene expression when the HBV DNA level was low by activating the ENI/Xp complex. Since poly(I∶C) could not activate the ENI/Xp complex in the absence of HBV genome (Figure 3B), this result also indicated an essential role of a low HBV genomic DNA level for poly(I∶C) to exert its enhancing effect. Note that, in the absence of poly(I∶C), the expression level of luciferase from the ENI/Xp reporter construct was higher in the presence of 20 µg HBV genomic DNA, and its expression level from the ENII/Cp reporter construct was higher in the absence of HBV DNA. The reason for these differences is unclear, but it might be related to the activities of HBV gene products on ENI/Xp and ENII/Cp complexes [11], [12], [13].
The ENI/Xp complex consists of multiple transcription factor binding sites [10], [14]. To further identify the IFN-α/β responsive element in this complex, we conducted the deletion-mapping experiments. The deletion of the sequence upstream of nt. 1136 was sufficient to abolish the stimulatory effect of poly(I∶C) on the ENI/Xp complex (Figure 3D), suggesting that the IFN-α/β responsive element resides at nt. 1115–1136. This sequence has previously been shown to contain HNF3 and STAT3 transcription factor binding sites [15], [16], [17]. To test whether the HNF3 binding site was indeed the IFN-α/β responsive element, we introduced mutations, which have previously been shown to abolish the HNF3 binding [15], into the ENI enhancer (Figure 3A). These mutations indeed abolished the response of ENI/Xp to IFN-α/β, confirming the role of the HNF3 binding site in mediating the effect of IFN-α/β (Figure 3D).
There are three isoforms of HNF3, which are HNF3α, HNF3β and HNF3γ. All these three isoforms could bind to the HNF3 site in the HBV ENI enhancer [15]. To determine whether the expression of these three HNF3 isoforms was affected by poly(I∶C), we conducted the Western-blot analysis on the nuclear extracts of the mouse liver. While poly(I∶C) had no effect on the levels of HNF3α, HNF3β and the control lamin-β protein in the liver of all four of our HBV transgenic mouse lines, poly(I∶C) specifically increased the level of HNF3γ in Tg08 and Tg31 mice but not in Tg05 and Tg38 mice (Figure 4A). To further study the role of these three different HNF3 isoforms on the HBV ENI enhancer activity, the expression plasmids for these three different isoforms were separately co-transfected with the ENI/Xp reporter into Huh7 cells, a human hepatoma cell line. The over-expression of HNF3γ, but not HNF3α or HNF3β, could enhance the ENI/Xp activity in a dose-dependent manner (Figure 4B).
The results suggested that IFN-α/β might induce the expression of HNF3γ to activate the ENI enhancer in mice with a low HBV level. To test this possibility, we decided to use the shRNA to suppress the expression of HNF3γ. We first verified that the HNF3γ shRNA could indeed suppress the expression of HNF3γ and reduce the ENI/Xp activity in Huh7 cells (Figure S4). We next co-delivered the expression plasmid of this HNF3γ shRNA with the ENI/Xp reporter construct and 8 µg 1.3mer HBV DNA into the mouse liver by hydrodynamic injection. Poly(I∶C) could activate the HBV ENI/Xp complex in the presence of the control shRNA but not in the presence of the HNF3γ shRNA (Figure 4C). Similarly, the activation effect of poly(I∶C) on HBV DNA replication and RNA transcription was also abolished by the HNF3γ shRNA in mice injected with 8 µg 1.3mer HBV DNA (Figure 4D). These results demonstrated that the enhancing effect of IFN-α/β on HBV replication was mediated by HNF3γ.
Previous studies indicated that HNF3 and STAT3 could bind to each other and cooperatively stimulate the HBV ENI enhancer [17]. Our finding that IFN-α/β induced the expression of HNF3γ to stimulate the HBV ENI enhancer prompted us to examine whether IFN-α/β also affects STAT3 in mice with a low HBV level. Although poly(I∶C) had no effect on STAT3 in Tg05 mice that produced a high level of HBV, it activated STAT3 in Tg08 mice, which produced a low level of HBV, as evidenced by the increased level of phosphorylated STAT3 (p-STAT3) and its association with the nuclear fraction (Figure 5A). To further determine the role of STAT3 in mediating the effect of IFN-α/β on HBV gene expression, we also conducted the shRNA-knockdown experiment to suppress the expression of STAT3. We first analyzed the effect of the STAT3 shRNA and demonstrated that it could reduce the expression level of STAT3 by approximately 40% and a similar level of the ENI/Xp activity in a reporter assay in Huh7 cells (Figure S5). We then injected mice with the expression plasmid for the STAT3 shRNA or a control shRNA, the ENI/Xp reporter construct and 8 µg HBV genomic DNA, and analyzed the effect of poly(I∶C) on the ENI/Xp activity. The STAT3 shRNA reduced the activation effect of poly(I∶C) on the ENI/Xp complex (Figure 5B). Similar to the HNF3γ results, the STAT3 shRNA, but not the control shRNA, also partially reduced the enhancing effect of poly(I∶C) on HBV DNA replication and RNA transcription (Figure 5C). These results indicated that STAT3 also played an important role in mediating the enhancing effect of IFN-α/β on HBV gene expression. The lack of complete inhibition of the effects of poly(I∶C) by the STAT3 shRNA was probably due to the inefficiency of this shRNA to knockdown the expression of STAT3 (Figure S5).
The observation that the effect of IFN-α/β on HBV is dependent on viral load prompted us to investigate how that effect may affect viral growth in vivo. We injected mice with either 4 µg or 20 µg1.3mer HBV genomic DNA and analyzed HBV surface antigen (HBsAg) and DNA levels in the mouse serum over a seven-week period of time. Although mice injected with 20 µg HBV DNA produced initially a higher serum level of HBsAg, this antigen became undetectable after a week. In contrast, mice injected with 4 µg HBV DNA produced a lower-level of surface antigen that persisted for well over a month (Figure 6A). Importantly, this prolonged antigenemia was abolished if mice injected with 4 µg HBV DNA were also injected with anti-IFN-α/β antibodies on a weekly basis, indicating a role of IFN-α/β in maintaining antigenemia. In contrast, although anti-IFN-α/β antibodies slightly increased the level of HBsAg in mice injected with 20 µg HBV DNA one week after DNA injection, they did not prolong antigenemia, indicating the possible involvement of other factors in limiting viral persistence. Since HBsAg could be masked by the antibodies that it elicited, we also analyzed the serum HBV DNA levels. Mice injected with 20 µg HBV DNA produced a higher level of serum HBV DNA than mice injected with 4 µg HBV DNA within the first four days after injection. However, this difference was not observed one week after injection. Furthermore, the serum HBV DNA level of mice injected with 20 µg DNA declined rapidly thereafter and became undetectable after three weeks, whereas this serum DNA level persisted for a much longer period of time in mice injected with 4 µg DNA (Figure 6B). Similarly, if mice injected with 4 µg HBV DNA were also injected with anti-IFN- α/β antibodies, their serum viral DNA level also became undetectable after three weeks. The anti-IFN-α/β antibodies had little effect on the serum HBV DNA level in mice injected with 20 µg HBV DNA, except at the earliest time point. When the alanine aminotransferase (ALT) was analyzed to monitor liver injury, high levels of ALT were observed only within the first few days, most likely caused by the hydrodynamic injection, which causes liver injury (Figure S6). The ALT levels were low after that, indicating minimal liver injuries. These results indicated that the low-dose inoculation of the HBV DNA could lead to a more persistent viral replication and this persistence was dependent on IFN-α/β.
Interferons are thought to play an important role in the control of viral infections. Indeed, previous studies using transgenic mice that produced a high level of HBV indicated that interferons could suppress HBV replication [18]. Our observation that IFN-α/β could enhance HBV replication in mice that produced a low level of HBV is thus rather intriguing. Our results indicated that this enhancement was due to the activation of the HNF3γ gene and STAT3, which then stimulate the HBV ENI enhancer activity. The induction of HNF3γ by IFN-α/β requires the presence of a low level of HBV, as such induction was not observed in the presence of a high level of HBV DNA (Figure 4A) or in naïve mice (data not shown). STAT3 is activated in the presence of a low level of HBV, but it was not activated in the presence of a high level of HBV (Figure 5A). Since STAT3 can also be activated by IFN- α/β in hepatocytes in the absence of HBV [19], [20], it appears that HBV at a high replication level can prevent the activation of STAT3 by IFN-α/β. These observations indicate an interesting interplay between HBV and the interferon signaling pathway. A model of how IFN-α/β enhances HBV replication in illustrated in Figure 7. In this model, the binding of IFN-α/β to its receptor activates STAT3, likely due to the phosphorylation by JAK, which is associated with the IFN-α/β receptor and is activated upon binding of IFN-α/β to its receptor. In the mean time, the activated interferon signaling pathway also interacts with HBV to induce the expression of HNF3γ, which then binds cooperatively with STAT3 to the HBV ENI enhancer to stimulate HBV gene expression and viral replication. How HBV may interact with the JAK-STAT signaling pathway to induce the expression of HNF3γ is still not clear. This effect is independent of the HBx protein, since Tg31 mice carried the X-null HBV genome but yet HNF3γ could be induced by poly(I∶C) in this mouse line (Figure 4A). Clearly, if another HBV gene product such as that of the S, C or P gene is involved, this gene product must exert a dose-dependent effect since only a low replication level of HBV could induce HNF3γ.
The observation that IFN-α/β enhances HBV replication when the HBV DNA level is low may represent a mechanism by which HBV uses to establish its infection in patients, as the viral level is expected to be low in patients during the early stage of HBV infection. This possibility is supported by our observation that the injection of a small amount of HBV genomic DNA into mice could lead to prolonged viremia in an IFN-α/β-dependent manner (Figure 6). As our mouse model does not allow the reinfection of hepatocytes by HBV, it is conceivable that, if reinfection is possible, viral replication can persist in mice for an even longer period of time. Indeed, it has been shown that low-dose (1 or 10 genome-equivalent copies) inoculations of HBV into chimpanzees would lead to the spread of the virus and result in the infection of 100% of hepatocytes and prolonged immunopathology [21]. In contrast, the inoculation of between 104 and 108 genome copies of the virus led to a limited spread of the virus in the liver and the speedier clearance of the virus. Based on our findings described in this report, it is conceivable that the initial IFN-α/β response to the low-level HBV inoculation enhanced viral replication and spread and prolonged viral infection in chimpanzees.
Although the injection of IFN- α/β antibodies could suppress viral persistence in mice injected with 4 µg HBV DNA, the injection of IFN-α/β antibodies did not prolong viral persistence in mice injected with 20 µg HBV DNA. This result indicated that the removal of IFN-α/β alone was not sufficient to maintain viral persistence when viral load is high. The reason for this is unclear, but it may involve other factors such as additional cytokines (e.g., IFN-γ) that may be induced by high HBV load [18].
Recent studies indicated that the HBx protein could bind to MAVS (also known as IPS-1, VISA or Cardiff), which is an important adaptor molecular of the RIG-I signaling pathway, to suppress the induction of IFN-β [22], [23]. It has also been reported that the HBV DNA polymerase could bind to the DDX3 deadbox RNA helicase to suppress its interaction with TBK1/IKKε and the induction of IFN-β [24], [25]. In contrast to HBV structural proteins, HBx and the HBV polymerase are produced at a much lower level during viral replication. For this reason, it is likely that these two HBV products can only efficiently suppress the induction of type I interferons when the viral replication level is high. It is conceivable that the lag period before the effective concentrations of HBx and polymerase are reached in HBV-infected cells will allow HBV to use the interferon response to stimulate its gene expression and replication. However, once HBV has replicated to a high level and interferons become a negative regulator for HBV replication, HBx and the viral polymerase will exert their anti-interferon activities to enhance the survival of the virus.
IFN-α has been used to treat HBV patients, but it generated sustained response in only a minority of patients. Our observation that IFN-α/β could have opposite effects on HBV in a viral load-dependent manner indicates that viral load may be one of the reasons why this therapy has generated inconsistent responses in HBV patients.
Our studies on mice were conducted in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. Our animal protocol was approved by the Institutional Animal Care and Use Committee of the University of Southern California.
HBV transgenic mouse lines Tg05 and Tg08 have been previously described [5], [26]. These two mouse lines carried the 1.3mer, over-length wild-type HBV genome. Tg31 and Tg 38 also carried 1.3mer HBV genome, with the exception that the expression of X protein was abolished by the introduction of an A-to-C mutation at nt.1377 to remove the initiation codon of the X protein and a C-to-T mutation at nt.1398 to introduce a premature termination codon in the X coding sequence. All of the experiments were conducted using age-matched male mice with the B6 genetic background.
The plasmid p1.3×HBV, which contains the 1.3mer over-length HBV genome, has been described before [27]. pCMV-HNF3α, pCMV-HNF3β and pCMV-HNF3γ, which express HNF3α, HNF3β and HNF3γ, respectively, have also been described [15]. The expression plasmids for mouse HNF3γ shRNA, STAT3 shRNA and control shRNA were purchased from Sigma-Aldrich. The HBV ENI/X reporter constructs pGL-3-1115-luc, pGL-3-1136-luc, pGL-3-1167-luc and pGL-3-1203-luc, which contained nt.1115–1355, 1136–1355, 1167–1355 and 1203–1355, respectively, of the HBV genome, were generated by PCR amplification of the HBV DNA fragment for cloning into the pGL3-basic vector (Promega). The plasmid pGL-3-ENII-luc was constructed by insertion the ENII/core promoter complex (nt.1403–1803) into the pGL3-basic vector. pRL-SV40 (Promega), which expresses the renilla luciferase, was included in the transfection studies to serve as the internal control to monitor the transfection efficiency.
Age and HBeAg-matched male HBV transgenic mice or naïve mice were injected intravenously with 200 µl saline with or without poly(I∶C) (200 µg/mouse). Mice were sacrificed 24 hours later. The serum was collected and the liver was harvested and stored at −80°C for analyses.
Nine-week old male mice were injected via the tail vein with p1.3×HBV in 5–8 seconds in a volume of saline equivalent to 8% of the body weight of the mouse. In all the injection experiments, the vector DNA pUC19 was included if necessary to ensure that the total amount of DNA used for injection is identical among different mice. 24 hours after the hydrodynamic injection, the serum was collected and HBeAg was assayed by ELISA. Mice matched by body weight, age and HBeAg levels were used for injection with poly(I∶C).
Rabbit anti-mouse IFN-α (PBL, New Jersey) and hamster anti-mouse IFN-β (Biolegend, San Diego) antibodies were used in this study. Purified rabbit IgG (Cell Signaling Tech.) and hamster IgG (Abcam) were used as the control antibodies. Rabbit anti-HNF3α (Abcam), anti-HNF3β (Cell Signaling Tech.), anti-HNF3γ (Sigma–Aldrich Co.), anti-STAT3 (Cell Signaling Tech.), anti-phosph-STAT3(Tyr705) (Cell Signaling Tech.) and anti-lamin-β (Abcam) antibodies were used for Western-blot.
Liver tissues were homogenized in DNA lysis buffer (20 mM Tris-HCl, pH 7.0, 20 mM EDTA, 50 mM NaCl, 0.5% SDS), incubated for 16 hours at 37°C with proteinase K (600 µg/ml) and then phenol/chloroform extracted for the isolation of DNA. The HBV RI DNA in the core particles was isolated using our previous protocol [27]. For RNA isolation, liver tissues were homogenized in Trizol (Invitrogen) and total RNA was isolated following the manufacturer's protocol. Both Southern and Northern blot analyses were conducted using the 32P-labeled HBV DNA probe. For Western-blot analysis, liver tissues were homogenized in the RIPA solution (10 mM Tris-HCl, pH 7.0, 150 mM NaCl, 1% Triton X-100, 1% sodium deoxycholate, and 0.1% sodium dodecyl sulfate) and, after a brief centrifugation to remove cell debris, the protein concentrations were determined by Bradford BCA (Biorad) and the Western blot was performed using our previous procedures [28].
10 µl mouse serum was added into 100 µl lysis buffer (20 mM Tris-HCl, 20 mM EDTA, 50 mM NaCl, and 0.5%SDS) containing 27 µg proteinase K. After incubation at 65°C overnight, viral DNA was isolated by phenol/chloroform extraction and ethanol precipitation. The DNA pellet was rinsed with 70% ethanol and resuspended in 10 µl TE (10 mM Tris-HCl [pH 7.0], 1 mM EDTA). For hydrodynamic injection studies, 10 µl serums was digested with 10 µg DNase I and micrococcal nuclease for 30 min at 37°C to remove free DNA. HBV DNA was then isolated as described above. For HBV real-time PCR analysis, the following primers were used: forward primer, 1552-CCGTCTGTGCCTTCTCATCTG-1572; and reverse primer, 1667-AGTCCTCTTATGTAAGACCTT-1646. The TaqMan probe used was 1578-CCGTGTGCACTTCGCTTCACCTCTGC-1603. The assays were performed as described [29].
The reporter constructs were delivered into the mouse liver by hydrodynamic injection. Nine-week old male mice were used in the studies. Twenty-four hours after injection, the serum was collected and HBeAg was assayed by ELISA. Mice with matched body weight and the HBeAg level were injected intravenously with 200 µl saline with or without poly(I∶C) (200 µg/mouse) at 42 hours after hydrodynamic injection and sacrificed 6 hours later. The mouse liver was isolated and stored at −80°C. The firefly luciferase and the renila luciferase activities were measured using the dual luciferase assay kit (Promega). The firefly luciferase activities were normalized against the renilla luciferase activities, which served as the internal control. All of the experiments were repeated at least three times.
Serum ALT levels were measured using the ALT kit (Cayman Chemical Company, USA). HBsAg and HBeAg were measured using their respective ELISA kit (International Immuno-Diagnostics, CA). All of these assays were conducted following the manufacturers' instructions.
|
10.1371/journal.pntd.0002143 | Structure-Activity Relationship Studies on the Macrolide Exotoxin Mycolactone of Mycobacterium ulcerans | Mycolactones are a family of polyketide-derived macrolide exotoxins produced by Mycobacterium ulcerans, the causative agent of the chronic necrotizing skin disease Buruli ulcer. The toxin is synthesized by polyketide synthases encoded by the virulence plasmid pMUM. The apoptotic, necrotic and immunosuppressive properties of mycolactones play a central role in the pathogenesis of M. ulcerans.
We have synthesized and tested a series of mycolactone derivatives to conduct structure-activity relationship studies. Flow cytometry, fluorescence microscopy and Alamar Blue-based metabolic assays were used to assess activities of mycolactones on the murine L929 fibroblast cell line. Modifications of the C-linked upper side chain (comprising C12–C20) caused less pronounced changes in cytotoxicity than modifications in the lower C5-O-linked polyunsaturated acyl side chain. A derivative with a truncated lower side chain was unique in having strong inhibitory effects on fibroblast metabolism and cell proliferation at non-cytotoxic concentrations. We also tested whether mycolactones have antimicrobial activity and found no activity against representatives of Gram-positive (Streptococcus pneumoniae) or Gram-negative bacteria (Neisseria meningitis and Escherichia coli), the fungus Saccharomyces cerevisae or the amoeba Dictyostelium discoideum.
Highly defined synthetic compounds allowed to unambiguously compare biological activities of mycolactones expressed by different M. ulcerans lineages and may help identifying target structures and triggering pathways.
| Buruli ulcer is a chronic necrotizing skin disease caused by Mycobacterium ulcerans. The characteristic histopathological features of Buruli ulcer, severe destruction of subcutaneous tissue with minimal inflammation in the core of the lesion, are primarily attributed to the cytotoxic activity of mycolactone, the macrolide exotoxin of M. ulcerans. Different geographical lineages of M. ulcerans produce different structural variants of mycolactone. By using highly defined synthetic mycolactones, including both naturally occurring molecular species and additional non-natural variants, we have assessed the influence of the structure of the C-linked upper side chain and the lower C5-O-linked polyunsaturated acyl side chain on biological activity. Changes in the lower side chain affected the cytotoxic activity against mammalian cells more profoundly than changes in the upper side chain. Mycolactone A/B had no antimicrobial activity against Gram-positive and Gram-negative bacteria and was also inactive against Saccharomyces and Dictyostelium.
| The macrolide exotoxin mycolactone is a key virulence factor of M. ulcerans and plays a central role in the pathogenesis of Buruli ulcer [1]. Mycolactones have been shown to act both in vivo and in vitro on various mammalian cell types, including fibroblasts [1]–[6], adipocytes [7], keratinocytes [8], myocytes [9], [10], macrophages [2], [6], [11]–[14], dendritic cells [15] and T-cells [16]–[18]. Effects caused by mycolactones include induction of apoptosis/necrosis, cytoskeletal rearrangements, impaired cytokine production and interference with cellular signaling.
The polyketide synthases required for mycolactone biosynthesis are encoded on the extrachromosomal plasmid pMUM [19], [20]. M. ulcerans has evolved from a common M. marinum ancestor by acquisition of this plasmid and has subsequently diverged into two principal mycolactone-producing lineages [21], [22]. The “classical” lineage includes M. ulcerans isolates associated with Buruli ulcer from Africa and Australia. The “ancestral” lineage includes both Buruli ulcer isolates from Japan, China and Mexico and isolates from fish and frogs previously also designated M. pseudoshottsii, M. liflandii or M. marinum [23], [24]. Mycolactones are composed of a 12-membered macrolide core and two attached side chains; a short upper, C-linked side chain (comprising C12–C20) and a longer lower, C5-O-linked polyunsaturated acyl side chain. While the macrolide core structure and upper side chain are conserved, mycolactone populations from different M. ulcerans sub-lineages vary in the length, the number and localization of hydroxyl groups and in the number of double bonds of the lower side chain. M. ulcerans strains may produce several molecular variants of mycolactone, with one or more species dominating [25]. The mycolactone repertoire seems to be highly conserved within a defined geographical sub-lineage of M. ulcerans [25]. Mycolactone A/B is produced by strains of the classical M. ulcerans lineage found in Africa and is regarded as the most potent toxin. Australian classical lineage strains produce - in addition to mycolactone A/B - mycolactone C, which lacks one hydroxyl group. Mycolactone D with an additional methyl group is produced by Chinese strains belonging to the ancestral lineage. M. ulcerans ancestral lineage isolates from fish and frogs have been found to produce the mycolactone variants E and F.
Mycolactones have previously been prepared from M. ulcerans cultures by a two-step extraction procedure, yielding preparations of acetone soluble lipids predominantly containing mycolactone. These extracts can be further purified by chromatographic methods [26]; nevertheless, the use of extracted mycolactones for comparative studies may be compromised by the heterogeneity of preparations. Therefore, biological studies with highly defined synthetic mycolactones represent an attractive alternative. Based on the established synthesis of the mycolactone core [27]–[30], different synthesis strategies have been pursued for the stereoselective partial and total synthesis of mycolactones [31]–[33]. In addition, simplified C8-desmethyl-mycolactone analogues have been synthesized, which were analyzed for their cytopathic potency by using cell rounding as a parameter to compare cytotoxic activities [34]. No systematic structure-activity relationship studies on larger sets of synthetic mycolactones have been published so far. Results of individual studies cannot be reliably compared, since different readouts, such as cell rounding [22], [34], [35], cytokine production [36] or flow cytometric parameters [35], [36] have been employed. Furthermore, different cell lines (such as Jurkat T-cells [36], murine fibroblasts [22], [34], [35] and sets of human tumor cell lines [37] have been used and cytotoxic activity has been assessed after different times, such as 24 hours [36] or 24 and 48 hours [35]. Most of these structure-activity studies have been limited to mycolactone A/B and a limited number of derivatives. A comparison of the activity of eight C8-desmethyl mycolactone analogues is hampered by the fact that lack of the C8-methyl substituent reduces the cytopathic activity by a factor of 125 [34]. Here, we have performed more systematic comparative studies using synthetically produced natural toxins and additional structural mycolactone variants that are not found in nature.
We recently reported the synthesis of mycolactone A/B [38]. Details of the syntheses of mycolactones C and F and of the six non-natural mycolactone derivatives will be published elsewhere (Gersbach et al., manuscript in preparation). Briefly, all mycolactones discussed here were prepared by the same overall strategy that we had previously developed for the synthesis of mycolactone A/B. Thus, a modified Suzuki coupling was employed to establish the C12–C13 bond and elaborate the full upper side chain and a Yamaguchi type acylation reaction was used to attach the lower side chain. All final products used for biological testing were purified by RP-HPLC; they were generally obtained as mixtures of (interconverting) double bond isomers. Analytical data for the synthetic compounds are provided as supplementary information (S1).
For biological testing, 0.5 mg/ml stock solutions of the mycolactones were prepared in cell culture grade DMSO (Sigma). Stock solutions were aliquoted and stored frozen at −20°C.
Murine L929 fibroblasts were grown in RPMI medium (Gibco) supplemented with 10% FCS (Sigma), 2 mM glutamine (Gibco) and 0.05 mM β-mercaptoethanol (Gibco) and incubated at 37°C and 5% CO2. Cells were passaged 3–6 times prior to use in cytotoxicity experiments. For flow cytometry analysis 24,000 cells were seeded into 24-well plates (Falcon) and allowed to adhere o/n. Medium was then aspirated and replaced by 500 µl medium containing different concentrations of mycolactone and 0.12% DMSO (vol/vol). After incubation for 24, 48 or 72 h, cells were detached from the culture plates by repeated gentle flushing through a pipette tip without use of Trypsin-EDTA. Harvested cells were centrifuged for 10 min at 1,200xg, resuspended in 300 µl binding buffer with 0.2 µg/ml Annexin V-FITC (AnnexinV kit, Calbiochem) and incubated for 30 min at 4°C. The cells were spun again and pellets were resuspended in 300 µl staining buffer containing 0.3 µg/ml propidium iodide (AnnexinV kit, Calbiochem). Cell suspensions were analyzed by flow cytometry using a BD FACS Calibur Flow Cytometer (Becton Dickinson) and apoptotic (A+/PI−) and necrotic (A+/PI+) cell populations were determined using the CellQuest Pro Software (Becton Dickinson). The experiments were set up in triplicates and performed at least twice. Mycolactone A/B in a concentration range of 3.75 to 120 ng/ml was included as control in all experiments. The mycolactone concentration at which 50% of the cells were killed (LC50) was determined by plotting the percentage of affected cells (sum of A+/PI− and A+/PI+ cells) against the log concentration of the individual mycolactones. In Fig. 1–3 only data for concentrations close to the LC50 are shown.
To measure proliferation of L929 fibroblasts, 24,000 cells were seeded into 24-well plates (Falcon) and allowed to adhere o/n. The medium was aspirated and replaced by 500 µl medium containing 60 ng/ml of mycolactone and 0.06% DMSO (vol/vol). At time point 0 and after 24, 48 and 72 h fibroblasts were harvested, diluted 1∶100 in isotonous solution and measured using an automated cell counting device (Casy®TT, Schärfe System). The experiment was set up in triplicates and performed twice.
24,000 cells were seeded on four-chamber glass slides (BD Falcon) with complete RPMI medium and allowed to adhere for 24 hours. The medium was aspirated and replaced by 500 µl medium containing different concentrations of mycolactone. After the specified incubation period, cells were washed once in PBS and fixed with 4% formaldehyde (Medite) for 20 min. Fibroblasts were washed again in PBS prior to permeabilization in Triton X-100 (0.1% in PBS) for 20 min. Cells were rinsed in PBS and blocked by incubation in 4% FBS in PBS for additional 20 min. The actin cytoskeleton was stained by incubating the cells for 1 h at room temperature with Texas Red-X phalloidin (3 units/ml, in blocking solution, Molecular Probes). Cells were washed in blocking buffer, then in PBS. ProLong Gold antifade reagent (Invitrogen) containing diamidino-2-phenylindole (DAPI) was used for nuclear counterstaining. Cover slips were mounted onto the slides and cell rounding as well as the staining of nuclei and actin cytoskeleton was qualitatively analyzed by fluorescence microscopy (Leica DM 5000 B). Mycolactone-induced changes in the pattern and intensity of the Texas Red-X phalloidin staining of the cytosolic actin cytoskeleton as well as in the uniform, round and clear-edged DAPI staining of nuclei in healthy cells were observed.
Metabolic activity of mycolactone-treated L929 fibroblasts was analyzed by performing Alamar Blue assays. Seeding of cells and addition of mycolactones were performed as described for the flow cytometry-based cytotoxicity assays. After mycolactone treatment, alamarBlue® reagent (Invitrogen) was added to the wells (10% v/v) and the cells were further incubated for 1 hour at 37°C and 5% CO2. Fluorescence intensities were measured using a SpectraMax Gemini XS (Molecular Devices) and the values were calculated referring to the DMSO control (0 ng/ml mycolactone) set at 100%. The experiments were set up in triplicates and performed at least twice. The concentration at which the metabolic activity of cells was inhibited by half (IC50) was determined by plotting the fluorescence intensity against the log concentration of the individual mycolactones.
Antimicrobial activity of mycolactone on Streptococcus pneumoniae (SP1, P1577), E. coli (DE(3)) and Saccharomyces cerevisiae was tested by applying the disk agar diffusion (Kirby-Bauer) method. Bacteria were pelleted, resuspended in PBS and spread on blood agar/LB agar. The plates were dried for 30 min and sterile paper disks were distributed circle-like onto the agar. Mycolactone A/B solutions of different concentrations (0.003 µg/ml to 10 µg/ml) were applied on the paper disks (40 µl). The agar plates were incubated o/n at 37°C and then analyzed for potential zones of inhibition.
The effect of mycolactone A/B on the growth of Dictyostelium discoideum DH1-10 was assessed by performing an Alamar Blue assay in a 24-well format. 2,000 cells were seeded in 500 µl medium containing mycolactone in the concentration range of 0.16 to 500 ng/ml. As controls, DMSO and blasticidin were used. After an incubation period of 3 days at room temperature, alamarBlue® reagent (Invitrogen) was added and the plate was further incubated for 18 hours at room temperature.
Recently we described a novel strategy for the synthesis of mycolactone A/B that is based on the stereoselective construction of the macrolactone core by ring-closing olefin metathesis and subsequent incorporation of the C- and O-linked side chains by suitable fragment couplings [38]. Taking this synthesis approach a set of natural mycolactones (mycolactone A/B, mycolactone C, mycolactone F) and additional derivatives displaying modifications in the lower or upper side chain (PG-119, PG-120, PG-155, PG-157, PG-165 and PG-182) were produced (see Figures 1, 2 and 3) for biological testing. The biological activity of these synthetic compounds on the murine L929 fibroblast cell line was assessed by flow cytometry. After treatment with different concentrations of synthetic mycolactones, cells were stained both with FITC-labeled annexin-V and with propidium iodide. Annexin-V binds to exposed phosphatidylserine residues translocated from the inner to the outer leaflet of the plasma membrane in cells undergoing apoptosis. Propidium iodide intercalates into the DNA of cells that have lost nuclear membrane integrity, serving as a marker for necrosis. Quadrant analysis was performed to determine apoptotic (A+/PI−) and necrotic (A+/PI+) cell populations.
While first signs of mycolactone A/B-mediated cell death were already detectable after 24 hours, significant effects were only observed after 48 hours [38]. For comparison with mycolactone A/B, the lethal concentration of mycolactone analogues at which 50% of the cells were affected (LC50) was therefore determined after 48 hours (Table 1). As described previously [38], mycolactone A/B was highly potent (Figure 1) with a LC50 of 12 nM (Table 1). For the two naturally occurring structural variants mycolactone F and mycolactone C, the LC50 values were 29 nM and 186 nM, respectively (Table 1). Mycolactone C differs from mycolactone A/B in lacking the hydroxyl group at position C12 of the lower side chain. Mycolactone F has a shorter side chain with also only two hydroxyl substituents (Figure 1). While these natural mycolactones retained cytotoxic activity, compound PG-155, a non-natural structural variant devoid of all hydroxyl groups in the lower side chain, showed only minor activity with a LC50 of 4550 nM (Table 1).
Apart from these mycolactone variants with modifications in the lower side chain, also analogues with modifications in the upper side chain were synthesized and tested (Figure 2). Introduction of a hydroxyl group at C20 in compound PG-165 had no major effect, since PG-165 had only a slightly higher LC50 (15 nM) than mycolactone A/B (Figure 2, Table 1). In addition, derivatisation of this hydroxyl group into an acetate (PG-157 with a LC50 of 45 nM) or into a bulky butyl carbamate (PG-182 with a LC50 of 50 nM) reduced cytotoxicity only about three-fold (Figure 2, Table 1). Thus, the upper side chain turned out to be relatively tolerant to a significant extension in length and to the presence of polar linker elements between the natural side chain and the extension module.
PG-120, a derivative with a significantly truncated lower side chain, showed some residual cytotoxic activity (LC50 = 3426 nM), whereas PG-119, a derivative with an acetyl residue as the lower side chain, showed no activity within the concentration range tested (Figure 3, Table 1). For all compounds, except PG-120, concentrations required for cytotoxic activity (as measured by flow cytometry), reduction of metabolic activity in an Alamar Blue-based assay, changes in the intensity and pattern of phalloidin-staining of the actin cytoskeleton and changes in the round, clear-edged and uniformly stained nuclear morphology of normal cells were in the same range. While the IC50 value for PG-120 (171 nM; Table 1), was twenty-fold lower than the LC50, the LC50/IC50 ratios of all other compounds with widely varying toxic potency ranged between 1.5 and 3.2 (Table 1). Furthermore, at such sub-lethal PG-120 concentrations a marked reduction in cell proliferation (Figure 4), and a transient effect on the actin cytoskeleton accompanied by the rounding up of the cells, without changes in nuclear morphology was observed (Figure 5A). A similar activity was not observed for PG-119 (Figures 4 and 5).
When analyzed for antimicrobial activity, mycolactone A/B was found inactive against all microbial species tested, including Gram-positive (Streptococcus pneumoniae) and Gram-negative (Neisseria meningitis, Escherichia coli) bacteria; it was also inactive against yeast (Saccharomyces cerevisae) and amoeba (Dictyostelium discoideum).
Our flow cytometric analyses of murine fibroblast L929 cells treated with a series of synthetic mycolactones reconfirmed that changes in the O-linked lower side chain can profoundly affect the biological activity. Activity of the synthetic mycolactone A/B was in the range reported for mycolactone preparations extracted from M. ulcerans cultures [1]. Mycolactone F was about two times less active and mycolactone C about 15 times less active than mycolactone A/B, respectively. For extracted mycolactone C an even far more pronounced difference in activity compared to mycolactone A/B has been described in assays determining L929 fibroblast rounding at 24 h and loss of monolayer at 48 h [25]. In addition to mycolactone C, Australian M. ulcerans strains also produce mycolactone A/B. Our data indicate that this mycolactone A/B portion may be more important for the pathogenesis caused by these strains than mycolactone C. In accordance with our findings, only a slightly lower activity was observed, when extracted mycolactone F was compared to mycolactone A/B in a L929 cell apoptosis assay at 24 h [3]. When the inhibition of IL2 production by activated Jurkat T-cells instead of cell death was used as readout, both mycolactones F and C were dramatically less potent than mycolactone A/B [36].
While we have investigated different types of modifications for the lower and upper side chains, it is clear that both the incorporation of polar substituents at C20 and the extension of the upper side chain by up to 7 heavy atoms, in contrast to most of the modifications of the lower side chain, does not lead to a substantial loss in cytotoxicity. It remains to be seen how the removal of hydroxyl groups from the upper side chain or its overall shortening would affect potency.
It has been proposed that mycolactones enter mammalian cells via passive diffusion and interact with cytosolic target(s) [39]. Reduced or abolished activity of structural variants of mycolactone may thus be related to lack of binding to target structure(s), inefficient triggering of activation pathways or reduced translocation across the cell membrane. Studies using isotopically labeled rather than fluorescence labeled structures with altered biophysical properties are required to gain better insight into mechanisms that allow mycolactones to cross biological membranes.
Our findings with the truncated mycolactone PG-120 shows that different biological effects of mycolactone can be dissociated by using structural variants. In line with these observations, sub-lethal doses of mycolactone A/B have been shown to alter trafficking and cytokine production of lymphocytes and macrophages [40], [41]. It remains to be elucidated whether different pathways and target structures are involved in the triggering of the biological effects of mycolactone.
Since a number of macrolides have antibiotic activity against a broad spectrum of bacteria it has been speculated that mycolactone secreted by M. ulcerans during active disease may prevent superinfection of BU wounds. However, synthetic mycolactone A/B showed no antimicrobial activity against any of the microorganisms tested here. In line with this observation, superinfection of Buruli ulcer lesions seems to be much more common than traditionally anticipated (Yeboah-Manu et al., personal communication).
Much has still to be learnt about the biophysical properties of mycolactones, their distribution and stability in biological systems, their target structures and triggering pathways in mammalian cells. Synthetic natural mycolactones, isotopically labeled derivatives and structural variants represent valuable tools to address these open questions in future.
|
10.1371/journal.pcbi.1004261 | Metabolic Needs and Capabilities of Toxoplasma gondii through Combined Computational and Experimental Analysis | Toxoplasma gondii is a human pathogen prevalent worldwide that poses a challenging and unmet need for novel treatment of toxoplasmosis. Using a semi-automated reconstruction algorithm, we reconstructed a genome-scale metabolic model, ToxoNet1. The reconstruction process and flux-balance analysis of the model offer a systematic overview of the metabolic capabilities of this parasite. Using ToxoNet1 we have identified significant gaps in the current knowledge of Toxoplasma metabolic pathways and have clarified its minimal nutritional requirements for replication. By probing the model via metabolic tasks, we have further defined sets of alternative precursors necessary for parasite growth. Within a human host cell environment, ToxoNet1 predicts a minimal set of 53 enzyme-coding genes and 76 reactions to be essential for parasite replication. Double-gene-essentiality analysis identified 20 pairs of genes for which simultaneous deletion is deleterious. To validate several predictions of ToxoNet1 we have performed experimental analyses of cytosolic acetyl-CoA biosynthesis. ATP-citrate lyase and acetyl-CoA synthase were localised and their corresponding genes disrupted, establishing that each of these enzymes is dispensable for the growth of T. gondii, however together they make a synthetic lethal pair.
| Understanding the metabolism of disease-causing microorganisms can guide drug design through the identification of metabolic enzymes whose activity is indispensable for important cellular functions. Such understanding can come from the reconstruction and computational analysis of metabolic networks. In this study we have focused on the metabolism of an opportunistic human pathogen, Toxoplasma gondii, which chronically infects about 30% of humans worldwide. Through a semi-automatic reconstruction we have developed ToxoNet1, a comprehensive metabolic model of Toxoplasma gondii, and we have performed an extensive computational analysis to explore the properties of Toxoplasma metabolism. In particular, we have identified and classified the minimal set of substrates the parasite utilizes for growth, along with the genes and pairs of genes that are essential for cellular functions such as growth and energy metabolism. We have validated our computational predictions that the genes encoding acetyl-CoA synthase and ATP-citrate lyase enzymes serve complementary function but their simultaneous disruption does not allow cell growth. This study presents a number of hypotheses generated using ToxoNet1, which can lead to the discovery of novel antiparasitic drug targets.
| The phylum of Apicomplexa comprises a large number of obligate intracellular parasites that can infect organisms across the whole animal kingdom. An important member of this phylum, Toxoplasma gondii, is a ubiquitous opportunistic pathogen responsible for one of the most common parasitic infections in humans and warm-blooded animals. It is estimated that up to 30% of the human population is chronically infected [1]. Toxoplasmosis is largely asymptomatic in healthy adults but can cause severe disease or even death in immunocompromised individuals and can lead to complications in development of the foetus, if primary infection occurs during pregnancy [2].
T. gondii possesses a complex life cycle, which is composed of an asexual replicative stage in the intermediate host and a sexually replicative stage within the definitive feline host. During the asexual phase, T. gondii can switch from a fast-replicative tachyzoite form, which causes acute disease, to a slow-growing bradyzoite stage, which forms cysts that are characteristic of chronic infection. The encysted, slow growing form is resistant to commonly used drugs and immune system attack. Few efficient medicines are available to treat toxoplasmosis and they mainly treat the acute phase of the disease. Furthermore, poor tolerance of these drugs promotes the search for novel drug targets.
Unlike other notorious apicomplexan parasites that infect a narrow range of host cell types, such as the Plasmodium species, T. gondii is able to invade and asexually replicate within virtually any nucleated cell of warm-blooded animals. The broad range of cells amenable to infection by T. gondii reflects the plasticity of the parasite’s metabolism and versatility in accessing and utilising nutrients to support its intracellular growth [3,4]. The complexity of decoupling the metabolic processes of this intracellular pathogen from those of the infected host limits the depth of our understanding about the metabolic capabilities of T. gondii.
Currently, no adequate experimental approaches exist to answer comprehensively the following important questions: 1) what substrates are available within the host cell that are necessary for T. gondii replication and which of these are dispensable; 2) in which intracellular compartments do the enzymatic activities annotated at the genome level occur; 3) which of these enzymatic activities are indispensable for replication or other vital processes of the parasite. While achievable, the application of high-throughput gene knockout or knockdown strategies to globally determine gene essentiality in T. gondii remains a major undertaking.
Computational (i.e. in silico) metabolic modelling coupled with systematic analyses facilitates the study of biological systems. This modern approach of systems biology has been extensively exerted to predict gene essentiality in various bacteria, including numerous pathogenic species [5]. In silico metabolic models offer a cost-effective pipeline to identify putatively indispensable metabolic processes that, in the case of pathogens, represent potential targets for therapeutic intervention [6]. Recently, models have been constructed for eukaryotic pathogens including for members of the phylum Apicomplexa [7–10].
In this study we have reconstructed a genome-scale metabolic model of T. gondii, ToxoNet1, aiming to address the abovementioned questions in a systematic way and to the extent possible with the currently available knowledge. Using flux-balance analysis we have identified genes, reactions and pairs of genes for which deletion renders production of biomass components impossible. Furthermore, we have assessed which precursors are necessary for each of the biomass components and have defined minimal sets of such precursors that allow in silico simulation of growth. To illustrate the applicability of ToxoNet1 in filling knowledge gaps regarding parasite metabolism, we experimentally challenged the model prediction regarding the alternative routes for generation of cytosolic acetyl-CoA. We have confirmed in vitro the functional redundancy and synthetic lethality of the two biosynthetic routes that involve the ATP-citrate lyase and acetyl-CoA synthase enzymes.
Here we present the full genome-scale in silico reconstruction of metabolism in T. gondii with manually refined gene-reaction associations—ToxoNet1. The model reconstruction process required completion of the following major steps (schematically shown in Fig 1): (1) reconstruction of the draft metabolic network; (2) compartmentalization of the intracellular space; (3) verification of the metabolic capabilities and manual literature-based corrections; (4) representation of the plausible exchanges of metabolites between the infected host cell and the parasite. In general terms, the reconstruction process was consistent with the workflow that has been previously used for semi-automated reconstruction of a genome-scale metabolic model for Penicillium chrysogenum by means of the RAVEN Toolbox [11]. All the necessary manual corrections were made in accordance to the conventional model reconstruction protocol [12].
For the compartmentalization process we consulted the models of Plasmodium falciparum [9,13] and used sequence-based predictors of subcellular localization [14–16] as well as the ApiLoc (http://apiloc.biochem.unimelb.edu.au/apiloc/apiloc) database. The databases used for identification of “gap” reactions were KEGG [17] and LLAMP [18]. Recon 2 [19] was used as a model of host cell metabolism to define the list of putatively host-supplied substrates.
Reconstruction of the metabolic network was achieved by combination of the state-of-the-art algorithm for a semi-automated generation of metabolic networks [11] and a comprehensive manual curation based on the relevant primary literature. Details of the reconstruction process are provided in the materials and methods section and the most important steps are discussed below.
The reconstructed metabolic network accounted for 527 open-reading frames (ORFs) that were linked to 867 unique metabolic reactions present in the KEGG [17] database. Each functional annotation in the model was assigned with two estimates (namely a bit-score and e-value), which indicated confidence of association for a given ORF with a corresponding enzymatic function.
In ToxoNet1, the majority of reactions were associated with genes that encode corresponding metabolic enzymes (Fig 2A). Intracellular metabolic reactions not associated with any genes comprised only 9.6% of all the metabolic reactions present in the model (6.7% of all the reactions in the model); they include spontaneous reactions (not enzyme-catalysed) and so-called gap-filling reactions that were added for correct functioning of the model. Most of the metabolite transport reactions, which were added to connect pathways segregated between subcellular compartments, also lacked known gene associations.
Among the metabolic enzymes encoded in the genome of T. gondii and present within ToxoNet1, the transferases (class 2) and oxidoreductases (class 1) were the most numerous (over 60% of all the enzymes). Hydrolases (class 3), lyases (class 4) and ligases were significantly less represented (Fig 2B). The least frequent were isomerases (class 5), which is in accordance with the models of other pathogens, such as Leishmania major [20] and Plasmodium falciparum [8].
Relatively straightforward experimental methods exist to determine the localization of proteins in T. gondii [21], however, it would be laborious and expensive to apply such methods for hundreds of enzymes. According to the ApiLoc database (http://apiloc.biochem.unimelb.edu.au/apiloc/apiloc), experimental data was available only for a limited number of proteins: 60 out of the 527 included in ToxoNet1. Thus the only reasonable option for building a compartmentalized metabolic model was to make use of sequence-based localization predictors. To define putative localizations of the enzymes with no experimental evidence we generated sequence-based predictions using three software algorithms: TargetP [16], MitoProt II [15], and ApicoAP [14]. We then manually reconciled output data of the three independent predicting algorithms and assigned the localization based on the manually determined consensus prediction, also considering recent primary literature whenever it was available (S1 Table).
Comparison of these predictions with 60 experimentally established subcellular localizations available from ApiLoc uncovered two issues: (i) not all the computational predictions matched their experimental data (highlighted in the S1 Table); (ii) some of the enzymes were reported to be present in two and, in one case (glutathione/thioredoxin peroxidase [22]), three different subcellular compartments defined in ToxoNet1. In these cases the sequence-based predictors were not efficient and suggested only one of the compartments. Therefore, we assigned compartments in a supervised manner considering all available evidence.
The information on compartmentalisation organised according to the global metabolic subsystems of ToxoNet1 is shown in the Table 1.
In ToxoNet1 a majority of the reactions (c.a. 69%) occur in the cytosolic compartment. We also considered this compartment as a default for enzymes with a subcellular localization remaining unclear from the in silico predictions. The mitochondrion accommodates 19% of all metabolic reactions of the model. The remaining 12% are localized to the apicoplast, a plastid-like non-photosynthetic organelle [23]. We connected these two compartments to the cytosol by 223 transport reactions. This enabled the corresponding number of metabolites to be transported across these boundaries, which delineate the organellar membranes (for further details see the materials and methods section). We allowed the metabolites present in the mitochondrion or apicoplast to be reversibly transported to and from the cytosol if they satisfied the following requirements: (i) metabolite has to be present (i.e. participate in at least one reaction) in both the cytosol and corresponding non-cytosolic compartment; (ii) it should be neither phosphorylated nor containing an acyl-carrier-protein or CoA moiety, unless supporting evidence of such transport is available. We also allowed: (i) import of phosphoenolpyruvate and dihydroxyacetone phosphate from the cytosol to the apicoplast [24]; (ii) export of ATP from the mitochondrion to the cytosol (TGME49_249900); (iii) export of isopentenyl pyrophosphate (IPP) and dimethylallyl pyrophosphate (DMAPP) from the apicoplast to the cytosol. The latter allowed us to observe an unusual feature of the isoprenoid biosynthesis pathway in T. gondii regarding the isopentenyl pyrophosphate isomerase (EC 5.3.3.2). This enzyme interconverts IPP into its more reactive isomer—DMAPP. In T. gondii, as well as in Plasmodium spp. this enzyme is absent [25]. Thus, both IPP and DMAPP have to be produced and transported from the apicoplast as separate entities for synthesis of long isoprenoid chains in the cytosol.
A major limitation of all algorithms for automatic reconstruction of metabolic models is the number of missing reactions (so-called gaps) that disrupt metabolic pathways. The RAVEN Toolbox offers a functionality called metabolic tasks (checkTasks and fitTask functions), which verifies that specific tasks are fulfilled and, if necessary, it fills the gaps in the metabolic pathways to meet the required functionality.
We created metabolic tasks for the synthesis of every metabolite included in the biomass reaction (complete formulation of all the tasks is provided in the S2 Table). The outcomes of the tasks were categorized in three groups shown in Table 2. The first group contains the metabolites that could be produced from glucose and inorganic compounds (i.e. de novo synthesis was possible in the model). The second group contains the metabolites whose synthesis required a precursor with a specific moiety (e.g. hypoxanthine or other source of purine moiety for purine nucleotides and their derivatives). The third group are the biomass constituents that could not be produced even with the maximum set of host-supplied substrates, and therefore they were directly taken up from the host (e.g. choline, arachidonic acid, cholesterol).
According to ToxoNet1, T. gondii has the capability to produce de novo a number of biomass precursors such as pyrimidine nucleotides and their derivatives, fatty acids, isoprenoids and about half of the proteinogenic amino acids (Table 2). When tasks of de novo production for certain biomass constituents failed we attempted to produce them by supplying additional precursors. In some cases uptake of one or more molecule(s) from the list of host-supplied molecules, enabled production of the necessary biomass precursors in ToxoNet1. Nevertheless, several biomass building blocks could not be produced in the model even when all the 230 host-supplied substrates were provided simultaneously, indicating that the parasite is likely auxotrophic for these molecules.
In the cases when metabolic tasks could not be accomplished despite literature evidence (mainly acquired from the LLAMP database [18]), we performed gap-filling. This is a conventional part of building genome-scale metabolic models [12] and it implies inclusion of a minimal number of reactions to complete pathways of interest. We used both the automated gap-filling function of the RAVEN Toolbox (fillGaps) and manual gap-filling based on the information from the LLAMP database [18]. For instance, the threonine biosynthesis pathway in ToxoNet1 was nearly complete but lacked a single enzyme, the homoserine dehydrogenase (E.C. 1.1.1.3). The corresponding reaction (R01773) was included in the model without a gene assigned (i.e. as a “gap-fill”) justified by the presence of four other enzymes in the pathway and the fact that the immediate downstream enzyme was identified in a proteomics study [26]. ToxoNet1 also retrieved a known issue of the lysine biosynthesis pathway: the bacterial-type pathway consisting of 9 enzymatic steps lacks enzymes annotated for 4 sequential steps, making the presence of the functional pathway debatable. While this issue remains unresolved, we decided not to gap-fill this pathway in the model, thus, leaving lysine an essential amino acid. The full list of gap-fill reactions included in ToxoNet1 is provided in S3 Table.
Importantly, some of the information necessary for a metabolic model could not be easily deduced from the genome sequences alone. This is why we merged the draft output metabolic network from the RAVEN Toolbox with a small-scale model of central carbon metabolism in T. gondii. We manually built such a model prior to this study based on the genome annotation from ToxoDB and refined it according to the relevant primary literature. This model contained a number of recent experimental findings as well as manually assembled complex gene-reaction associations. Among these were the pyruvate dehydrogenase activity that can be carried out by the branched-chain keto-acid dehydrogenase complex (BCKDH) in the mitochondrion [27], mitochondrial pyruvate transporter [28–30] and the GABA-shunt connected to the tricarboxylic acid (TCA) cycle [3].
One of the major challenges in simulation of parasite metabolism in silico is the unknown range of metabolites available for the parasite within the host cell. As extracellular replication (axenic cultivation) of T. gondii is not possible, it remains undefined which of the substrates are dispensable for the parasite and which are not.
To address this issue in silico, we developed a method to enumerate all of the smallest/ minimal metabolite sets (further referred to as in silico minimal medium or IMM), which could enable simulation of growth in ToxoNet1. In brief, the algorithm applied iterative rounds of biomass production using the fewest number of substrates possible. After each of the iterations we added a constraint to ensure that the next IMM included at least one substrate uncommon from the preceding ones (further details on the formulation of the algorithm are described in the materials and methods section).
The results of these simulations showed that as few as 19 substrates were sufficient for simulation of T. gondii replication in ToxoNet1. We also observed a relatively large number (2592) of alternative sets of 19 substrates with 10 metabolites being constitutively present in all the IMMs. The other 9 substrates could be substituted by at least one other host-supplied metabolite (Table 3). Despite the very large theoretical number of combinations (in the case of 9 variable substrates picked from 220) we observed only 2592 alternative IMM of 19 substrates. This indicates the ability of ToxoNet1 to substitute one substrate with another from the set of 230 is rather limited. Indeed, the majority of the 2592 IMMs arise from flexibility in the carbon source (one out of 7 available) and a source of nicotinate moiety (one out of 3). The other metabolites are either non-substitutable (10) or substitutable with one single alternative. However, in this particular analysis, we excluded the possibility of substituting one metabolite with simultaneous uptakes of several others, as it would lead to more than the minimal number of substrates used.
We simulated in silico the outcomes of systematic removal of genes and reactions in ToxoNet1 to explore which of them represent indispensable metabolic functions. With the full set of 230 host-supplied substrates out of the 527 genes in ToxoNet1, 53 genes were predicted to be essential (Table 4 and literature evidence [31–38]). Considering that most of the transport reactions do not have known gene-reaction associations, we also simulated single reaction deletions. This allowed us to assess the dispensability of metabolite transport across the compartments, as well as metabolite exchanges between T. gondii and its host cell (Table 4 and S5 Table).
In addition to single gene and reaction essentiality, we also simulated double gene deletions to reveal the pairs in which genes are not essential for replication on their own, yet deleterious when disrupted together. A total of 20 pairs that caused such synergistic effect (synthetic lethality) are listed in S5 Table.
Gene essentiality predictions depend on the following important aspects of metabolism as represented in the model: (1) range of substrates that can enter the model (i.e. molecules that the parasite can take up from the infected host cell); (2) composition of the parasite cell represented as a set of biomass precursor molecules (the concept of biomass objective functions is explained in [39]); (3) the presence of alternative metabolic routes to produce biomass building blocks from the different substrates. Assuming a very permissive range of 230 substrates as potentially accessible for the parasite we predicted a minimal set of essential genes and reactions. In consequence, the number of genes predicted by ToxoNet1 as non-essential is likely to be overestimated. Yet with this assumption the probability of incorrect prediction of genes to be essential is lower.
In order to evaluate whether ToxoNet1 predictions closely reflect the metabolic capabilities of T. gondii observed experimentally in tissue culture, we chose to assess the importance of the two independent routes of cytosolic acetyl-CoA production in the rapidly dividing tachyzoite stage (Fig 3A). Acetyl-CoA is an important molecule in central carbon metabolism that is involved in many biochemical processes such as fatty acid synthesis (FAS type I pathway), fatty acid chain elongation and acetylation of proteins, in particular histones. ToxoNet1 identified two enzymes that can produce acetyl-CoA in the cytosol: (1) from acetate through the acetyl-CoA synthase (ACS) reaction (TGME49_266640), (2) from TCA-derived citrate by ATP-citrate lyase (ACL) (TGME49_223840), as shown in Fig 3A. Acetoacetate-CoA ligase (TGME49_219230) produces acetoacetyl-CoA that can be converted to acetyl-CoA by the acetyl-CoA acetyltransferase (TGME49_301120), which belongs to the fatty acid degradation pathway in mitochondrion. We thus concluded that this is unlikely to be producing cytosolic acetyl-CoA production due to the predicted mitochondrial localization of the two enzymes. In further support of the importance of cytosolic acetyl-CoA production is the presence of a putative ortholog of the human acetyl-CoA transporter (AT1) in the T. gondii genome [40]. This transporter was shown to localize at the endoplasmic reticulum (ER) membrane and to be essential for the survival of eukaryotic cells by allowing import of acetyl-CoA from the cytosol to the ER to acetylate proteins within this compartment. ToxoNet1 predicted ACS- and ACL-encoding genes to be fully dispensable when knocked out individually, however, their simultaneous knockout was predicted to be lethal (S5 Table).
To determine the localization and level of expression of ACS and ACL in T. gondii, we modified the endogenous locus (knock-in) by introducing a 3xTy-epitope tag at the C-terminal end of both genes in the RHku80ko (Ku80ko) background strain, which limits random integration in the genome hence facilitating recovery of homologous recombination events (Fig 3B). ACS is clearly cytosolic and nuclear whereas ACL appears to localize mostly to the cytosol, whilst a fainter nuclear staining can also be detected by indirect immunofluorescence assay (IFA) (Fig 3C). Localization of the acetyl-CoA transporter AT1 in the perinuclear region of the ER in T. gondii was validated using the same knock-in tagging strategy (Fig 3B and 3C). Expression of the epitope-tagged proteins was further validated by Western blot analyses, where ACS runs at the expected molecular weight of ~80 kDa, ACL at ~140 kDa and AT1 at ~65 KDa (Fig 3D). Comparative signal intensity observed on Western blots suggests that ACS is significantly more abundant than ACL and AT1.
To functionally assess the importance of both routes to produce cytosolic acetyl-CoA and challenge the predictions made by ToxoNet1, individual deletion of the genes encoding ACL and ACS were achieved using a double homologous recombination strategy (ACLko and ACSko respectively) in Ku80ko (Fig 3E). For both genes, transgenic parasites were readily obtained and cloned. Absence of the ACL and ACS ORFs and their replacement by a selection cassette in individual clones was validated by genomic PCR (Fig 3E), thus, supporting the prediction made by the model that these genes are both dispensable for the survival of T. gondii tachyzoites. No significant defect could be observed in the overall lytic cycle of these knockout mutants as represented by plaque assays performed in human foreskin fibroblast (HFF) monolayers. Indeed, both ACLko and ACSko parasites formed lysis plaques of similar sizes when compared to wild type (Ku80ko) parasites (Fig 3F). Moreover, after 24h of intracellular growth, most vacuoles of the ACLko strain contained 4 to 8 parasites, which is comparable to the number of parasites per vacuole for cells infected with Ku80ko or ACSko (Fig 3G). Finally, growth competition assays between mutants and wt parasites, which would detect mild loss of fitness, showed no significant defect either.
In order to assess whether a double knockout of ACS and ACL is lethal for T. gondii as suggested by ToxoNet1, we first attempted to disrupt ACL by single homologous recombination in the middle of the ORF in the ACSko parasite background (same strategy as presented in Fig 3B but leading to a truncation of the protein and removal of the catalytic site). While we were able to interrupt the ACL gene in Ku80ko parasites, we failed to generate such a mutant in the ACSko background. This result strongly suggested that ACS and ACL together are critical for the biosynthesis of cytosolic acetyl-CoA. To further confirm the synthetic lethality between ACS and ACL we generated a conditional knockdown of ACL in Ku80ko and in ACSko by U1 snRNP-mediated gene silencing with Cre-recombinase dependent positioning of U1, as has been recently developed in T. gondii [41] (generating ACL-lox and ACSko/ACL-lox respectively; Fig 4A). Following Cre-mediated recombination, the endogenous 3’untranslated region is excised and a U1 recognition site is placed adjacent to the termination codon. Consequently the ACL pre-mRNAs are cleaved at the 3’-end and degraded, leading to a highly efficient knockdown of the gene. Loss at the protein level can be assessed by immuno-detection of the C-terminal 3Ty-tag (Fig 4A). Correct integration of the construct was confirmed by genomic PCR (Fig 4B). Importantly, upon deletion of ACS the level of endogenous ACL protein was significantly increased compared to ACL-lox (Fig 4C). This change in level of ACL in the absence of ACS was reproducibly confirmed by generation of a second independent transgenic parasite line where the ACS gene was disrupted in the ACL-lox background strain (Fig 4D).
To conditionally disrupt ACL in the ACSko, the ACL-lox and ACSko/ACL-lox parasites were transfected with a plasmid transiently expressing the Cre recombinase. While ACL-lox excised parasites could be readily propagated in culture the ACSko/ACL-lox excised parasites were lost after the first passage as monitored by genomic PCR analyses of the two excised parasite populations. While genomic recombination of excised parasites could be readily detected in the original transfected parasites (P0), the signal was lost immediately after the first passage of these cultures (P1 and P2) (Fig 5A) indicating a rapid deleterious effect. Thirty hours post transfection of Cre recombinase, the loss of ACL-3Ty-tagged protein was evident by IFA in about 50% of the ACL-lox and ACSko/ACL-lox vacuoles. While ACL-lox parasites lacking ACL could be propagated in culture, parasites lacking ACL in the ACSko/ACL-lox strain were immediately lost in the first culture passage (Fig 5B). While ACL-lox excised parasites appeared normal, most vacuoles from excised ACSko/ACL-lox parasites exhibited a severe morphological defect and impairment in the parasite division process with a loss of pellicle integrity as seen by perturbation of GAP45 staining (Fig 5C). Furthermore these parasites appear to continue dividing their nuclei, mitochondrion and apicoplast, but fail to form daughter cells, resulting in organelle accumulation in the vacuolar space (Fig 5C).
To determine whether the deleterious phenotype and loss of pellicle integrity could be attributed to a block in type I fatty acid synthesis following depletion in cytosolic acetyl-CoA, the gene coding for T. gondii FASI was disrupted in wild type RH strain parasites by CRISPR/Cas9 mediated genome editing [42]. Trangenic parasites were obtained following double-stranded breaks generated by Cas9 at a position downstream of the FASI ATG. Two independent clones were sequenced to confirm the introduction of frame shift mutations (S1A Fig) No defect in the lytic cycle could be observed in these parasites (S1B Fig), which is in accordance with the ToxoNet1 prediction that FASI activity is not essential for parasite survival. Taken together, these data firmly support the predictions made by the model regarding the individual dispensability of ACS and ACL and the synthetic lethality of ACL in the absence of ACS.
ToxoNet1 constitutes a full genome-scale reconstruction of metabolism within T. gondii, which can simulate growth of the parasite in silico and infer essentiality of its genes. It considerably extends the scope of previous work [10] and contributes to better understanding of the limits in the metabolic capabilities of this opportunistic human and animal pathogen. While the modelling efforts of Song et al were aiming to reveal strain-type specific differences in metabolism of T. gondii, in the present study we reconstructed an independent model that represents the potential scope of the metabolic capabilities of the parasite independently of the life-stage and strain-type. Aiming at the most reliable gene essentiality predictions we used many alternative assumptions on the range of accessible substrates and transportability of metabolites. We have also achieved more comprehensive coverage of all the metabolic capabilities of the parasite and implemented functional gene-protein-reaction associations enabling rigorous gene deletions studies. Hereafter we provide a more in depth discussion on the key aspects of our approach.
There are two distinct approaches for model reconstruction process that are commonly referred to as top-down and bottom-up [43]. The top-down approach is usually applied when detailed experimental data about the majority of individual system components is scarce. Conversely, for the bottom-up approach, a significant body of relevant primary literature is a necessary prerequisite. These two approaches are very much complementary and this is why increasingly more studies combine them in order to achieve the best results. This was also the case in the ToxoNet1 reconstruction efforts: the first draft model was a hybrid of the output generated by RAVEN Toolbox and a significantly smaller manually reconstructed (and highly curated) model, which was built upon the similar model of P. falciparum [44] used as a template. This bottom-up small-scale model allowed capturing of the features that were not identified by functional annotations, such as formation of multi-enzyme complexes for certain enzymatic activities (e.g. the mitochondrial BCKDH complex that carries out the pyruvate dehydrogenase function [27]) as well as cofactor specificities for the enzymes.
Genes with metabolic functions in ToxoNet1 represent a modest fraction (c.a. 9%) of the T. gondii protein-coding genes, which is comparable but greater than in the genome-scale models of P. falciparum (7%) [8], L. major (6.7%) [20] and Cryptosporidium hominis (5.5%) [7]. ToxoNet1 provides a broader and more complete coverage of the metabolic capabilities of the parasite with 145 additional enzyme-coding genes (38% greater) compared to the earlier metabolic model of T. gondii [10]. Comparison of the number of metabolic reactions with the genome-scale models of the malaria parasite P. falciparum [8,9,13] confirms the common view that T. gondii possesses broader metabolic capabilities (Table 5). These differences in metabolic capabilities could be partially responsible for the observations that T. gondii can infect a very broad range of cell types for asexual replication, while Plasmodium spp. can only replicate within specific tissues.
ToxoNet1 contains confidence estimates (e-values and bit-scores) for each gene annotated as encoding certain metabolic activities. This has allowed us to suggest a number of metabolic functions for genes that did not have annotations with E.C. numbers in the ToxoDB database (S4 Table). For example, TGME49_237140 (an “ethylene-inducible protein” in ToxoDB without an E.C. identifier) was annotated as pyridoxal 5'-phosphate synthase pdxS subunit (E.C. 4.3.3.6) with very high confidence estimates (e-value 5.30x10-141 and bit-score 479). Accordingly, an orthologous gene (PF3D7_0621200) is annotated as pyridoxine biosynthesis protein (PDX1) in P. falciparum, yet without an E.C. number assigned. There was also a set of genes with low confidence in their function, despite their annotation in ToxoDB (e.g. TGME49_305840 had a low sequence identity to known nicotinate-nucleotide adenylyltransferases with an e-value of 4.70x10-15 and a bit-score of 60.6). These two examples demonstrate the potential of the RAVEN Toolbox to improve genome annotation and produce models with confidence estimates that enable evaluation and, potentially, future corrections of the existing models.
ToxoNet1 contains 260 unique dead-end metabolites, which currently can be only produced or only consumed in the network. We kept them in the model in order to allow future developments and expansion of the scope of the model. Further definition of the metabolic routes can utilise these dead-end metabolites and contribute to a more complete understanding of the metabolic capabilities of the parasite. For instance, the tRNAs loaded with the respective amino acids were included to enable future extension of the model to the representation of protein synthesis.
We chose to impose relaxed constraints in terms of subcellular compartments due to rather high uncertainty in subcellular localisation of the enzymes as well as largely unknown capabilities of the parasite to transport metabolites across its organellar membranes. We assigned the enzymes with the corresponding reactions to their putative compartments and allowed a broad range of metabolites to be transported across the compartment boundaries. This approach to compartmentalisation ensures minimal bias in our results due to underestimated transport capabilities or incorrect assignment of enzymes to subcellular compartments. We expect subsequent refinements of ToxoNet1 with more stringent compartmentalisation as more reliable, high-throughput computational and experimental methods are becoming available for subcellular localisation of enzymes and functional annotation of transporter proteins in T. gondii. Our objective of this work is to provide ToxoNet1 as a resource for the community and therefore we avoided imposing constraints and hypotheses, which are not well-tested and confirmed that could contaminate the model and our results.
To understand the nutritional requirements of T. gondii, we implemented an algorithm that identifies and ranks the minimal number of substrates required for growth in ToxoNet1. The results indicate that uptake of as few as 19 of 230 substrates allows the model to produce all biomass building blocks. Moreover, we identified 2592 sets of 19 substrates that can allow growth, and 10 substrates common to all these sets (Table 3). These 10 substrates are compounds that are not synthesized de novo by T. gondii and are the precursors for biomass building blocks and essential cofactors. Interestingly, the uptake of S-adenosyl-homocysteine provided precursors for multiple biomass building blocks simultaneously (purine nucleotides, threonine, methionine and their derivatives). Among the alternative substrates we identified 7 carbon sources: five hexoses and two pentoses (ribose and deoxyribose), which can be incorporated into the metabolism by a ribokinase enzyme. Uptake of pentoses through a hexose transporter has been reported in the protozoan parasite L. major and may potentially represent an additional level of versatility in meeting the need of a carbon source, similarly to the previously observed in T. gondii utilization of glutamine [45,46]. We could envision such need as the parasite uses various host cells, where the set of available substrates may vary between cell types.
Validation of genome-scale models using metabolic tasks is an approach developed to evaluate metabolic capabilities of the models in a systematic manner. To date it has been applied to several models [11,47,48] for testing whether the major metabolic functions can be fulfilled and the biochemical pathways that support these functions were represented correctly. ToxoNet1 is the first metabolic model of an apicomplexan parasite where the reconstruction process involved the validation of metabolic tasks. Using this approach we explored the capability of this pathogen to produce de novo biomass building blocks from glucose and inorganic substances. Interestingly, almost all the cofactors of metabolic enzymes, with the exception of pyridoxal phosphate cannot be produced de novo. Their synthesis required precursors that contained certain chemical moieties (Table 2). This suggests that virtually all the metabolic functions of the parasite depend on an adequate supply of specific precursors by the infected host cell. Furthermore, we found that T. gondii requires uptake of almost all the amino acids, which are essential for the host, with the possible exception of threonine (discussed below). Notably, T. gondii lacks the pathway for arginine synthesis, which can be produced by human cells, however arginine is growth limiting for human cells because its de novo synthesis is insufficient and thus supplemental uptake is required [49]. This suggests that during infection and growth of the parasite, the biosynthetic capabilities of the host can be significantly compromised due to competition for essential and growth limiting amino acids and vitamins.
Threonine is the only amino acid essential for human cells and which T. gondii can potentially produce. However, this pathway includes the enzyme homoserine dehydrogenase (E.C. 1.1.1.3), the encoding gene for which is not identified to date in the genome of the parasite. Data from ToxoDB suggests that the genes encoding the enzymes of the threonine biosynthesis pathway, namely aspartokinase (TGME49_227090), threonine synthase (TGME49_220840), homoserine kinase (TGME49_216640) and aspartate-semialdehyde dehydrogenase (TGME49_205420), are expressed at low levels in the tachyzoite stage. Homoserine kinase has been detected in the proteome of oocysts [50], the infective forms of the parasite that can survive for extensive periods of time outside the host cell [51]. Taken together, these observations indicate that the whole pathway from aspartate to threonine might not be functional in certain life stages and therefore T. gondii could be a conditional threonine auxotroph. Should threonine biosynthesis also be active in the other life stages it could represent a selective drug target for inhibition of parasite replication without affecting the mammalian host cells. Maintenance of some metabolite pools in T. gondii may be a result of both uptake and de novo production. Hence, if the complete set of genes of a metabolic pathway is present in the genome of the parasite, this pathway is not necessarily active and will not meet the demands of the parasite regarding the products of the pathway. Fox et al. showed that salvage of pyrimidine nucleotides from the host cell can support in vitro but not in vivo survival of T. gondii mutants with disrupted de novo pyrimidine synthesis pathways [52]. Therefore, accurate consideration of metabolism in the host and pathogen together is central for a better understanding of the metabolic needs and capabilities of T. gondii.
Our study represents the first model-based gene essentiality predictions for T. gondii. The earlier efforts to infer essentiality of genes in this parasite were reported by Gautam et al. [53]. Their approach for producing the list of putatively essential genes was largely dissimilar to ours and relied on conservation of the enzymes across parasitic and free-living species with further pruning based on the literature data. We believe that our modelling approach, which takes into account many more inputs and constraints, produces more reliable gene essentiality predictions, and it follows the standard modelling procedures established by the community [12,54]. Nevertheless, we consider that the part of the workflow proposed by Gautam et al. can be complementary to our study, which also includes larger number genes with the updated functional annotation we performed here. This future study will prioritise the putatively essential genes according to their apparent utility as drug targets with minimal risk of off-target effects on the host cell metabolism.
Using ToxoNet1 we predicted a minimal set of 53 enzyme-encoding genes to be indispensable for parasite replication within human cells. The majority of these genes (49 out of 53) have orthologs in the malaria parasite P. falciparum and 32 of them were predicted as essential in the existing metabolic models [38]. Evidence of essentiality for metabolic enzymes in T. gondii is currently scarce, thus, we could not test the majority of our essentiality predictions (Table 5). However, new technologies such as CRISPR/Cas9-mediated genome editing, hold promise of forthcoming high-throughput gene knockout studies in various organisms including T. gondii [42]. The dataset of experimentally established gene essentiality will provide an important validation of ToxoNet1 and a prerequisite for its further refinement and expansion.
We predicted synthetic essentiality of 20 pairs of genes, which represent two distinct cases. In the first case the pair of genes encode two isoenzymes that catalyse the same reaction in the same compartment. For example, the genes TGME49_318580 and TGME49_285980 encode isoenzymes, which catalyse the phosphoglucomutase reaction (E.C. 5.4.2.2) and both were experimentally localised to the cytosol [33]. This reaction produces glucose-1-phophate in the cytosol, which is an indispensable precursor for starch and nucleotide-sugar metabolism. Therefore, simultaneous deletion of these two genes is a synthetic lethal. In the second case of synthetic essentiality each gene encodes for an enzyme of a different reaction but with a common product, which is necessary for biomass synthesis. Thus, the two different enzymes can substitute for each other (sometimes indirectly) for production of an essential metabolite. For example, we established experimentally the synthetic lethality between the genes encoding acetyl-CoA synthase (ACS) and ATP-citrate lyase (ACL), both of which produce acetyl-CoA in the cytosol of the parasite. This central metabolite participates notably in fatty acid synthesis, fatty acid chain elongation and in acetylation of proteins. We were able to rule out experimentally that the severe phenotype observed upon depletion of cytosolic acetyl-CoA was due to an impact on FASI since parasites lacking FASI did not exhibit any significant impairment in the lytic cycle.
Moreover given that deletion of the genes coding for elongases previously reported [55] did not phenocopy the lethality observed when attempting to delete ACS and ACL simultaneously, we suspect that blockage of protein acylation might be responsible for the severe consequences of deletion of cytosolic acetyl-CoA. In this context, we experimentally confirmed that the T. gondii acetyl-CoA transporter (AT1) is localized in the membrane of the ER and is anticipated to deliver acetyl-CoA to the secretory pathway. It would be interesting to examine the importance AT1 for parasite survival. In contrast, the mitochondrion and the apicoplast should not be affected since these compartments have their own routes for production of acetyl-CoA: (i) in the apicoplast it is produced by the pyruvate dehydrogenase (PDH) complex [56], and (ii) in the mitochondrion synthesis of acetyl-CoA is carried out by the BCKDH complex [27]. The predicted synthetic lethality between ACS and ACL suggests that these are the only two significant sources of acetyl-CoA in the cytosol, which was not assessed by previous experimental studies and is formally demonstrated here.
Furthermore this synthetic lethality suggests that there is no, or no significant, transport of acetyl-CoA to the cytosol from the mitochondrion or the apicoplast. This further confirms that, unlike acetate, acetoacetate could not be a source of cytosolic acetyl-CoA and therefore the metabolic role of acetoacetyl-CoA, considered to be a dead-end metabolite in ToxoNet1, requires further investigation. Unexpectedly, the deletion of the gene coding for ACS leads to a reproducible and immediate increase in abundance of ACL protein, a possible compensatory adaptation that has not been reported previously. It is unclear whether the capacity of the parasite to respond to the absence of ACS results from an increase at the level of transcription or protein stability but this intriguing mechanism of adaptation deserves further investigation. ToxoNet1 can assist in comprehensively embracing the various routes that T. gondii employs to produce acetyl-CoA in different subcellular compartments.
Beyond this study ToxoNet1 can be used as a global metabolic context for integration and interpretation of various high-throughput experimental data, similarly to the studies made in P. falciparum and other eukaryotic pathogens [38]. Of particular interest is the integration of experimental data collected on different life stages of T. gondii that will allow the model to yield context-specific predictions and, potentially, reveal valid drug targets as well as fundamental knowledge regarding the stages implicated in persistence and transmission of this important human and animal pathogen.
The reconstruction process started with generation of a draft metabolic network of T. gondii based on the annotation of its protein sequences as extracted from the ToxoDB database [57]. Within the framework of the RAVEN Toolbox these protein sequences were compared to the hidden Markov models (HMM) generated for each KEGG [17] orthology group [11]. In the cases when e-values for the matches between a T. gondii protein and an HMM were smaller than the specified e-value cutoff (10–20), the enzymatic reactions associated with the corresponding KEGG orthology ID were added to the draft metabolic network. As a result we have obtained a set of metabolic reactions linked to genetic loci of T. gondii. At this stage the model did not contain information about either subcellular compartments, or about transport of metabolites. We next merged this model with a manually curated, small-scale metabolic reconstruction that we previously built based on the P. falciparum model [44].
Similarly to Huthmatcher et al [9]., we removed reactions with generic metabolites (such as “protein”, “dNTP”) and replaced non-unique metabolite identifiers with unique ones when they corresponded to the biochemically equivalent entities (e.g. (S)-Lactate and L-lactate were replaced with L-lactate).
The parasitophorous vacuole (PV) space, which secludes T. gondii from the host cell cytosol, is the outermost compartment represented in ToxoNet1. Extracellular compartment of the model corresponds to the PV space. A eukaryotic organization of intracellular space of the parasite also includes a number of compartments, among which we chose to represent in ToxoNet1 the most relevant to metabolism, namely: mitochondrion, apicoplast and cytosol. The following sequence-based localization predictors were used to suggest putative subcellular localization of the enzymes: TargetP [16] (version 1.1), MitoProt II [15] (version 1.101) and ApicoAP [14] (version 1)). As an input data we used sequences of ORFs extracted from ToxoDB (v.9, strain ME49) for the genes included in ToxoNet1. Computational localization predictions were reconciled with the literature evidence for T. gondii and P. falciparum in the cases when the latter were available (see S1 Table).
Transport reactions were included to link the majority of the metabolites in non-cytosolic intracellular compartments with their cytosolic counterparts; such transports were not created for phosphorylated metabolites or those that contained [acyl-carrier protein] ([ACP]) or Coenzyme A (CoA) moiety attached (with the exception of the apicoplast-to-cytosol transport of IPP, and DMAPP as well as the mitochondrion-to-cytosol transport of ATP).
Testing metabolic tasks is a built-in functionality of the RAVEN Toolbox [11] meant for verification of correct metabolic capabilities in genome-scale models. Formulation of this function allows one to test production of a certain metabolite(s) (or flux through particular reactions) with strictly specified inlet and outlet of metabolites. For instance, a metabolic task “de novo synthesis of ATP” was formulated as following: given unlimited uptake of glucose, oxygen, inorganic phosphate, sulfate and ammonium, can the model produce ATP. The model failed to accomplish this task, which is consistent with the literature knowledge about auxotrophy of T. gondii for purines [58]. Thus, the task could be passed only when we added to the list of available compounds hypoxanthine, adenine or another molecule containing a purine moiety. In a similar manner we have tested over a 60 metabolic tasks (all in S2 Table) to ensure realistic metabolic capabilities of the model.
We simulated metabolism of the host cell using the most recent tissue-unspecific model of human metabolism [19]. The Recon2 model was modified in terms of available substrates to reflect growth on the defined minimal medium (Dulbecco's Modified Eagle's Medium with glutamine and glucose). The PV which secludes the parasite from the host cell cytosol had been reported as being permeable to small-molecule metabolites with molecular weights below 1500 Da [59] and thus imposes no relevant constraint to the metabolites we considered in ToxoNet1. Thus, all the molecules that could be produced from the medium components in the cytosol of the host cell were assumed to be potentially accessible for the parasite provided that they satisfy the following criteria: host-supplied substrates were only small molecules (below 1.5 kDa) that were not phosphorylated or bound to-CoA,-[ACP] or carnitine. We also assumed that the parasite could potentially dispose of a wide range of metabolic by-products from its cytosol into the host cell. Thus, we added sink reactions for the same 230 metabolites that were assumed to be host-supplied. As an exception we also allowed direct uptake into the cytosol of the following five generic metabolites: “reduced acceptor”, “sulfur donor”, “acyl-CoA”, “carboxylate” and “1-acylglycerol”; two non-generic metabolites: selenite and myo-inositol, and the appearance of apoprotein in the mitochondrion and the apicoplast.
In order to simulate replication of the parasite and assess the essentiality of genes, reactions and substrates, we assumed maximisation of flux through the biomass reaction to be the objective function of ToxoNet1. It represents cellular replication as a reaction that consumes pre-defined amounts of metabolites defined as small-molecule biomass precursors as well as energy in the form of ATP. We used the biomass reaction from the previous study [10] as a template and introduced the following modifications: we extended this biomass reaction with the following cofactors—NAD, NADP, FAD and lipoylated protein necessary for pyruvate dehydrogenase (PDH) activity in the mitochondrion and the apicoplast; we changed the compartment for lipid precursors from endoplasmic reticulum (in Song et al. [10]) to the cytosol (no ER compartment in ToxoNet1), and for L-lysine from mitochondrial to cytosol. The presence of the mitochondrial pathway for L-lysine production remained obscure, thus in ToxoNet1 it is acquired from the host instead of gap-filled de novo production in the parasite and sequestered towards biomass from the cytosol.
FBA is a standard computational approach for exploration of the metabolic capabilities represented in constraint-based models; principles, computational implementation of FBA as well as the key assumptions are extensively described elsewhere [60].
In the absence of clear knowledge on the scope of substrates that the parasite can take up from the host we made the following assumption: all the metabolites present in the host cell cytosol are potentially accessible for the parasite except those that are phosphorylated, bound to coenzyme A, acyl-carrier protein or carnitine. A list of molecules that satisfy these criteria was generated using the recent tissue-unspecific model of human metabolism Recon2 [19] (the list of the substrates is in S6 Table).
Minimal sets of metabolites necessary for the production of biomass were explored using an in-house developed mixed-integer linear programming algorithm. For each of the exchange reactions that allowed uptake of a substrate into the model (i.e. 230 host-supplied metabolites) we added one binary variable that denoted its utilization. Our algorithm solved the model subject to minimization of the sum of the binary variables thus yielding a minimal set of uptakes that allowed the model to simulate growth. After each iteration the algorithm added one new constraint to the model to assure that the following set would include at least one different uptake reaction compared to all the previously generated ones. The iterations were repeated until no more alternative sets of the same length could be found.
We performed simulation of gene deletion using a conventional approach [61] that implies evaluation of gene-reaction associations that include the gene of interest, preventing flux through corresponding metabolic reactions and an attempt to achieve a doubling time of 4.5 hours (specific growth rate of 0.15 h-1). Similar approaches were used for double gene deletion studies—simultaneously blocking reactions associated with all pairs of genes that were not predicted as singularly essential. In reaction deletion simulations we blocked flux through every single reaction in ToxoNet1 one at a time. Subsequent attempts at achieving the wild-type doubling time in the absence of the reaction indicated whether the gene was dispensable or not.
T. gondii tachyzoites (RHku80-ko (Ku80ko), RHku80-ko/ACS-ko (ACSko), RHku80-ko/ACL-ko (ACLko), RHku80-ko/ACL3Ty-LoxP3’UTRLoxP-U1 (ACL-lox), RHku80-ko/ACS-ko/ACL3Ty-LoxP3’UTRLoxP-U1 (ACSko/ACL-lox), RH/FASIko (FASIko)) were grown in confluent Human Foreskin Fibroblasts (HFF) and maintained in Dulbecco’s Modified Eagle’s Medium (DMEM, Life technology, Invitrogen) supplemented with 5% foetal calf serum, 2 mM glutamine and 25 μg/ml gentamicin in a humidified incubator at 37°C with 5% CO2.
All amplifications were performed with LA Taq (TaKaRa) polymerase and primers used are listed in S7 Table.
Knock-in of ACS: (pKI-ACS-3Ty) The genomic fragment of ACS (TGME49_266640) was amplified using primers ACS-1 and ACS-2 prior to digestion with with KpnI and SbfI and subsequent cloning in the same sites of pTub8MIC13-3Ty-HXGPRT [62] to introduce 3 Ty-tags at the C-terminus of the endogenous locus. Before transfection pKI-ACS-3Ty was linearized with SnaBI.
Knock-in of AT1: (pKI-AT1-3Ty) The genomic fragment of AT-1 (TGME49_215940) was amplified using primers AT1-1 and AT1-2 prior to digestion with KpnI and NsiI and subsequent cloning in the same sites of pTub8MIC13-3Ty-HXGPRT [62] to introduce 3 Ty-tags at the C-terminus of the endogenous locus. Before transfection pKI-AT1-3Ty was linearized with HindIII.
Knock-in and knock-down of ACL: (pKI-ACL-3Ty and pKI-ACL-3Ty-LoxP-3’UTR-LoxP-U1) Genomic fragment of ACL (TGME49_223840) was amplified using primers ACL-1/ACL-2 and subsequently digested with KpnI and NsiI prior to cloning in the same sites of the pTub8MIC13-3Ty-HXGPRT for 3Ty-tag knock-in [62], or the modified C-terminal destabilization vector pG152-3Ty-LoxP-3’UTRSag1-HXGPRT-LoxP-U1 [41]. Prior to transfection both plasmids were linearized with XhoI.
Knockout of ACS: (pTub-CAT-ACS-ko) around 1.5kb of the 5’ and 3’ flanking regions of ACS were amplified using primers ACS-3/ACS-4 and ACS-5/ACS-6 respectively. The 5’ flanking region was then cloned between KpnI and XhoI restriction sites of the pTub5-CAT and the 3’ flanking region between the BamHI and NotI sites. The plasmid was cut with KpnI and NotI prior to transfection.
Knockout of ACL: (pTub-HXGPRT-ACL-ko) around 1.5kb of the 5’ and 3’ flanking regions of ACL were amplified using primers ACL-3/ACL-4 and ACL-5/ACL-6 respectively. The 5’ flanking region was then cloned between the KpnI and XhoI sites of the p2855-HXGPRT plasmid and the 3’ flanking region between the BamHI and NotI sites. The plasmid was cut with KpnI and NotI restriction enzymes prior to transfection.
Frameshift knockout of FASI using CRISPR/CAS9 plasmid [42]: This vector has been generated using the Q5 site-directed mutagenesis kit (New England Biolabs) with the vector pSAG1::CAS9-U6::sgUPRT as template (a gift from Dr. L.D. Sibley). The UPRT-targeting gRNA was replaced by the FASI (TGME49_294820) specific gRNA using the primer pairs gRNA-FASI/gRNA-rev (gRNA highlighted in red in S7 Table).
Parasite transfections were performed by electroporation as previously described [63].The hxgprt gene was used as a positive selectable marker in the presence of mycophenolic acid (25 μg/mL) and xanthine (50 μg/mL) for pKI-ACS-3Ty, pKI-AT1-3Ty, pKI-ACL-3Ty, pKI-ACL-3Ty-LoxP-3’UTR-LoxP-U1 and pTub-HXGPRT-ACL-ko vectors transfected in Ku80ko or ACSko, as previously described [64].
Ku80ko was transfected with pTub-CAT-ACS-ko and 20μM chloramphenicol was used to select resistant parasites.
Resistant parasites were cloned by limiting dilution in 96 well plates and clones were assessed by genomic PCR.
To efficiently disrupt the FASI locus, 30 ug of the FASI gRNA-specific CRISPR/CAS9 vector was transfected into wild type RH parasites. 48 hours after transfection, GFP positive parasites were sorted by flow cytometry and cloned into 96-well plates using a Moflo Astrios (Beckman Coulter). Individual clones were then analysed by sequencing.
Genomic DNA was prepared from tachyzoites using the Wizard SV genomic DNA purification system (Promega). Correct integration of the different constructs into the genome of the various strains was determined by genomic PCR using GoTaq Green Master Mix (Promega).
The antibodies used in this study were described previously as follows: polyclonal rabbit anti-GAP45, rabbit anti-TgProfilin [65], monoclonal mouse anti-Ty (BB2), mouse monoclonal anti-F1-ATPase beta subunit (P. Bradley, unpublished) (5F4), mouse monoclonal anti-ATrx1 11G8 [66]. For Western blot analyses, secondary peroxidase conjugated goat anti-rabbit or mouse antibodies (Molecular Probes) were used. For immunofluorescence analyses, the secondary antibodies Alexa Fluor 488 and Alexa Fluor 594-conjugated goat anti-mouse or rabbit antibodies (Molecular Probes) were used.
Parasite-infected HFF cells seeded on cover slips were fixed with 4% paraformaldehyde/0.05% glutaraldehyde (PFA/Glu) in PBS. Fixed cells were then processed as previously described [67]. Confocal images were generated with a Zeiss (LSM700, objective apochromat 63x/1.4 oil) laser scanning confocal microscope at the Bioimaging core facility of the Faculty of Medicine, University of Geneva. Stacks of sections were processed with ImageJ and projected using the maximum projection tool.
Parasites were lysed in PBS-1% Triton X-100 and mixed with SDS–PAGE loading buffer under reducing conditions. The suspension was subjected to sonication on ice. SDS-PAGE was performed using standard methods. Separated proteins were transferred to nitrocellulose membranes and probed with appropriate antibodies in 5% non-fat milk in PBS-0.05% Tween20. Bound secondary peroxidase conjugated antibodies were visualized using the SuperSignal (Pierce).
Plaque assays: HFF monolayers were infected with parasites and let to develop for 7 days before fixation with PFA/Glu and Crystal Violet staining to visualize plaques.
Intracellular growth assays: HFFs were inoculated with parasites, washed 4h post infection and coverslips were fixed at 24 h post-infection with 4% PFA/Glu and stained by IFA with rabbit anti-TgGAP45. Number of parasites per vacuole was counted in triplicate for each condition (n = 3). More than 200 vacuoles were counted per replicate.
Flux-balance analysis was performed using MATLAB (Version R2012b, The MathWorks) with CPLEX (ILOG IBM, version 12.51) and Mosek (version 7) solvers; RAVEN Toolbox was used within Matlab environment (version 1.07, downloaded from http://129.16.106.142/tools.php?c=raven). Input data was taken from KEGG database (www.kegg.jp, up to date as of 18/11/2013), ToxoDB (www.toxodb.org, version 9), ApiLoc database (http://apiloc.biochem.unimelb.edu.au/apiloc/apiloc, version 3), and LLAMP portal (www.llamp.net, no evident version tracking).
|
10.1371/journal.pcbi.1000874 | A Model for a Correlated Random Walk Based on the Ordered Extension of Pseudopodia | Cell migration in the absence of external cues is well described by a correlated random walk. Most single cells move by extending protrusions called pseudopodia. To deduce how cells walk, we have analyzed the formation of pseudopodia by Dictyostelium cells. We have observed that the formation of pseudopodia is highly ordered with two types of pseudopodia: First, de novo formation of pseudopodia at random positions on the cell body, and therefore in random directions. Second, pseudopod splitting near the tip of the current pseudopod in alternating right/left directions, leading to a persistent zig-zag trajectory. Here we analyzed the probability frequency distributions of the angles between pseudopodia and used this information to design a stochastic model for cell movement. Monte Carlo simulations show that the critical elements are the ratio of persistent splitting pseudopodia relative to random de novo pseudopodia, the Left/Right alternation, the angle between pseudopodia and the variance of this angle. Experiments confirm predictions of the model, showing reduced persistence in mutants that are defective in pseudopod splitting and in mutants with an irregular cell surface.
| Even in the absence of external information, many organisms do not move in purely random directions. Usually, the current direction is correlated with the direction of prior movement. This persistent random walk is the typical way that simple cells or complex organisms move. Cells with poor persistence exhibit Brownian motion with little displacement. In contrast, cells with strong persistence explore much larger areas. We have explored the principle of the persistent random walk by analyzing how Dictyostelium cells extend protrusions called pseudopodia. These cells can extend a new pseudopod in a random direction. However, usually cells use the current pseudopod for alternating right/left splittings, by which they move in a persistent zig-zag trajectory. A stochastic model was designed for the persistent random walk, which is based on the observed angular frequencies of pseudopod extensions. Critical elements for persistent movement are the ratio of de novo and splitting pseudopodia, and, unexpectedly, the shape of the cell. A relatively round cell moves with much more persistence than a cell with an irregular shape. These predictions of the model were confirmed by experiments that record the movement of mutant cells that are specifically defective in pseudopod splitting or have a very irregular shape.
| Eukaryotic cells move by extending pseudopodia, which are actin-filled protrusions of the cell surface [1]. Pseudopod formation by Dictyostelium cells, like many other moving cells, shows a typical pseudopod cycle: upon their initiation, pseudopodia grow at a constant rate during their first ∼15 s and then stop. The next pseudopod is typically formed a few seconds later, but sometimes commences while the present pseudopod is still growing, giving rise to a cell with two pseudopodia. The fate of the pseudopod after its initial growth phase determines its role in cell movement: the pseudopod is either retracted, or is maintained by flow of the cytoplasm into the pseudopod thereby moving the cell body. The frequency, position and directions of the maintained pseudopodia form the basis of cell movement, because they determine the speed and trajectory of the cell. An important aspect of cell motility is the ability of cells to respond to directional cues with oriented movement. Gradients of chemicals give rise to chemotaxis [2]. Other directional cues that can induce oriented movement are temperature gradients (thermotaxis) or electric fields (electrotaxis) [3], [4]. These signals somehow modulate basal pseudopod extension such that, on average, cells move in the direction of the positional cues. In this respect, studies on cell movement are critical for understanding directional movement.
Cells in the absence of external cues do not move in random directions but exhibit a so-called correlated random walk [5]–[9]. This tendency to move in the same direction is called persistence, and the duration of the correlation is the persistence time. Cells with strong persistence make fewer turns, move for prolonged periods of time in the same direction, and thereby effectively penetrate into the surrounding space. Other search strategies for efficient exploration are local diffusive search and Levi walks [8], [10]. Can we understand the cell trajectory by analyzing how cells extend pseudopodia?
To obtain large data sets of extending pseudopodia we developed a computer algorithm that identifies the cell contour and its protrusions. The extending pseudopod is characterized by a vector that connects the x,y,t coordinates of the pseudopod at the beginning and end of the growth phase, respectively [11]. A picture of ordered cell movement has emerged from the analysis of ∼6000 pseudopodia that are extended by wild type and mutant cells in buffer [12]. Dictyostelium cells, as many other eukaryotic cells, may extend two types of pseudopodia: de novo at regions devoid of recent pseudopod activity, or by splitting of an existing pseudopod [12], [13]. Pseudopod splitting occurs very frequently alternating to the right and left at a relatively small angle of ∼55 degrees. Therefore, pseudopod splitting may lead to a persistent zig-zag trajectory [14]. In contrast, de novo pseudopodia are extended in all directions and do not exhibit a right/left bias, suggesting that de novo pseudopodia induce a random turn of the cells. We observed strong persistence for cells that extend many splitting pseudopodia. Conversely, mutants that extend mostly de novo pseudopodia have very short persistence time and exhibit a near Brownian random walk [12].
In this report we investigated the theory of correlated random walks in the context of the observed ordered extension of pseudopodia. The aim is to define the descriptive persistence time or average turn angle with primary experimentally-derived pseudopod properties. First we obtained detailed quantitative data on the probability frequency distributions of the size and direction of pseudopod activity. We then formulated a model that consists of five components: pseudopod size, fraction of splitting pseudopodia, alternating right/left bias, angle between pseudopodia and variance of this angle due to irregularity of cell shape. We measured the parameter values of these components for several Dictyostelium mutants with defects in signaling pathways or cytoskeleton functions. Subsequently, we used these observed parameters in Monte Carlo simulations of the model and compared the predicted trajectories with the observed trajectories of the mutants. The results demonstrate two critical components in these correlated random walks: the ratio of pseudopod splitting relative to de novo pseudopodia, and the shape of the cell.
The strains used are wild type AX3, pi3k-null strain GMP1 with a deletion of pi3k1 and pi3k2 genes [15], pla2-null with a deletion of the plaA gene [16], sgc/gca-null cells (abbreviated as gc-null cells) with a deletion of gca and sgc genes [17], sgc/pla2-null cells with a deletion of sgc and pla2A genes [18], and ddia2-null cells lacking the forH gene encoding the Dictyostelium homologue of formin [19]. Cells were grown in HG5 medium (contains per liter: 14.3 g oxoid peptone, 7.15 g bacto yeast extract, 1.36 g Na2HPO4⋅12H2O, 0.49 g KH2PO4, 10.0 g glucose), harvested in PB (10 mM KH2PO4/Na2HPO4, pH 6.5), and allowed to develop in 1 ml PB in a well of a 6-wells plate (Nunc). Movies were recorded at a rate of 1 frame per second for at least 15 minutes with an inverted light microscope (Olympus Type CK40 with 20× objective) and images were captured with a JVC CCD camera. Cell trajectories were recorded as the movement of the centroid of the cell as described [20].
Images were analyzed with the fully automatic pseudopod-tracking algorithm Quimp3, which is described in detail [11]. Briefly, the program uses an active contour analysis to represent the outline of the cell using ∼150 nodes [21]. Extending pseudopodia that satisfied the user-defined minimum number of adjacent convex nodes and the minimum area change were identified. The direction of each extending pseudopod was identified by the x,y and time coordinates of the central convex node of the convex area at the start and end of growth, respectively. The tangent to the surface at the node where the pseudopod started was calculated using the position of the adjacent nodes. The automated algorithm annotates each pseudopod as de novo versus splitting using the criterion that the convex area of the new pseudopod exhibits overlap with the convex area of the current pseudopod or is within a user-defined distance. The output files containing the x,y-coordinates of the start and end position of the pseudopod, the tangent of the surface and the annotation of the pseudopod were imported in Excel to calculate pseudopod size, interval, direction to gradient, direction to tangent, etc for de novo and splitting pseudopodia (see Fig. 1), as well as fraction s of pseudopod splitting and alternating Right/Left bias a (RL +LR)/total splitting; Table 1).
The cell shape parameter Ψ was determined as follows: Using the outline of the cell with ∼150 nodes, two ellipsoids were constructed, the largest ellipse inside the cell outline and the smallest ellipse outside the cell outline. Then an intermediate ellipse was constructed by interpolation of the inner and outer ellipse. This intermediate ellipse makes several intersections with the cell outline, thereby forming areas of the cell that are outside the intermediate ellipse (with total surface area O), and areas of the intermediate ellipse that do not belong to the cell (with total surface area I; see Fig. S4). The intermediate ellipse was positioned in such a way that (this also implies that the surface area of the cell (T) is identical to the surface area of the intermediate ellipse). The cell shape parameter is defined as ; it holds that . For a cell with a regular shape that approaches a smooth ellipsoid, the surface areas O and I are very small and Ψ approaches zero. In contrast, O and I are larger for a cell with a very irregular shape; the largest value observed among ∼600 cells was Ψ = 0.92.
With the exception of 5h starved cells, each database contains information from 200–300 pseudopodia, obtained from 6–10 cells, using two independent movies. For 5h starved cells, we collected a larger database containing 835 pseudopodia from 28 cells using 4 independent movies, and typical databases for each mutant. The data are presented as the means and standard deviation (SD) or standard error of the means (SEM), where n represents the number of pseudopodia or number of cells analyzed, as indicated in Table 1.
The probability density functions of angles can not be analyzed as the common distribution on a line. Angular distributions belong to the family of circular distributions, which are constructed by wrapping the usual distribution on the real line around a circle. The data were analyzed with two circular distributions, the von Mises distribution (vMD), which matches reasonably well with the wrapped normal distribution, and the wrapped Cauchy distribution (WCD), which has fatter tails [22]. The vMD is given by(1)where I0(κ) is the modified Bessel function of the first kind of order zero(2)The WCD is given by(3)
Pseudopod extension is an ordered stochastic event [12]. The position of the tip of the formed pseudopodia depends on pseudopod size λp, splitting fraction s, Left/Right alternating ratio α, angle between split pseudopodia φ and variance of this angle σφ. A Monte Carlo simulation starts with a random angle α(1) of the first pseudopod. For the next and all subsequent pseudopodia the simulation uses four uniformly distributed random numbers Ri,n (i = 1, .., 4) to calculate α(n), the angle of the nth pseudopod: with the decision to split if R1,n <s; with the decision for alternating splitting if R2,n <a; for direction of split after de novo with decision right if R3,n <0.5; and for the direction of the de novo pseudopod. These probabilities result in a projected angle of extension in degrees. Finally, the actual pseudopod direction is drawn from a wrapped von Mises distribution with this projected angle as mean and σφ2 as variance (κ = 1/σφ2; variance converted to radians). The obtained α(n) and the pseudopod size λp are used to calculate the x,y coordinates of the tip of the pseudopod, followed by a next round of four random numbers to calculate α(n+1). In the simulations reported here we did not include stochastic variation in pseudopod length and pseudopod frequency, since we observed that they had only minor effects on the trajectories over several cell lengths.
Please note that in the simulations the direction of the simulated de novo pseudopodia is random; consequently, a small fraction of de novo pseudopodia are in the same direction of the previous pseudopod, which would be recognized in experiments as splitting pseudopodia. Conversely, a small fraction of the simulated splitting pseudopodia have angles much larger than 55 degrees and would be recognized in experiments as de novo pseudopodia. From the geometry of the cell, we estimate that the number of simulated de novo in the current pseudopod and the number of splitting pseudopodia outside the current pseudopod are approximately the same, suggesting that the simulations represent the observed ratio of splitting and de novo pseudopodia.
The angles between pseudopodia were analyzed in detail and the results are presented in Fig. 2. For splitting pseudopodia, the angle between the current and next pseudopod (φ1,2) has a clear bimodal distribution (Fig. 2A). A probability density function (PDF) of angles belongs to the family of circular or wrapped distributions. The data reported in this study were all fitted well by a von Mises distribution (vMD), which is the circular analog of the normal distribution. The wrapped Cauchy distribution has fatter tails and provided a poorer fit of the data (data not shown). The bimodal vMD presented in Fig. 2A is symmetric, yielding two means (φ1,2 = +/−55) that have the same variance κ = 1/σφ2; σφ1,2 = 28 degrees). Figure 2B shows the PDF of the angle between the current and next-next pseudopod (φ1,3), which is best described by a single vMD with a mean of φ1,3 = 2 degrees and σφ1,3 = 42 degrees. Figure 2C reveals that there is no significant correlation between the magnitude of angles between first/second pseudopod and the magnitude of the angles between second/third pseudopod (thus e.g. splitting at a larger angle is not followed by a split at a smaller angle). The extension of splitting pseudopodia is summarized in Fig. 2D, and is based on the previous observation that a pseudopod split to the right is frequently followed by a split to the left and visa versa [12]. Thus the next pseudopod is extended at an angle of ∼55 degrees to the right or left relative to the current pseudopod, and the next-next pseudopod is extended in roughly the same direction as the current pseudopod.
The angle between a de novo pseudopod and the previous pseudopod shows a very broad distribution (Fig. 2E). Nearly all angles between −180 and +180 are well represented with a somewhat lower abundance of angles around 0 degrees. This suggests that a de novo pseudopod can be extended in any direction, but with slightly lower probability of the direction of the current pseudopod.
To investigate the consequence of the observed ordered extension of pseudopodia for cell movement on a coarse time scale for many pseudopodia we recorded the movement of Dictyostelium cells during 15 minutes; in this period about 30 pseudopodia are extended. Previously we have presented the cell trajectories for several strains and developmental stages [12] (see also Fig. S1). The mean square displacement as a function time, , exhibits a slow approach to a linear function (Fig. 3A), which is typical for a transition of a correlated random walk at short times to a Brownian random walk after longer times [6], [23]. Previously, the often used equation for a correlated random walk were fit to the data points to estimate persistence time and speed of the cells [12]. The aim of the present study is to analyze the mechanism of cell movement from the perspective of the extending pseudopodia, which have a specific length and direction. A correlated random walk in two dimensions can also be described with steps and turns [24], [25]. With the replacement of the number of steps (n) in Eq. 7 in reference [25] for n = Ft we obtain(4)where λ is the step size in µm, F is the step frequency, and γ is the correlation factor of dispersion (0<γ<1), defined as the arithmetic mean of the cosine of the turn angle θ between steps(5)With three variables (F, λ, γ) the estimates of the parameters become uncertain. Fortunately, the step size can be deduced accurately from experimental data. As will be shown below in Eq. 10, the step size is given by λ = λpcos(φ/2), where measurements for λp and φ are presented in Table 1. Using this value for λ, the dispersion data were fitted to obtain the observed correlation factor of dispersion (γobs) with the corresponding turn angle (θ). In cells starved for 1 or 3 hours the correlation factor is only ∼0.5 with turn angle of ∼60 degrees. At 5 and 7 hours of starvation, cells move with much stronger persistence (correlation factor of 0.74 and 0.81 and a small turn angle of 42 and 36 degrees). Deletion of PLA2 or guanylyl cyclases prevents this increase of correlation factor, persistence is very low and cells disperse poorly.
How is pseudopod extension related to the observed correlation factor of dispersion γobs? As previously stated (see Fig. 2), Dictyostelium cells may extend either de novo pseudopodia in nearly random directions, or splitting pseudopodia in a direction similar to the previous direction. Therefore, cells that extend exclusively de novo pseudopodia are expected to exhibit a random walk with γobs = 0 (turn angle θ = 90 degrees), whereas cells extending exclusively splitting pseudopodia will exhibit strong persistence with large γ and small turn angle θ. As a consequence, γobs is expected to depend on the ratio s of splitting/de novo pseudopodia. Fig. 3B demonstrates that within experimental error this relationship is approximately linear; this holds true for the mutants as well as for wild type cells at different stages of development. The linear regression of all data yields γobs = 0.921s−0.044. Thus, when all pseudopodia are de novo (s = 0) the correlation factor is small (γobs = −0.044) giving a turn angle = 93 degrees, close to the expected value of 90 degrees for random turns. In contrast, when all pseudopodia are the result of splitting (s = 1) the correlation factor is large (γobs = 0.88) yielding a small turn angle ( = 29 degrees). The implication of small turn angles for splitting pseudopodia will be discussed later.
The alternating right/left extension of splitting pseudopodia can be used to simplify a description of the movement of Dictyostelium cells over longer distances. In this approach, the simplification may be valid for movement on a longer time scale only, as we study here, but may not be appropriate over shorter time scales of a few pseudopodia. Because pseudopodia are frequently extended alternating right/left, we consider movement by pairs of two pseudopodia.
Figure 4 shows four possibilities of pairs of splitting pseudopodia, which are the RL, LR, RR and LL, each with corresponding probabilities and angles as indicated. In addition to these splitting pairs, three combinations with de novo pseudopodia are possible: split-de novo, de novo-split, and de novo-de novo. The correlation factor of dispersion yields for the seven pairs:(6)De novo pseudopodia are extended in a random direction, i.e. , and equal zero. The turn angles of the four splitting pairs are 0, φ and 2φ, as indicated in Fig. 4A, and the variance is approximately 2σφ2 (see Fig. 2B). Consequently Eq. 6 reduces to:(7)where denotes the expected value of the cosines of the angles on a circle with weights given by the vMD with mean φ and variance given by κ = 1/(2σφ2). Since all splitting pseudopodia show the same variance this can be further reduced to(8)In this equation is obtained by calculating the probabilities of all turn angles on a circle with the vMD using Eqs. 1 and 2 and then taking the weighted average of the cosines of these angles. Although this procedure is straightforward, Eq. 8 can be further simplified, because for σφ smaller than ∼50 degrees a good approximation is (see Fig. S2). Finally, on a longer time scale and averaged over many steps, the correlation factor of pairs is related to the correlation factor of its underlying two steps by . With these replacements we obtain the analytical expression for the correlation factor(9)Thus, the correlation factor γ is the product of three terms: the splitting ratio s, a noise term with the variance σφ, and a term with right/left bias a and angle φ.
Finally, by considering movement in pairs of steps, Fig. 4 reveals that the step size of the displacement is given by(10)
We used Monte Carlo simulations to investigate how λ and γ depend on the pseudopod parameters size λp, splitting fraction s, alternating ratio a, angle between split pseudopodia φ and variance of this angle σφ2. These simulations are also useful to inspect whether step size λ and correlation factor γ are correctly described by Eqs. 8–10. The direction in which a pseudopod is extended appears to be an ordered stochastic event [12] that depends on multiple decisions according to the following scheme: The next pseudopod is a splitting or de novo according to the ratio s. A splitting pseudopod is extended with angle φ to the right or left relative to the previous splitting according to the alternating right/left bias a. A de novo pseudopod is extended in a random direction. The splitting pseudopod that appears after a de novo pseudopod has an equal probability to be extended to the left or right. Finally, the direction of the emerging pseudopod has a variance σφ2. The Monte Carlo simulation starts with a random angle α(1) of the first pseudopod and then uses the probabilities for splitting fraction s, alternating ratio a, angle between split pseudopodia φ and variance of this angle σφ2 to stochastically simulate the angle of the next pseudopod (see methods and Table 1 for pseudopod parameters of wild type and mutant cells). The simulated trajectories are qualitatively similar to the experimentally observed trajectories (see Fig. S1B): Fed wild type cells or mutants with abundant de novo pseudopodia make many turns and have small displacement, whilst the trajectories of starved wild type cells with abundant pseudopod splitting are more persistent with large displacements.
To investigate how the correlation factor γ depends on pseudopod parameters, the displacement was calculated from 100,000 trajectories obtained by MC simulation using a unit pseudopod size and different values of s, a, φ and σφ. The obtained displacement was then fitted to Eq. 4 to obtain estimates for the step size λ and Monte Carlo correlation factor, γMC. The symbols in Fig. 5 show the results of the MC simulation, whereas the curves are the result of Eqs. 8–10.
We first investigated the angle φ between splitting pseudopodia and the alternating right/left bias a. When all splitting pseudopodia are alternating (a = 1), the cells make a nearly perfect zig-zag trajectory, and therefore the angle φ has very little effect on the persistence factor γ (Fig. 5A). When splitting pseudopodia are extended in a random fashion to the right or left (a = 0.5), the persistence factor γMC decreases sharply as φ becomes larger than ∼30 degrees. At an intermediate right/left bias (a = 0.75) the persistence factor γMC remains relatively high as long as the angle between pseudopodia is below 60 degrees. The results of the MC simulation appear to be described very well by the simplified model (Eqs. 8–10). Furthermore, at the observed angle of φ = 55 degrees and alternating factor of a = 0.77, the deduced persistence factor γMC is 0.88 (see asterisk in Fig. 5A).
The fraction of splitting pseudopodia has a major impact on the persistence factor γ. In the MC simulations, the value of γ declines approximately linearly with the value of s (Fig. 5B), as was also observed experimentally (Fig. 3B), and obtained in Eqs. 8 and 9. Finally, we investigated the contribution of the variance σφ2 of the splitting angle to the persistence factor γ. This reveals that the persistence decreases strongly with increasing variance (Fig. 5C), with γMC following an approximately linear relationship with cos(σφ). The MC simulations are well described by Eq. 8, but deviate from Eq. 9 at σφ>30 degrees, as expected (see Fig. S2).
We also used these Monte Carlo simulations to obtain an estimate of the step size λ. It appears that λ does not to depend on s and a, but depends on φ according to (Inset Fig 5A), as was obtained in Eq. 10.
In summary, the obtained correlation factor from the MC simulation (γMC) are nearly identical to the correlation factor calculated with Eq. 9 (γstep). This suggests that the movement of Dictyostelium cells is qualitatively and quantitatively described very well by the model of persistent steps and random turns with the observed pseudopod parameters λp, s, a, φ and σφ.
How does the movement of pseudopodia relate to the movement of the centroid of the cell? The data presented in Table 1 reveal that the observed correlation factor γobs of the centroid for different cell types correspond well with the deduced correlation factors of the pseudopods (γMC and γstep), but is always larger by ∼15% (Table 1). Apparently, the observed turn angle of the cell's centroid is smaller than the turn angle of the extending pseudopod. Inspection of movies of 5h starved AX3 cells confirm this notion: the average angle between splitting pseudopodia is 55±28 degrees (Fig. 2A), while the centroid moves during period at an angle of only 31±23 degrees (mean and SD). Equation 9 reveals that the correlation factor γstep increases by 15% when φ = 55±28 degrees for the pseudopod is replaced by φ = 31±23 degrees for the cells centroid.
Probably two phenomena are responsible for the difference between pseudopod and centroid: extension of multiple pseudopodia and geometry of cells. When cells extend multiple pseudopodia it is likely that at any given instant of time, the front of the cell moves with a fixed fraction of the vector sum of velocities possessed by the pseudopodia active at that instant in time. The temporal overlap of two pseudopodia was deduced from the measured probability distributions of pseudopod extensions (Fig 2F in [26]), which reveal that ∼25% of the pseudopodia overlap with another pseudopod during on average ∼40% of their extension time. This suggest that the tip of the cell moves at an angle that is ∼6 degrees smaller than 55 degrees. Secondly, geometry predicts that the rear of the cell makes smaller changes of direction than the tip of the cell, comparable to the differences in curvature made by the front and rear wheels of a car. Figure S3 indicates that for a stereotypic pseudopod at 55 degrees the directional change of the centroid is ∼40 degrees (see Fig. S3). Together, multiple pseudopodia and cell geometry can explain observed difference between pseudopod and centroid changes of direction, leading to the small 15% difference between deduced pseudopod correlation factor (γMC and γstep) and observed centroid correlation factor (γobs).
The directional displacement is the displacement after n steps in the direction of the first step. An expression for the directional displacement is especially relevant when the organism is exposed to positional cues leading to a drift in one direction, such as during chemotaxis. The directional displacement of a cell after extending one pseudopod at an angle θ is , and for a population of cells . By Eqs. 3 and 10, the displacement at the first step may be written as , and at the ith step , see Eq. 6 of reference [25]. The cumulative displacement after n steps is(11a)which at is given by(11b)In essence, this equation describes the displacement of a cell population in which all cells extend the first pseudopod in the same direction. Subsequent pseudopodia are extended with a bias, which reduces geometrically with each step; the correlation factor γ indicates how many pseudopodia have correlated direction and therefore how far the population will disperse in the direction of the first pseudopod. Figure 6 presents the directional displacement as observed experimentally in wild type cells. The displacement in the direction of the first pseudopod slowly decreases at each subsequent pseudopod, approaching random movement after ∼10 pseudopodia. On average a cell moves ∼15 µm in the direction of the first pseudopod, which is the equivalent of about 3 pseudopodia (given a pseudopod size of ∼5 µm). This figure also presents the directional displacement as modeled by Eq. 11a with observed data for λp, φ and γ, which is in very close agreement with experimental data, again suggesting that the movement of a cell is satisfactory described by the model with five pseudopod parameters.
The variation in pseudopod direction σφ2 plays an important role in Eqs. 8–11 describing cell dispersal. Previously [12] we have shown that the next pseudopod emerges at a specific distance d from the tip of the current pseudopod, and is then extended perpendicular to the cell surface (i.e. perpendicular to the tangent to the surface curvature at the position where the pseudopod emerges). The pseudopod direction is expected to have high confidence for cells with a smooth ellipsoid shape, because the local bending is very predictable. However, this confidence is much smaller for cells with a very irregular shape. We investigated the role of cell shape using three experiments. First we demonstrate that the variance σφ2 indeed depends on the variance of the tangent and the normal to the tangent. Second, we show that wild type or mutant cells with irregular shape exhibit increased variance σφ2. Finally we show that, due to the increased variance, the mutant exhibits poor dispersal.
Quimp3 was used to construct the tangent to the surface curvature at the position where the pseudopod emerges. We first determined for wild-type cells the angle αt of this tangent relative to the previous pseudopod (αt = 34.5±24.9 degrees), and the angle β of the new pseudopod relative to this tangent (β = 89.1±13.3 degrees). As mentioned above, the observed angle of the new pseudopod relative to the previous pseudopod is φ = 55.2±27.8 degrees. We expect that the angle of the tangent relative to the previous pseudopod is independent from the angle of the pseudopod relative to the tangent; therefore we expect . Indeed, the observed standard deviation of 27.8 degrees is close to this expected value of 28.2 degrees. Importantly, the largest contribution to σφ2 is derived from the variance of the tangent σt2, which is related to the local shape of the cell.
In the collection of Dictyostelium mutants, we selected a strain with an irregular shape. Mutant ddia2-null with a deletion of the forH gene encoding the formin dDia2 has a star-like shape (Fig. 7C). In this mutant, new pseudopodia are extended at about the same frequency and distance from the present pseudopodia as in wild type cells, pseudopodia also grow perpendicular to the surface, and are extended roughly in the same direction of φ = 55 degrees as wild type cells (Fig. 7A). However pseudopodia exhibit much more variation in direction (σφ = 47 degrees compared to σφ = 28 degrees for wild type cells). Finally, we determined a shape parameter Ψ that indicates how much the cell outline deviates from an ellipse (see method section and Fig. S4). Figure 7B reveals that cells with increased irregular shape, either being wild-type or mutant, exhibit strongly increased variance σφ2. Importantly, the distance d and angle φ of the pseudopodia does not change with cell shape (Fig. 7A).
Using the observed values for s, a, φ, and σφ for ddia2-null cells we expect from Eq. 9 to obtain γstep = 0.43, significantly lower compared to γstep = 0.69. for wild type cells. Fig. 7D shows that the dispersion of ddia2-null cells is strongly reduced. The observed mean square dispersion was fitted to Eq. 4 yielding a correlation factor of γobs = 0.53 (Table 1), close to the value that was predicted from the extension of pseudopodia from an irregular surface.
In summary, these and previous results [12] suggest that a splitting pseudopod is induced at some distance d from the tip of the current pseudopod, and then grows perpendicular to the surface. In a cell with a regular shape, the tangent and therefore pseudopod direction can be approximated using the distance d; alternating R/L extensions lead to a relative straight zig-zag trajectory, providing strong persistence of movement. In a cell with a very irregular shape, the local curvature of the membrane at distance d is unpredictable. Consequently, alternating R/L splitting occur with large variation of directions, leading to frequent turns and poor persistence.
The movement of many organisms in the absence of external cues is not purely random, but shows properties of a correlated random walk. The direction of future movement is correlated with the direction of prior movement. For organisms moving in two dimensions, such as most land-living organisms, this implies that movement to the right is balanced on a short term by movement to the left to assure a long-term persistence of the direction. In bipedal locomotion, the alternating steps with the left and right foot will yield a persistent trajectory. Amoeboid cells in the absence of external cues show ordered extension of pseudopodia: a new pseudopod emerges preferentially just after the previous pseudopod has stopped growth [12]. Importantly, the position at the cell surface where this new pseudopod emerges is highly biased. When the current pseudopod has been extended to the left (relative to the previous pseudopod), the next pseudopod emerges preferentially nearby the tip at the right side of the current pseudopod. Since pseudopodia are extended perpendicular to the cell surface, this next pseudopod is extended at a small angle relative to the current pseudopod [12]. Therefore, this (imperfect) alternating right/left pseudopod splitting resembles bipedal locomotion. Cells may also extend a de novo pseudopod somewhere at the cell body, which is extended in a random direction. In starved Dictyostelium cells, the probability of extending a de novo pseudopod is ∼10-fold lower than of pseudopod splitting (probability calculated per µm circumference of the cell [12]).
The model for pseudopod-based cell dispersion depends on five parameters, the pseudopod size (λ), the fraction of split pseudopodia (s), the alternating left/right bias (a), the angle between pseudopodia (φ) and the variance of this angle (σφ2). With these parameters the experimental data on mean square displacement and directional displacement are well-explained using Eqs. 9 and 11, respectively. Pseudopodia are the fundamental instruments for amoeboid movement. The notion that the trajectories are described well by the five pseudopod parameters probably implies that we have identified the basic concept of the amoeboid correlated random walk: persistent alternating pseudopod splitting and formation of de novo pseudopodia in random directions.
The cells may modify one or more of these five pseudopod parameters in order to modulate the trajectories (see Table 1). Nearly all mutants, as well as wild type cells at different stages of starvation and development, have approximately the same average pseudopod size λp. In addition, the alternating right/left bias (a) fluctuates between 0.67 and 0.82, and the angle between splitting pseudopodia (φ) between 50 and 62 degrees. [pla2-null cells are the only exception [12]; emerging pseudopodia in pla2-null cells exhibit longer growth periods (∼27 s) than wild type cells (∼13 s), and are thus longer]. This suggests that all strains use the same mechanism for pseudopod splitting. In contrast to these constant properties of split pseudopodia, the fraction of split pseudopodia (s) changes dramatically upon starvation, and appears to be regulated by cGMP and PLA2 signaling. Well-fed cells extend pseudopodia that are predominantly de novo in random directions, leading to a nearly Brownian random walk [20]. Upon starvation, the appearance of cGMP and PLA2 signaling enhances splitting and suppresses de novo pseudopod extensions, which leads to more persistent movement. The important role of the fraction of splitting pseudopodia for cell movement is also depicted by the linear dependence of the correlation factor γ on the fraction s of splitting pseudopodia (Fig. 3B and Eq. 9).
The variance of the angle of pseudopod extension (σφ2) plays an important role in movement. In wild type cells, as well as in many mutant strains, σφ is about 28 degrees. The primary source of the variance of pseudopod angles lies in the variation of cell shape, by which the normal to the cell surface at a specific position on this surface will have significant variation. Since the direction of pseudopodia is given by this normal, it is predicted that a cell with irregular shape should have more variation in pseudopod direction, and consequently shows poor dispersion. The experiments with mutant ddia2-null cells strongly support this interpretation. Wild-type cells have a relatively regular spherical shape by which two nearby pseudopodia are extended in nearly the same direction (small σφ). In contrast, mutant ddia2-null cells have an irregular star-like shape; therefore, two nearby pseudopodia are often extended in very different directions (large σφ). The variance σφ2 can be regarded as the noise of the system. It indicates how fast a cell that extends only alternating splitting pseudopodia (a = 1 and s = 1) will lose correlation of directionality. With σφ = 28 degrees for wild type cells it follows from Eq. 10 that after ten pseudopodia the correlation of direction is still ∼0.5. In contrast, for ddia2-null cells we obtained σφ = 46.5 degrees, which implies that already after four pseudopodia the correlation of direction has declined to ∼0.5. Supported by Monte Carlo simulations using the parameters of the mutant, we conclude that poor dispersion of ddia2-null cells is due to the increased variance of pseudopod angles, which is caused by its irregular shape.
The correlation factor γ is the product of three terms (see Fig. 5 and Eq. 9), namely: splitting fraction (s), alternating pseudopod angles (a and φ), and the SD of the pseudopod angle (σφ). Strong persistence of cell movement is attained when all three terms are large and about equal in magnitude. Starved wild type cells follow this strategy: each term is ∼0.9, resulting in the observed correlation factor of 0.74. Mutants in which one of these terms is compromised, such as reduced splitting in sgc/pla2-null cells or enhanced noise of ddia2-null cells, have poor dispersion.
In summary, the correlated random walk of amoeboid cells is well described by the balanced bipedal movement, mediated by the alternating right/left extension of splitting pseudopodia. Cells deviate from movement in a straight line due to noise and because cells occasionally hop or make random turns. The turns in particular are used by the cells to modulate the persistence time, thereby shifting between nearly Brownian motion during growth and strong persistent movement during starvation.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.